Datasets:

babytreecc

/

ICLR2025review

like 7

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "The approach’s transparency in generating adversarial examples for LLMs raises ethical concerns, particularly around misuse, and would benefit from a clearer discussion of safeguards to prevent potential exploitation." }, "flag_for_ethics_review": { "value": [ "Yes, Potentially harmful insights, methodologies and applications" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See Weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- TIARA presents a tokenizer-independent framework for transferable adversarial attacks, which effectively bypasses constraints associated with fixed tokenization schemes and gradient access, making it adaptable across diverse LLM architectures.\n- TIARA employs a multi-stage candidate selection process that iteratively refines adversarial candidates based on both validation and test metrics, maximizing attack success rates across multiple target models.\n- TIARA offers a systematic analysis of adversarial strategies, identifying categories such as formatting manipulation and instruction/context manipulation, which advance understanding of adversarial patterns that exploit LLM vulnerabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the Tokenizer-Independent Adversarial Red-teaming Approach (TIARA ), a framework designed to improve AI safety by automating the red teaming of large language models (LLMs). TIARA allows for the generation of transferable adversarial examples without the need for gradient access or fixed tokenizers, facilitating effective attacks across diverse model architectures. Its contributions include a tokenizer-agnostic method for generating adversarial inputs, a gradient-free optimization technique for exploring token-level perturbations, and an automated approach for discovering model vulnerabilities, which aim at identifying potential risks and advancing the development of more secure Chatbot systems." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- TIARA’s multi-stage candidate selection and extensive perturbation sampling require significant computational resources, making it less feasible for deployment in real-time scenarios.\n- While TIARA is gradient-free, it relies on direct access to model logits for loss computation, which may limit applicability to models where even logit-level access is restricted.\n- The paper does not extensively test TIARA against models equipped with advanced, system-level defensive mechanisms, making it unclear how resilient the approach is in more robustly defended environments.\n- TIARA lacks an evaluation against input transformation defences, such as synonym replacement or back-translation, which could easily disrupt the token-level perturbations TIARA relies on, potentially reducing its success rate in adversarial bypass attempts.\n- TIARA’s reliance on token-level perturbations to generate effective adversarial examples may limit its success on models or tasks where semantic-level manipulations are more impactful. This limitation suggests a need for a more flexible perturbation strategy that balances token- and semantic-level alterations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. I wonder the intuition and design insight.\n2. How to scale to LLM with tokenizer of much larger vocabulary?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ An effective attack that outperforms previous jailbreak.\n+ Evaluation is quite thorough" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel jailbreak method towards LLM based on the attack transferability between the source LLM and the target LLM to achieve tokenzier-agnostic.\nComparing to previous gradient based attack, the attack does not require gradient to optimize jailbreak prompt.\nInstead, at each step the attack select best perturbed candicate and use multiple LLM to enhance the attack.\nFinally, through experiments, the authors show that the proposed attack outperforms pervious methods and has high attack success rate on commercial LLM like GPT-3.5." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The idea of candicate search has been explored in previous work and the usage of several LLMs to increase the attack transferability is a common technique (which is widely used to increase the adversarial tranferability for image classification models).\nSpecifically, BEAST attack [1] also investigates how to generate jailbreak prompt without using the gradient. Their beam search-alike method is similar to the TIARA attack here because it essentially selects the potentially best candidates to the target LLM. The differences I see include: TIARA uses multiple source LLM, multiple loss (teacher forcing and autoregressive) and disconnect the tokenizer and source LLM. However, these improvements are not milestone techniques in further improvement jailbreak attack.\n\n- Some design choices, intuition and attack insight, are not clear. For example, how to design the teacher-forcing loss and autoregressive loss? why choose these two loss functions? The use of multiple source LLM can increase the jailbreak attack success rate, but what are the potential reason behind it? There is neither qualitative nor quatitative analysis.\n\n- The tokenizer-agnostic is somehow overclaimed and limited when the target LLM has larger vocabulary. As show in Figure 2, the GPT-4 has lowest ASR comparing to the rest. As the api's tiktoken shows, GPT-4's vocabulary size is much larger than existing opensoured LLMs (which are similar to or better than the GPT-3.5). If future LLM (that should be red-teamed) has larger vocabulary, TIARA may be limited in this case. In other words, my concern is that TIARA may be only effective in red-teaming existing LLMs.\n\n[1] Fast Adversarial Attacks on Language Models In One GPU Minute. ICML 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- What was the typical number of iterations for an attack success to be found? Was there a relationship between the longer the optimization and the higher attack transferability?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The method is empirically effective, simple to implement, and investigates a realistic threat model. Furthermore, with the current perturbation function in the paper we can see that enough random perturbations are sufficient to eventually converge to a jailbreak. I think this is an interesting find that jailbreaks can essentially be brute forced in such a manner on a large scale with sufficient (though not unattainable) compute resources. \n\n- Using the loss function proposed in the paper combined with an evolutionary search strategy is interesting, and is a different approach compared to popular methods using LLM as a judge to guide evolutionary based strategies (e.g. TAP/PAIR) - potentially this indicates that powerful LLMs as fitness functions in this style of attack may not be needed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an algorithm for generating attacks that are highly transferable between open source models and closed source target models. With gradients not being computable on the closed source models (and potentially query numbers being limited), relying on transfer attacks can be a realistic and practical attack method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It would have been useful too compare with different perturbation functions: the paper shows a random approach, but as highlighted in the work it could be any function. Indeed the random perturbation function is very computationally expensive - 1024 pertubrations per iterations, up with to 1k iterations: in other words, it could be over 1 million separate model calls. Further, additional perturbation functions could highlight if the proposed loss is the most effective under different perturbation strategies.\n\n- The paper flow and balance could be improved: for example, the description of the Multi-Stage Candidate Selection which is a key competent of the algorithm (and is listed as a key contribution of the paper) is in the appendix, and could use a clearer explanation of the steps. \n\n- It is unclear why Llama3 was used as a transfer model in combination with others, but not used either stand alone, or not evaluated against direct attacks in Table 2." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Excellent transfer attack performance.** The reported results are inspiring, especially those against GPT models.\n2. **The practicality of the method.** Although an ensemble of multiple models is quite an old talk for universal adversarial attack, it is arguably the most feasible method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores transfer attacks against LLMs. Unlike previous works that utilize gradient to optimize a suitable prompt, this work employs a perturbation-based strategy. The method employs additional models as oracle models and judges the fitness of the current adversarial prompt with the weighted scores of multiple source models. The scoring is based on two loss functions, in which teacher-forcing loss ensures the completion of the target prompt and the auto-regressive loss avoids the occurrence of the deviant tokens. The attack performance is demonstrated via comprehensive experiments. The attacks against GPT models is impressive." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Inadequate exposure of method details and experiment details.** What about the setting of temperature and top_k for both the auto-regressive loss and evaluation? What about the choice of the subset of candidate tokens that will be sampled mentioned in line 262. After reading the paper, I am still not sure whether the perturbation tokenizer corresponds to the tokenizer of the source model or the one of the target model.\n2. **Ad-hoc attack configuration.** The choice of source models is ad-hoc and will impact the attack performance a lot, as mentioned in Section 4.1. Experiments about the lower-bound and upper-bound performance are in demand, corresponding to the weakest source model and the strongest source model.\n3. **The analysis of tokenizer behaviors is not in-depth enough.** Although different models use distinct tokenizers, it does not mean they exhibit totally different behaviors. For example, for the same tokenization algorithm, a larger vocabulary of tokenizer A may contain the smaller vocabulary of tokenizer B. Analyzing the overlap between different tokenizer behaviors, especially on the crafted prompts, will help us understand whether the transfer attack leverage semantics or just token combinations to make sense.\n4. **Exploration of Attack stability.** As demonstrated in Table 2, the TIARA yields an almost 100\\% ASR. As TIARA is in essence an evolutionary method, the stability of the success is amazing. Can the authors give more details about the evolutionary trajectories and analyze the countable failed cases?\n5. **Unclear standpoint.** Although the paper emphasizes the tokenizer-agnostic feature of their attack, it is unclear about the difference between tokenizer-agnostic and model-agnostic due to the superficial exploration of the tokenizer behavior. The authors should add a discussion section that clearly defines and distinguishes between tokenizer-agnostic and model-agnostic approaches, and explicitly positions your method within this framework.\n\nI would like to raise my score if the authors can answer my concerns, especially the one about tokenizer behavior (weakness 3)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce TIARA, a tokenizer-independent method for generating transferable adversarial attacks on language models for enhanced red teaming." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024tokenizeragnostic,\ntitle={Tokenizer-Agnostic Transferable Attacks on Language Models for Enhanced Red Teaming},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4GcZSTqlkr},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) have become increasingly prevalent, raising concerns about potential vulnerabilities and misuse. Effective red teaming methods are crucial for improving AI safety, yet current approaches often require access to model internals or rely on specific jailbreak techniques. We present TIARA (Tokenizer-Independent Adversarial Red-teaming Approach), a novel method for automated red teaming of LLMs that advances the state-of-the-art in transferable adversarial attacks. Unlike previous token-level methods, TIARA eliminates constraints on gradient access and fixed tokenizer, enabling simultaneous attacks on multiple models with diverse architectures. By leveraging a combination of teacher-forcing and auto-regressive loss functions with a multi-stage candidate selection procedure, it achieves superior performance without relying on gradient information or dedicated attacker models. TIARA attains an 82.9\\% attack success rate on GPT-3.5 Turbo and 51.2\\% on Gemini Pro, surpassing previous transfer and direct attacks on the HarmBench benchmark. We provide insights into adversarial string length effects and present a qualitative analysis of discovered adversarial techniques. This work contributes to AI safety by offering a robust, versatile tool for identifying potential vulnerabilities in LLMs, facilitating the development of safer AI systems." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Adversarial Attacks", "Red Teaming", "Transferable Attacks", "AI Safety", "Large Language Models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/61cc817d83bfdd1fe57925009f9c91c4ecc765ba.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/a32b4ace946848188da27d0b33504f6cdf3119c3.zip" }, "title": { "value": "Tokenizer-Agnostic Transferable Attacks on Language Models for Enhanced Red Teaming" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- From Section 3 and 4, only one attribution mode is considered. For more complex or practical scenarios, all attributions may have their mode variables. Whether the SD-HC can efficiently and effectively address them?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper establishes a novel theory for supervised disentanglement with hidden correlations. It consists of both necessary and sufficient conditions, serving as a significant improvememt for existing works.\n- The experiments are comprehensive and the results are good enough." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel supervised disentanglement method, SD-HC, designed to address hidden correlations in multi-modal data, which traditional DRL methods often overlook. SD-HC focuses on preserving essential mode information within attributes by minimizing mode-based conditional mutual information, thus achieving better disentanglement and predictive accuracy under both hidden and attribute correlations. The authors theoretically prove that mode-based conditional mutual information minimization is both necessary and sufficient for effective disentanglement in complex, correlated datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The motivation of the problem setting is not clearly discussed. In Figure 2, to study the hidden correlations among attributions, why not directly study $a_k$ and $a_{-k}$? The role or advantage of introducing the variable $m_k$ is not clear for me.\n- It seems that the performance of SD-HC heavily relies on the mode label estimation. First, in real applications, how to determine the number of modes $N_m$? In addition, for complex data pattern, the data representations are not high-quality and the clutering is not accurate accordingly. In such case, is there any techniques to guarantee good performance?\n- There are some ambiguous concepts in this paper. For example, 1) \"Data mode\". I do not find its concrete definition. This term is not common in machine learning and its meaning varies with contexts. 2) \"Multi-modal\". It seems not to refer to the commonly used meaning, such as text and image. More strict definitions is required to make this paper more clear.\n- The writting needs to be further improved. For example, in line 155, $c^a$ and $c^m$ should be boldface." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See the weakness part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper studies an important problem in disentangled representation learning by considering hidden correlations, which are prevalent in real-world data.\n2. The experimental evaluation is thorough, including both synthetic and real-world datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the problem of disentangled representation learning in the presence of hidden correlations, which are common in real-world data. The authors consider a multi-modal data distribution where certain attributes may induce complex modes correlated with other attributes. To tackle this issue, they first investigate the necessary and sufficient conditions for supervised disentanglement under the proposed data-generating process. Subsequently, they introduce the SD-HC method for learning disentangled representations with hidden correlations. The proposed method is validated on both synthetic data and five real-world datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The major issue lies in the lack of detailed theoretical explanation:\n - The causal graph presented in Figure 3 is confusing. Since $z = f(x)$ is a function of $x$, there should be a directed edge from $x$ to $z$. However, the graph shows $z$ as a parent of $x$, which is misleading.\n - The claim that \"mode-based conditional mutual information (CMI) minimization is the necessary and sufficient condition for supervised disentanglement\" lacks a detailed formal list of necessary assumptions. This claim heavily depends on the data-generating process, and the specific assumptions should be clearly highlighted.\n - The link between Proposition 2 and the independence requirement in Definition 2 is unclear. Proposition 2 appears to be a mutual information result that is not directly related to causality. Although the authors attempt to bridge this gap in Lines 242 to 244, a more rigorous analysis with detailed assumptions and proofs is needed.\n2. The mode label estimation step is confusing. In this step, mode labels are estimated by clustering on the corresponding attribute. Consequently, there would be no confounders that affect $m_k$ and $a_{-k}$ while not affecting $a_k$, which violates the assumptions depicted in Figures 2 and 3.\n3. The paper lacks detailed explanations regarding the choice of attributes and the modes in the real-world datasets. Providing more context would help in understanding the experimental results better.\n4. There are typos in Lines 332 and 336. \"Figure 7a\" and \"Figure 7b\" should be replaced with \"Figure 5a\" and \"Figure 5b,\" respectively." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The studied problem is open and important -- latent confounding is prevalent and proper efforts are deserved to this problem.\n2. The experiments are extensive and the proposal looks promising." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a method called Supervised Disentanglement under Hidden Correlations (SD-HC) to improve disentangled representation learning (DRL). Traditional DRL methods struggle with hidden correlations in multi-modal data. SD-HC addresses this by discovering data modes under certain attributes and minimizing mode-based conditional mutual information. This approach ensures disentanglement while preserving mode and attribute information. The paper provides theoretical proofs and empirical evidence showing SD-HC outperforms existing methods on various datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper may benefit from a clear introduction of the task, especially the instantiation with real-world examples. Although I appreciate several examples present in the paper, a coherent running example would aid the readability much more if it includes the objective of the task, the problem of the hidden correlation, and the instantiation of the proposal in this example.\n2. The graphical model Figure 2 is not well-defined for me. $z$ variables are functions of the observed data $x$. How would they appear in the true data-generating process? To me, this is not even a causal graph -- if I intervene on attribute $a_{i}$ in the true data-generating process, this influences will not be blocked by estimated variables $z$'s.\n3. The main arguments are elusive to me. Especially, what does Proposition 2 imply? Its connection to the disentanglement is unclear. To me, it states that if you can find the modality variable $m_{1}$ and condition on it, the associated estimated variable $z_{1}$ will be d-separated from other attributes $a_{-1}$. First, there is no guarantee to discover the true modality variable $m_{1}$ (the proposal is somehow heuristic); second, how would this lead to the disentanglement conclusion?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- (minor) Why renaming? The current name SD-HC sounds more like a problem setting, not a specific method." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "(Disclosure: I've reviewed this paper before, and I used a PDF comparison tool to check the differences.)\n\n- This paper studies a challenging problem in disentangled representation learning.\n- The motivating example in Figure 1 is easy to follow and highlights the need for this method. It also links to the experiments on the human activity recognition task in Section 5.\n- The symbols and technical terms are properly defined. The definitions of disentangled data generation process and disentangled representations are clearly stated in Definitions 1 and 2.\n- The paper provided sufficient empirical evidence on both image and time series data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses disentangled representation learning for data with correlated, multimodal attributes and proposes Supervised Disentanglement under Hidden Correlations (SD-HC), which infers hidden modes within attributes to achieve disentanglement while preserving essential information by minimizing mode-based conditional mutual information (CMI). The proposed approach is validated through theoretical analysis and experiments on illustrative toy data, image data (colored MNIST), and time series datasets (UCI-HAR, RealWorld, HHAR, and MFD), showing improved performance over existing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The connections between the theory and learning are still unclear to me. A better way to characterize the proposed method is to say something like \"Under x assumptions, given x supervision, when x loss is optimized or x condition is satisfied, the learned representations are disentangled in the sense of definition x\".\n- The propositions and corollaries can be written in a more self-contained manner, even if the symbols have been defined in the text above." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024supervised,\ntitle={Supervised Disentanglement Under Hidden Correlations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4HAXypZfsm},\nnote={under review}\n}" }, "abstract": { "value": "Disentangled representation learning (DRL) methods are often leveraged to improve the generalization of representations. Recent DRL methods have tried to handle attribute correlations by enforcing conditional independence based on attributes. However, the complex multi-modal data distributions and hidden correlations under attributes remain unexplored. Existing methods are theoretically shown to cause the loss of mode information under such hidden correlations. To solve this problem, we propose Supervised Disentanglement under Hidden Correlations (SD-HC), which discovers data modes under certain attributes and minimizes mode-based conditional mutual information to achieve disentanglement. Theoretically, we prove that SD-HC is sufficient for disentanglement under hidden correlations, preserving mode information and attribute information. Empirically, extensive experiments on one toy dataset and five real-world datasets demonstrate improved generalization against the state-of-the-art baselines. Codes are available at anonymous Github https://anonymous.4open.science/r/SD-HC." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Disentangled representation learning", "Supervised representation learning", "Mutual information", "Causal graph analysis", "Hidden Correlations" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/6a30d3a132ddef7f71d9506e3c7ac7382e2b9824.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Supervised Disentanglement Under Hidden Correlations" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I think it'd be important to test adaptive attacks, especially considering that the same techniques used in this paper have been shown ineffective with standard classifiers." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper extends the use of several detection techniques to vision-language models, which might be interesting as making these models robust is a relevant challenge. The question of whether detection in multi-modal models is easier than with classifier is also interesting.\n\n- The experimental evaluation includes several datasets and tasks, and different architectures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper benchmarks the effectiveness of several methods for detecting adversarial input images in multi-modal models. In particular, these techniques exploit the geometry of the features extracted by the vision and multi-modal encoders to distinguish between clean and adversarial images, and are agnostic of the downstream task. In the experimental evaluation, the different methods are compared on several datasets for both classification and retrieval tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Detection methods for image classification have often been shown ineffective against adaptive attacks: for example, LID was bypassed in [A], Mahalanobis distance was shown non-robust to adversarial perturbations in [B], and several detection methods are bypassed in [C]. Thus, the proposed methods should be tested against attacks targeting the detection mechanism, e.g. as discussed in [C]. Moreover, the ablation study in Sec. 5.3 about generalization to different attacks only uses methods which are very close to the PGD attack (with only 10 steps) used for tuning the detection methods, and optimize the same loss function: then the generalization to them is not surprising.\n\n- The techniques used for detection are from prior works, which limits the technical novelty of the paper.\n\n- I think it'd be useful to report the success rate of the attack before detection.\n\n[A] https://arxiv.org/abs/1802.00420 \\\n[B] https://arxiv.org/abs/2201.07012 \\\n[C] https://openreview.net/forum?id=af1eUDdUVz" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "This paper should consider further feedback regarding the lack of rigorous mathematical proof, computational scalability concerns, and vulnerability to adaptive attacks." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper is easy to follow, and the structure is well-claimed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes GAD-VLP, a geometric adversarial detection framework for vision-language pre-trained models. The method leverages geometric approaches including local intrinsic dimensionality, k-nearest neighbors distance, Mahalanobis distance, and kernel density estimation to identify adversarial examples in VLPs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper lacks rigorous theoretical analysis of why geometric methods work better for VLPs compared to traditional models. While empirical results are shown, there's no formal proof or theoretical guarantees about the method's effectiveness, especially claiming why their approach is fundamentally sound for VLPs.\n\n2. The paper doesn't adequately address the computational costs of calculating geometric metrics, especially for large-scale deployments. Computing k-NN, Mahalanobis distances, and KDE for high-dimensional embeddings can be computationally expensive. \n\n3. The paper doesn't consider sophisticated adaptive attacks that might specifically target the geometric detection mechanism. Adversarial methods often adapt to known defense mechanisms, and the lack of evaluation against such adaptive attacks raises questions about the method's robustness in real-world scenarios. \n\n4. The authors don't thoroughly examine the false positive rates and their impact on model usability. In real-world applications, high false positive rates could lead to unnecessary rejection of legitimate inputs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- What is the setup for retrieval is it top-1, top-5? \nNo other questions, some questions on choice of experiments can be inferred from Weaknesses section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The problem being studied is important - Detecting adversarial inputs in multi-modal setup.\n- The related work exploration seems sufficient.\n- Diversity of evaluation tasks (classification, retrieval) and models is reasonable" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper looks at detection of adversarial examples for multi-modal models (CLIP, ALBEF). They extend uni-modal detection methods to vision-language based models. Experiments are conducted on classification and Image-retrieval tasks using different 'geometry of sample' based approaches to show GAD-VLP (proposed framework) works well when using different methods for detection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Overall the paper lacks technical novelty as previously used methods (MMD, KDE etc) for uni-modal detection methods are just transferred to multi-modal setup.\n\n- The evaluation/testing setup in re to adversarial setup is not sufficient.\n\n- The adversarial attacks do not seem strong, looking at Fig. 2 a naturally strong attack would be to add a regularizer term that enforces the image embedding (under perturbation) to stay close to the clean embedding. Testing method on such strong attacks would be nice.\n\n- The attacks are based on10-step PGD, and versions of FGSM all at eps=8/255 (other values should have been looked at). A lot of new attacks (see the ones in [1, 2]) for CLIP like models have been proposed - testing which would have been also a valuable addition to the submission.\n\n- For the binary classifier using LID, attacking the classifier (detector) would also be a reasonably strong attack.\n\n[1] Schlarmann, Christian, and Matthias Hein. \"On the adversarial robustness of multi-modal foundation models.\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.\n\n[2] Mao, Chengzhi, et al. \"Understanding Zero-shot Adversarial Robustness for Large-Scale Models.\" The Eleventh International Conference on Learning Representations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Please see the weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1.\tThe writing is clear. The formulas are correct.\n2.\tThe experiment is multi-dimensional.\n3.\tThe research topic is important for VLP." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose the GAD-VLP, a method for detecting adversarial attacks in VLP Models. It uses geometric metrics like Local Intrinsic Dimensionality (LID), k-Nearest Neighbors (k-NN), Mahalanobis distance, and Kernel Density Estimation (KDE) to distinguish adversarial examples. GAD-VLP works across various VLP architectures, whether they use separate or combined embeddings for vision and language inputs. The method demonstrates high detection accuracy against different attacks and is applicable to tasks like zero-shot classification and image-text retrieval. The study highlights GAD-VLP's robustness and versatility across multiple domains." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tWhile the proposed method is effective, the metrics used are standard and not specifically tailored for the Vision-Language Pre-training (VLP) task.\n2.\tThe paper does not sufficiently highlight the distinctions between unimodal models and VLP models, resulting in a lack of justification for the choice of metrics.\n3.\tClassical VLP models, such as BLIP [1] and X-VLM [2], are missing from the analysis.\n4.\tThe adversarial attack methods utilized are limited and do not include more popular approaches like Set-level Guidance Attack (SGA) [3].\n\n[1] BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation\n\n[2] Multi-Grained Vision Language Pre-Training: Aligning Texts with Visual Concepts\n\n[3] Set-level Guidance Attack: Boosting Adversarial Transferability of Vision-Language Pre-training Models" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024gadvlp,\ntitle={{GAD}-{VLP}: Geometric Adversarial Detection for Vision-Language Pre-Trained Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4HL2aiDV97},\nnote={under review}\n}" }, "abstract": { "value": "Vision-language pre-trained models (VLPs) have been deployed in numerous real-world applications; however, these models are vulnerable to adversarial attacks. Existing adversarial detection methods have shown their efficacy in single-modality settings (either vision or language), while their performance on VLPs, as multimodal models, remains uncertain. In this work, we propose a novel aspect of adversarial detection called GAD-VLP, which detects adversarial examples by exploiting vision and joint vision-language embeddings within VLP architectures. We leverage the geometry of the embedding space and demonstrate the unique characteristics of adversarial regions within these models. We explore the embedding space of the vision modality or the combined vision-language modalities, depending on the type of VLP, to identify adversarial examples. Some of the geometric methods do not require explicit knowledge of the adversary's targets in downstream tasks (e.g., zero-shot classification or image-text retrieval), offering a model-agnostic detection framework applicable across VLPs. Despite its simplicity, we demonstrate that these methods deliver a nearly perfect detection rate on state-of-the-art adversarial attacks against VLPs, including both separate and combined attacks on the vision and joint modalities." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Adversarial detection", "Geometric Distance", "Multimodal models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/877092dc5c513fa23e4f52d99216fea29577425c.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "GAD-VLP: Geometric Adversarial Detection for Vision-Language Pre-Trained Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- Why do the segment encoder q, decoder p, and prior f all use the same notation, θ, for their parameters?\n- How does the distribution mismatch between Dp and Du cause problems, and how does HPL address this issue (through improved credit assignment) in the main benchmark experiments? Is there any analysis on these tasks other than the example of gambling MDP?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper introduces a framework for hindsight preference learning, with a practical implementation using a VAE architecture.\n- I think using the example of a gambling MDP greatly helps in understanding the motivation for using hindsight preference learning.\n- Experimental results support the effectiveness of HPL." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new preference learning method that utilizes hindsight information. By incorporating future states into the reward through a VAE structure, HPL can generate more robust rewards and overcome the limitations of Markovian rewards." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Although I understand the benefits of hindsight preference learning over Markovian rewards, the authors do not clearly position their paper within the broader research on non-Markovian rewards. As mentioned in the related work, both the Preference Transformer and Hindsight PRIOR assume non-Markovian rewards and consider the full trajectory when calculating rewards. Given this, why does the paper primarily motivate its approach as an improvement over Markovian rewards, despite the existence of several approaches that do not rely on this assumption?\n- For example, what is the fundamental difference between this paper and the Hindsight PRIOR paper, as Hindsight PRIOR considers “the importance of a state in a trajectory given that a future state was reached” [1]? Please clarify the novelty of this work in comparison to other hindsight-based PbRL papers.\n- The gambling example is very helpful in illustrating the motivation. However, I am concerned that in more stochastic settings, hindsight modeling may introduce noise and add excessive complexity. The ablation study on k, while focusing on a different message, seems to somewhat support this concern.\n\nReferences \n\n[1] Verma, M., & Metcalf, K. (2024). Hindsight PRIORs for Reward Learning from Human Preferences." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The preference signals on the MetaWorld platform do not actually take into account future information. Why is the algorithm proposed in this paper considered superior?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "**Originality**: The paper studies a very important problem in PbRL: the nature of human annotations, which is inherently more complex than scripted annotations. By proposing a VAE structure with marginalization techniques, this paper a novel and interesting method.\n\n\n**Quality**: Experiments are conducted on existing human-annotated dataset where the proposed method shows greater performance.\n\n\n**Clarity**: The paper is well written with illustrative examples, clear text and figures.\n\n\n**Significance**: The studied problem: the mismatch between annotations and chosen reward models is of very importance and shall be beneficial for the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies preference-based reinforcement learning (PbRL). Notably, it proposes to consider human annotations from a holistic perspective; that is the annotations would include the considerations of future outcome instead of in-trajectory information only. Inspired by this mismatch, authors propose to model human preference with hindsight information via a VAE structure, which is further marginalized for rewarding labeling and following RL downstream tasks. Experiments are conducted on annotations from human and scripted teacher and advantageous performance is demonstrated." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The primary concern lies in the **mismatch** between motivation and downstream methods. \n\nThe paper aims to incorporate additional future information when learning the $r_{\\psi}$. However, actually, the $r_{\\psi}$ that takes future information into account should be the Q-function rather than the reward function. Or more rigorously, the $r_{\\psi}$ actually encompasses far more information than reward function, but it is not necessarily a Q-function. When downstream algorithms utilize the $r_{\\psi}$, the $r_{\\psi}$ is considered as reward function, employing a method that does not consider future information." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In addition to questions about potential weaknesses, I would like to raise a few more questions:\n- When computing the reward using the prior distribution, how does the effect depend on the number of samples? (N=20 in appendix)\n- What about using both preference data and unlabeled data when training the VAE? Wouldn't this improve the representation across multiple points?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The problem to be solved is clearly specified, and although the approach is simple, it effectively addresses aspects that previous papers have overlooked.\n- In the experiments, HPL outperforms the existing SOTA algorithms on most datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors study the problem of offline preference-based RL. One major problem with existing SOTA approaches, as cited by the authors, is the evaluation of trajectory segments from a global perspective, making the obtained reward implicitly dependent on the future part of the segment. The contributions of the paper can be summarized as follows:\n- The authors propose HPL, which models human preferences using a reward function conditioned on the state, action, and future trajectory, referred to as hindsight information.\n- HPL leverages the trajectory distribution of the unlabeled dataset to learn a prior over future outcomes, providing robust and advantageous rewards during labeling." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I understand that the use of a gambling MDP in Section 3.1 highlights the importance of this study. However, given the very short trajectory length (two steps) in the example, it remains unclear whether hindsight information was effectively applied to achieve accurate credit assignment using future trajectory information.\n- Figure 10 shows that HPL is already 2 to 10 times slower than other models. There seem to be several reasons for this slower speed. \n - Equation 8 requires summing all steps of each trajectory, which likely takes considerable time.\n - A significant amount of time is spent on sampling in Equation 9.\n - If the running time includes VAE modeling time, it is expected to take even longer.\n- The core idea of HPL seems to lie not in using future trajectory to train the model, but rather in the learned VAE that represents the future trajectory. As the authors mentioned, when the future length exceeds a certain threshold, it becomes difficult to accurately represent longer future segments, which lowers HPL's performance. If future information is available, the performance should ideally be more accurate. How might this limitation be addressed?\n\nMinor comments:\n\n- typo on line 403: ''$\\hat{P}(s_{good}|s_{1},a_{1})$ and $\\hat{P}(s_{good}|s_{1},a_{1})$\" -> \"$\\hat{P}(s_{good}|s_{1},a_{1})$ and $\\hat{P}(s_{bad}|s_{1},a_{1})$\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. The experiments utilize a customized split of the original dataset, using medium-expert quality for the preference dataset and medium quality for the unlabeled dataset. Has there been any evaluation of preference shift on real-world, large-scale datasets?\n2. Equation 9 includes the number of samples N in the reward function. How does N influence performance, and what is the practical range for this parameter?\n3. The concept of utilizing future segments appears similar to graph search and retrieval-based offline RL methods [1,2]. Could HPL achieve comparable results?\n[1] Deyao Zhu, Li Erran Li, and Mohamed Elhoseiny. Value memory graph: A graph-structured world model for offline reinforcement learning. arXiv preprint arXiv:2206.04384, 2022. \n[2] Yin Z H, Abbeel P. Offline Imitation Learning Through Graph Search and Retrieval. arXiv preprint arXiv:2407.15403, 2024." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The paper addresses a critical challenge in offline preference-based reinforcement learning, presenting well-motivated arguments and clearly illustrating the reward modeling issue through compelling examples.\n2. The authors make the insightful observation that human preferences often stem from final outcomes. Their proposed hindsight preference learning leverages future segments to better model these preferences, resulting in a more robust reward function, particularly for datasets with outcome shifts due to multiple modalities or approaches. The integration of conditional VAE makes the algorithm practical for complex tasks like visuomotor policy learning.\n3. The experimental evaluation strongly supports the core ideas, encompassing multiple popular benchmarks and competitive baselines. The comprehensive ablation studies examining future segment length, dataset sizes, and reward ensemble provide valuable insights into HPL's advantages." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Hindsight Preference Learning (HPL), a novel reward model learning method for offline preference-based reinforcement learning. The authors address the issue of distribution shift between preference-labeled and unlabeled datasets, which can bias the Markovian reward assumption. To tackle this problem, they propose a reward function that is conditioned not only on states and actions but also on future segments. To handle the high-dimensional nature of future segments, the authors employ a conditional VAE to transform them into embedding vectors that better integrate with the reward function. The effectiveness of the method is demonstrated through experiments on Gym-MuJoCo, Adroit, and Meta-World benchmarks, comparing against existing preference-based reinforcement learning algorithms. The results show significant improvements under preference distribution shift, while ablation studies validate the effectiveness of both the VAE and hindsight reward function." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The VAE implementation is limited to encoding future segments, whereas it could potentially be extended to data augmentation, prior distribution modeling, and other applications.\n2. While the authors focus on \"preference distribution shift\" between labeled and unlabeled datasets, they assume stability in other distributions. The paper does not adequately address potential shifts in visual appearance, dynamics, and sensor noise - common challenges in real-world datasets." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is clearly written, easy to understand, and provides detailed, well-functioning code.\n2. The proposed VAE method incorporates hindsight future information from trajectories into the reward model, which effectively mitigates the reward learning issue caused by the distributional shift between preference data $D_p$ and offline data $D_u$.\n3. The experimental performance is impressive, and the ablation studies are thorough." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Hindsight Preference Learning (HPL), a novel method for offline PbRL. It addresses the shortcomings of existing methods that rely on Markovian reward assumptions, by incorporating future trajectory outcomes in reward learning. HPL uses a variational auto-encoder (VAE) to encode future information and improves credit assignment by considering the future consequences of actions. The authors provide extensive empirical results across a variety of tasks (locomotion and manipulation) and demonstrate that HPL outperforms several state-of-the-art PbRL methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation example might be unreasonable: As far as I know, in the same task, $D_u$ is often sampled from $D_p$, so they follow the same marginal distribution. When such an extreme example occurs, it indicates a significant distribution difference between $D_u$ and $D_p$. Although Section 5.3 demonstrates the effectiveness of HPL under distributional shift, this example may lack generality.\n2. In fact, this method can be viewed as leveraging the unlabeled dataset to improve the robustness of the reward model. The idea is somewhat similar to previous work [1], which learns the reward model from labeled data and then labels the unlabeled data to enhance offline RL performance. A further theoretical analysis, such as exploring the theoretical performance bounds of learning the hidden dynamics from the unlabeled data in the reward model, would enrich the paper (though this could be very challenging, so it's just a suggestion).\n3. In Table 7, the performance of OPPO and CPL is quite poor. Is this mainly because the preference data $D_p$ is too small?\n4. In Figure 6c, reducing the size of $D_p$ leads to worse performance. Why is that?\n5. The baselines and ablations lack verification of the core idea of the paper: If the authors aim to verify that the performance improvement comes from the richer environment dynamics provided by the unlabeled data, they should compare with an ablation version where the VAE learns future representations from $D_p$ itself rather than from $D_u$, to further clarify the source of the performance gain. Additionally, in Table 1, a bi-directional transformer version of PT could be used to model future information as another baseline for comparison.\n[1] The Provable Benefits of Unsupervised Data Sharing for Offline Reinforcement Learning. ICLR23." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose to model the effect of future to achieve better credit assignment when learning from human preferences." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024hindsight,\ntitle={Hindsight Preference Learning for Offline Preference-based Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4HNfKrGlSJ},\nnote={under review}\n}" }, "abstract": { "value": "Offline preference-based reinforcement learning (RL), which focuses on optimizing policies using human preferences between pairs of trajectory segments selected from an offline dataset, has emerged as a practical avenue for RL applications. Existing works rely on extracting step-wise reward signals from trajectory-wise preference annotations, assuming that preferences correlate with the cumulative Markovian rewards. However, such methods fail to capture the holistic perspective of data annotation: Humans often assess the desirability of a sequence of actions by considering the overall outcome rather than the immediate rewards. To address this challenge, we propose to model human preferences using rewards conditioned on future outcomes of the trajectory segments, i.e. the hindsight information. For downstream RL optimization, the reward of each step is calculated by marginalizing over possible future outcomes, the distribution of which is approximated by a variational auto-encoder trained using the offline dataset. Our proposed method, Hindsight Preference Learning (HPL), can facilitate credit assignment by taking full advantage of vast trajectory data available in massive unlabeled datasets. Comprehensive empirical studies demonstrate the benefits of HPL in delivering robust and advantageous rewards across various domains." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Offline Reinforcement Learning", "Preference-based Reinforcement Learning", "Preference Model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/507f77a4fb6dc4609994e8e7fdf79ab6a700f24d.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/f1c7de91e29cde278d50dba428a67c78a15b8e46.zip" }, "title": { "value": "Hindsight Preference Learning for Offline Preference-based Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- In Line 280, is this a typo on $g(\\theta)$? Looks like this is a missing $\\Delta$." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The proposed normal deflection fields, which correct inaccurately estimated normals from normal maps, seem reasonable to me. \n- Table 6 and Figure 5 effectively showcase the contribution of the method's design through comprehensive ablation studies.\n- The experimental results demonstrate that this proposed pipeline achieves state-of-the-art reconstruction quality in indoor scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a pipeline for multi-view 3D reconstruction of indoor scenes. The pipeline is based on recently popular differentiable volume rendering methods using Signed Distance Functions (SDF), such as VolSDF and NeuS. To enhance reconstruction quality in indoor scenes, particularly in textureless regions, the paper introduces monocular priors for additional supervision, inspired by recent works like MonoSDF and NeuRIS.\n\nA core contribution of this paper is the introduction of a normal deflection field, which helps the model identify and correct inaccuracies in reference normal images estimated from monocular models, such as the Omnidata model. Experimental results show that this deflection field aids in reconstructing fine geometric details, such as thin structures.\n\nAdditionally, the authors introduce techniques to further improve geometry reconstruction quality, including (1) an adaptive normal loss weight and (2) unbiased volume rendering on thin objects.\n\nThe experimental results are promising and include comprehensive ablation studies. Overall, the paper is well-argued and substantiated." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Using unbiased volume rendering for thin structures seems reasonable. However, why not apply this unbiased volume rendering weight uniformly across all regions? The authors mention potential convergence issues with this approach, but it would be helpful if they provided a more in-depth analysis explaining the nature of these convergence issues and offered insights into why they occur." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. would it make sense to consider depth maps together to decide if the normal prior is accurate or not? For example, in most complex geometry areas such as thin structure areas, the depth is not continuous, and one may not be able to hope the normal is accurate there. \n2. what is the reason for the artifacts (which I mentioned in weakness 4), is that because without supervision the model struggles to recover the details?\n3. to deal with 2, maybe add some smooth loss / geometric preserving loss on these areas that the normal prior is categorized as bad? \n4. I don't see a reason why this method only works indoor, has the author also tried it on outdoor datasets?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-written and easy to follow. \n2. The normal deflection detection design is under reasonable assumptions and it offers a solution to deal with non-accurate normal prior. \n3. The angle deviation loss design fits the situation. \n4. The author provides complete experiments to show their method's performance. Including comparison with previous work quantitatively and qualitatively and ablation studies to show each module improves the results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a method to improve prior normal-guided neural SDF indoor scene reconstruction. Instead of supervising the normal uniformly across the whole domain, the author proposes to use a network to learn the deflect normal and desired adaptive angle prior to loss according to the deviation of the deflect normal to the prior normal. The intuition is that large deviation happens in complex structure areas where the normal prior in general is bad quality. The author also proposes to use the re-weighted loss on normal, depth, and color rendering. The experiments show the proposed method performs well qualitatively and quantitatively." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The bad quality normal prior problem seems only solved half. The author reduced the loss weight on the part that the normal prior is not accurate, but not the other half, what to do to improve the method itself performance in these areas.\n2. Like the other prior-based methods, the quality of the prior matters. The author assumes smooth area the prior is relatively accurate. \n3. Like the other SDF-based methods, the reconstruction seems to still struggle with the thin structures. \n4. I noticed that in most cases the method outperforms previous methods, but sometimes it creates noise and artifacts." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The inaccuracy and bias are widely existed in monocular depth and normal priors, therefore, the idea of exploring the uncertainty of monocular clues to improve the reconstruction performance is reasonable and valuable.\n 2. The normal deflection field is effectively incorporated throughout the training process, enabling rendered deflected normal images to be supervised by estimated normals and adaptively applied in ray sampling, photometric loss, and unbiased rendering.\n3. The numerical and visualization results are attractive and impressive, which highly demonstrate the effectiveness of the proposed method.\n4. The ablation studies are thorough, clearly showcasing the effectiveness of each proposed module." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes ND-SDF, which learns a normal deflection field to reduce the significant bias in monocular normal and depth priors. The normal deflection field can be rendered through volume rendering to further guide adaptive prior loss, ray sampling and unbiased rendering. The numerical and visualization results are compelling." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of citation. A similar approach of learning rotated normals to mitigate the bias between normal priors and ground truth normals was first proposed in NC-SDF[1]. While I know that NC-SDF has not released the source code so it’s difficult to directly compare with it, an appropriate clarification is necessary to distinguish the proposed idea from NC-SDF.\n2. Lack of quantitative comparisons with DebSDF. DebSDF also predicts the uncertainty of monocular priors and is able to reconstruct details such as chair legs. The numerical result of DebSDF is included in the table but visual comparison with DebSDF is absent, which is insufficient.\n\nThe methodology and experiments in this paper are solid. I like this paper, however, the two major concerns above prevent me from giving a higher score. I am willing to raise my score if the concerns are well addressed.\n\n[1]. Chen Z, Wu X, Zhang Y. NC-SDF: Enhancing Indoor Scene Reconstruction Using Neural SDFs with View-Dependent Normal Compensation. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 5155-5165." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. For Figure 3, what about the comparison with unbiased density function in NeuS, TUVR, DebSDF, HF-NeuS[3], and NeuRodin(optional, because NeuRodin is the concurrent work)[4]? These works also propose an unbiased density strategy. (visual comparison is better for good understanding)\n\n2. In Figure 7, The deflection angle map appears to have a higher value at complex structural regions. What's the visual comparison between edge mask(EaNeuS), NCC mask(Equation 6 in NeuRIS), and uncertainty-aware mask(DDPNeRF[5], DebSDF, H2OSDF)?\n\n3. For Equation 12 of this paper, it seems that when $\\hat{\\mathbf{N}}^d(\\mathbf{r})$ and $\\hat{\\mathbf{N}}(\\mathbf{r})$ perfectly fits the prior $\\mathbf{N}(\\mathbf{r})$, then this loss will become a perfect zero. The reviewer is concerned that this situation that causes zero loss doesn't seem to be in a good state. Is there a more comprehensive explanation about this training objective and why this training objective will work to help solve the concerns above?\n\n4. Does the proposed normal deflection field have some advantages over the normal uncertainty field, like in DebSDF and H2OSDF?\n\nReference:\n\n[3] HF-NeuS: Improved Surface Reconstruction Using High-Frequency Details. Wang et al. (Neurips 2022)\n\n[4] NeuRodin: A Two-stage Framework for High-Fidelity Neural Surface Reconstruction. Wang et al. (Neurips 2024)\n\n[5] Dense Depth Priors for Neural Radiance Fields from Sparse Input Views. Roessle et al. (CVPR 2022)" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper writing is easy to follow.\n\n2. The experiment covers lots of different dataset for a comprehensive comparison, like Scannet, TanksandTemples, Scannet++, Replica.\n\n3. The ablation study indicates the effectiveness of the deflection field and adaptive angle prior when compared to its base model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces ND-SDF and proposes to learn a Normal Deflection field to represent the angular deviation between the scene normal and the normal prior.\n\nThe paper also additionally introduces a ray sampling strategy based on the deflection angle to improve the surface quality further. The paper also proposes to use an unbiased function inspired by TUVR." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major weakness:\n\n1. Although the dataset included is various, the compared methods are not so much comprehensive(except the Scannet). The quantitative comparison on TanksandTemples, Scannet++, Replica only includes three works in 2022(NeuRIS, MonoSDF, VolSDF) and one work in 2021(Unisurf). This leads to an insufficient experiment. For example, what about the comparison with HelixSurf, TUVR, DebSDF, H2OSDF on TanksandTemples, Scannet++, Replica? The reviewer thinks the comparison without newer baselines is insufficient on TanksandTemples, Scannet++, Replica. Although Table 1 includes most newer baselines, usually only 4 scenes in Scannet are used for indoor scene reconstruction experiments. This is why the reviewer thinks that more quantitative experiments are required here.\n\nDetails: The reviewer understands that it's not necessary to run all previous baselines in all datasets. But just place here as an example. Since the unbiased weight technique is used, the reviewer thinks **at least** TUVR(CVPR2023) is a newer baseline that should be included in both the quantitative and qualitative comparisons on TanksandTemples, Scannet++, Replica datasets. (But more comparison with other newer methods will enhance the paper's experiment part.)\n\n2. The qualitative comparison is also insufficient (only compared with older baselines). It's better to include the qualitative comparison with newer baselines in Figure 4, Figure 11, Figure 12, Figure 13.\n\n3. Some works have both pure MLP version and hash grid version (like MonoSDF). The performance and efficiency of different versions have some differences. Since ND-SDF includes instant-NGP in its implementation, it's better to clearly indicate the version to compare within the table. (For example, \"MonoSDF(mlp)\", \"MonoSDF(grid)\".)\n\nMinor weakness:\n\n4. For Section 3.4, similar ideas of better ray sampling strategy have been investigated in some previous work. Like edge-aware sampling in EaNeuS[1], GaussianRoom[2], and uncertainty-aware sampling in DebSDF. Please consider adding a brief discussion about these previous methods.\n\n5. In Figure 2, \"GT depth\" and \"GT Normal\" should be changed to \"depth prior\" and \"normal prior\".\n\n6. The number of selected scenes in each dataset for the experiment has not been reported.\n\n7. The dataset conducted in the ablation study has not been reported.\n\nReference:\n\n[1] Edge-aware Neural Implicit Surface Reconstruction. Li et al. (ICME 2023)\n\n[2] GaussianRoom: Improving 3D Gaussian Splatting with SDF Guidance and Monocular Cues for Indoor Scene Reconstruction. Xiang et al. (arxiv 2024)" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024ndsdf,\ntitle={{ND}-{SDF}: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4HRRcqE9SU},\nnote={under review}\n}" }, "abstract": { "value": "Neural implicit reconstruction via volume rendering has demonstrated its effectiveness in recovering dense 3D surfaces. However, it is non-trivial to simultaneously recover meticulous geometry and preserve smoothness across regions with differing characteristics. To address this issue, previous methods typically employ geometric priors, which are often constrained by the performance of the prior models. In this paper, we propose ND-SDF, which learns a Normal Deflection field to represent the angular deviation between the scene normal and the prior normal. Unlike previous methods that uniformly apply geometric priors on all samples, introducing significant bias in accuracy, our proposed normal deflection field dynamically learns and adapts the utilization of samples based on their specific characteristics, thereby improving both the accuracy and effectiveness of the model. Our method not only obtains smooth weakly textured regions such as walls and floors but also preserves the geometric details of complex structures. In addition, we introduce a novel ray sampling strategy based on the deflection angle to facilitate the unbiased rendering process, which significantly improves the quality and accuracy of intricate surfaces, especially on thin structures. Consistent improvements on various challenging datasets demonstrate the superiority of our method." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Normal Deflection Fields", "High-Fidelity Indoor Reconstruction" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/98b35d836ddaea3b2dbfe627cd9ca78db6ed06ae.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/61b33d9dfb7ccb6053e2840a710eca961f72fa28.zip" }, "title": { "value": "ND-SDF: Learning Normal Deflection Fields for High-Fidelity Indoor Reconstruction" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Do the mean-max aggregation lines in table 1 use optimal solutions in training as well? I find it a little strange that max aggregation performs better than the algorithm without the optimal solutions. Shouldn't the problem-specific aggregations do better?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- If I understand correctly, this primal dual approach will necessarily yield a certificate of solution quality since the dual objective bounds the primal, which is not easy to obtain with neural nets on those problems.\n- Ablations and experiments on real world data are helpful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes to simulate the primal-dual framework for approximation algorithms by Goemans and Williamson using a graph neural network. The paper follows the standard neural algorithmic reasoning blueprint with an encoder-processor-decoder approach and it leverages a gnn that performs message passing on a bipartite graph representation of the problem. The model is used to solve NP-hard problems such as minimum vertex cover, minimum hitting set, and minimum set cover. The model also leverages optimal solutions to the problem as additional supervision and is tested on real and synthetic graphs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concerns are the following:\n- The experimental comparisons are inadequate. For example, MVC (and its complement problem of max independent set) have been tackled by several works in the literature (examples can be found in [1,2,3]). Warm-starting solvers has also been done (e.g., see 4). In my view, discussion and comparisons with some work from the literature on the core experiments as well as the warm start ones would help clarify where the model stands overall. Furthermore, the experiments on synthetic instances are rather small-scale (especially table 1). [2,3] provide an experimental setting with hard synthetic instances of larger sizes for MVC. It would be good to see how the model performs against some of those baselines on those instances.\n- On the topic of scale, it would be good to see what the performance looks like if the synthetic instances were scaled up an order of magnitude. Right now, it seems like it would take a lot of supervision to train a scalable version of this algorithm, especially if one wanted to leverage optimal solutions, which become costly for NP-hard problems as the instance size grows.\n- A central goal of this paper is to have a machine-learning model that closely aligns with the primal-dual framework. The numbers seem to agree in table 1 when it comes to MVC and MHS but for MSC we see that not having access to the optimal solutions seems to lead to results worse than the primal-dual algorithm itself. This points to the broader issue that even though the model is designed to closely align with this framework, the approximation guarantee is not secured in practice.\n\nOverall, this is an interesting approach to neural combinatorial optimization. However, I find the empirical results unconvincing at this stage and ultimately the algorithm does not seem to preserve the approximation guarantee in practice so I am unsure about the contribution as a whole. At this stage. I cannot recommend acceptance but I am open to reconsidering after the rebuttal.\n\n\n\n\n1. Khalil, Elias, et al. \"Learning combinatorial optimization algorithms over graphs.\" Advances in neural information processing systems 30 (2017).\n2. Brusca, Lorenzo, et al. \"Maximum independent set: self-training through dynamic programming.\" Advances in Neural Information Processing Systems 36 (2023)\n3. Yau, Morris, et al. \"Are Graph Neural Networks Optimal Approximation Algorithms?.\" arXiv preprint arXiv:2310.00526 (2023).\n4. Benidis, Konstantinos, et al. \"Solving recurrent MIPs with semi-supervised graph neural networks.\" arXiv preprint arXiv:2302.11992 (2023)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to \"Weakness\"." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper proposes an approach that is the first NAR method to surpass the learned algorithm.\n2. Author conducted several experiments with good numerical results favoring the proposed framework.\n3. Theoretical proofs are given to validate the effectiveness of the proposed framework." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an approach that integrates the primal-dual method with GNNs to solve NP-hard combinatorial optimization problems more efficiently. \nIt establishes theoretical guarantees that the proposed framework can replicate the classical primal-dual method. \nBy refining the optimal solutions for these problems, the model surpasses the performance of conventional algorithms. \nAdditionally, by utilizing predictions as warm starts, the search time for commercial solvers can be reduced.\nNumerical results showcase the scalability and effectiveness of the proposed framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the proposed framework is designed to address NP-hard combinatorial optimization problems, the proposed framework primarily operates on linear programming (LP) relaxation, which appears to be an extension to existing methods, such as those referenced in [1] and [2]. \nBesides, there are also many works that directly predicts solutions to CO problems. e.g. [3] and [4].\nAn extended discussion, especially with experiments, on how the proposed approach differs from them would strengthen the authors' claims.\n\n[1]. Bingheng Li, Linxin Yang, Yupeng Chen, Senmiao Wang, Qian Chen, Haitao Mao, Yao Ma, Akang Wang, Tian Ding, Jiliang Tang, and Ruoyu Sun. Pdhg-unrolled learning-to-optimize method for large-scale linear programming, 2024.\n\n[2]. Ziang Chen, Jialin Liu, Xinshang Wang, Jianfeng Lu, and Wotao Yin. On representing linear programs by graph neural networks, 2023.\n\n[3]. Vinod Nair, Sergey Bartunov, Felix Gimeno, Ingrid von Glehn, Pawel Lichocki, Ivan Lobov, Brendan O’Donoghue, Nicolas Sonnerat, Christian Tjandraatmadja, Pengming Wang, Ravichandra Addanki, Tharindi Hapuarachchi, Thomas Keck, James Keeling, Pushmeet Kohli, Ira Ktena, Yujia Li, Oriol Vinyals, and Yori Zwols. Solving mixed integer programs using neural networks, 2021.\n\n[4]. Dinghuai Zhang, Hanjun Dai, Nikolay Malkin, Aaron Courville, Yoshua Bengio, and Ling Pan. Let the flows tell: Solving graph combinatorial optimization problems with gflownets, 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- The three problems are closely related. Can the method be applied on other problems that can work with the primal-dual algorithm?\n\n- In table 2, why use geometric mean for total time while arithmetic mean for optimal time?\n\n- what is the relation of PDGNN and GAT, GCN and SAGE? How are they combined as shown in Table 4? What are the inputs and outputs of each module? Since the main text does not mention how are PDGNN works with other models, the results confused me.\n\n- what does the claimed OOD generalizability comes from? There is no module designed for the purpose. So it confuses me why the method has good OOD generalizability.\n\nSince I am not an expert in these problems, I would consider raising my score based on the rebuttal." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Writing is good with many details. The method looks sound. Experiments are comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper design a GNN model based on the primal-dual algorithm to solve CO problems including Minimum Vertex Cover, Minimum Set Cover, and Minimum hitting set. The retrained model can be used to warm start a commercial solvers, and also applicable to other problems closely connected with the studied problems, eg node classification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- To me, it is not clear whether the work is important in literature, especially in dealing with the three problems.\n\n- Important baselines are missing. For example, other neural solvers for the three problems are not provided. GIN is not designed with the purpose of solving the problem. So the comparison with merely GIN can be unfair. To demonstrate the effectiveness, it is necessary to include more solvers for the problem.\n\n- Authors give results of model performance on OOD data in Table 2. However, since authors do not give results of other baselines, the OOD generalizability cannot be proved directly. By the way, no OOD results on the minimum set cover problem are provided." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Refer to \"Weaknesses\".\n\n2. In the experiments, how many layers does PDGNN has?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The observation that the primal-dual algorithm naturally align with message-passing GNN is interesting. The integer programming studied in this paper is a quite general problem. The writing is clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose a new GNN architecture, termed PDGNN, for solving the integer programming $\\min \\vec{w}^T\\cdot\\vec{x}$ s.t. $A\\vec{x}\\geq \\vec{1}$ and $x_i\\in \\{0,1\\}$, which can model many combinatorial optimization problems, such as set cover and hitting set. This GNN is designed by unrolling a primal-dual approximation algorithm. Numerical experiments are conducted to evaluate the performance of PDGNN." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. About Section 1.1: The phenomenon that GNNs designed by unrolling an algorithm surpass the performance of the algorithm itself is not surprising (since the algorithm is an instantiation of the GNN, so the GNN has enough expressive power to surpass). In fact, the phenomenon is general: see. e.g. \"PDHG-unrolled learning-to-optimize method for large-scale linear programming\" in ICML 2024. As in\n real-world scenarios, PDGNN will commonly be used as a warm start, so the paper could highlight how much computational time can be saved by employing PDGNN as a warm-start.\n\n2. As shown in Table 3, by employing PDGNN as a warm-start, the commercial solver achieve a $\\leq 1.1\\times$ speedup compared to the vanilla version. The improvement did not impress me, as I had been expecting a $\\geq 2\\times$ speedup.\n\n3. The paper proves that PDGNN can exactly replicates the primal-dual algorithm, but lacks explicit theoretical guarantees on the efficiency and effectiveness of PDGNN. For example, I would like to see theorems like \"PDGNN of xx layers has enough expressive power to output an xx-approximate solution for xxx problems\"，or discussions on the challenges in deriving such theoretical guarantees or intuitions about what kind of guarantees might be possible given the current framework." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a general GNN-based framework using primal-dual approximation algorithms to approach NP-hard combinatorial optimization problems, advancing neural algorithmic reasoning beyond the polynomial-time-solvable domain." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024primaldual,\ntitle={Primal-Dual Graph Neural Networks for General {NP}-Hard Combinatorial Optimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4Hd7u3LHlZ},\nnote={under review}\n}" }, "abstract": { "value": "Neural algorithmic reasoning (NAR) seeks to train neural networks, particularly Graph Neural Networks (GNNs), to simulate and generalize traditional algorithms, enabling them to perform structured reasoning on complex data. Previous research has primarily focused on polynomial-time-solvable algorithmic problems. However, many of the most critical optimization problems in practice are NP-hard, exposing a critical gap in NAR. In this work, we propose a general GNN-based framework for NP-hard optimization problems, built on the classical primal-dual framework for designing efficient approximation algorithms. We enhance this framework by integrating optimal solutions to these NP-hard problems, enabling the model to surpass the performance of the approximation algorithms it was initially trained on. To the best of our knowledge, this is the first NAR method explicitly designed to surpass the performance of the classical algorithm on which it is trained. We evaluate our framework on several NP-hard combinatorial optimization problems, demonstrating its ability to generalize to larger and out-of-distribution graph families. In addition, we demonstrate the practical utility of the framework in two key applications: as a warm start for commercial solvers to reduce search time, and as a tool to generate embeddings that enhance predictive performance on real-world datasets. Our results highlight the scalability and effectiveness of GNNs for tackling complex combinatorial optimization problems, advancing their utility beyond the scope of traditional polynomial-time-solvable problems." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "neural algorithmic reasoning", "graph neural networks", "combinatorial optimization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/8f7ac14732a0b6a8fd1789254c6c91eaff7868f7.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Primal-Dual Graph Neural Networks for General NP-Hard Combinatorial Optimization" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The author's point-by-point discussion is well-executed but lacks an explanation of the logical relationships between the three different issues, which makes them feel somewhat disconnected. What is the logical relationship among these three issues—is it progressive or parallel?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The introduction of the PMBC strategy enables parallel processing in SNNs, representing an innovative advancement. Additionally, the proposed LIF neuron model incorporates a reset-refractory mechanism, enhancing both biological interpretability and dynamic computational capabilities.\n2. The article is well-structured, making the authors' arguments and experimental content easy to follow. The experimental results partially demonstrate the suitability of the proposed method as an SNN architecture for long-sequence tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents SPikE-SSM, a novel spiking state space model designed to address key challenges in long-sequence learning with spiking neural networks (SNNs). The authors introduce a boundary compression strategy (PMBC) to accelerate spiking neuron model inference, enabling parallel processing for long sequence learning. Additionally, they propose a new LIF neuron model with a reset-refractory mechanism to exploit the temporal dimension for biologically interpretable dynamic computation. The model's evaluation on long-range arena benchmarks and the large language dataset WikiText-103 demonstrates the potential of dynamic spiking neurons for efficient long-sequence learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The scalability of SPikE-SSM to larger datasets and more complex tasks has not been thoroughly discussed. Additionally, there is a lack of comparison with non-SSM-based SOTA SNN architectures in terms of computational efficiency and energy consumption. \n2. The presentation of PMBC in Figure 2 is unclear, making it difficult for readers to grasp the design rationale and core concepts being conveyed.\n3.What does \"x,h,x,...,x\" represent in Figure 3? It is not explained in the caption, which may lead to confusion.\n4. I couldn't find a complete diagram of the network architecture, nor any information on parameters or settings, in either the main text or the appendix.\n5. To my knowledge, there are several existing strategies for parallel training of SNNs[1][2], which the authors did not compare in this paper. What are the advantages of the proposed approach compared to these existing methods?\n[1] Fang, Wei, et al. \"Parallel spiking neurons with high efficiency and ability to learn long-term dependencies.\" Advances in Neural Information Processing Systems 36 (2024).\n[2] Zhang, Shimin, et al. \"Tc-lif: A two-compartment spiking neuron model for long-term sequential modelling.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 15. 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In Figure 1, the input to the GLU is a floating-point number. Is the multiplication of floating-point numbers unavoidable in this case? Or is there a more effective way to replace this module?\n2. While the article introduces sparsity through the reset mechanism and refractory period, is this nonlinear transformation interpretable? Specifically, how do the reset modules and refractory periods enhance model performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. A method similar to fixed-point iteration is proposed to solve the output based on the input sequence efficiently. This method can run in parallel, accelerating the training process.\n2. Better accuracy and sparsity compared with other spikingSSMS." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a spiking neural network model based on SSM. It employs a method similar to fixed-point iteration to address the challenge of training the reset mechanism in parallel. Additionally, the model introduces a refractory period for neurons, enhancing biological interpretability and resulting in sparser spike patterns." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Figure 1, the input to the GLU is a floating-point number. Is the multiplication of floating-point numbers unavoidable in this case? Or is there a more effective way to replace this module?\n2. While the article introduces sparsity through the reset mechanism and refractory period, is this nonlinear transformation interpretable? Specifically, how do the reset modules and refractory periods enhance model performance?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "The paper presents an incremental improvement over existing SNN-SSM methods but lacks the novelty, theoretical rigor, and experimental results to justify acceptance. While the proposed PMBC strategy and refractory neuron model are interesting, their practical benefits are not convincingly demonstrated. Additionally, the paper’s clarity issues further detract from its overall impact. A stronger focus on theoretical backing, real-world applications, and improved clarity would be needed for future consideration." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Detailed questions you can refer to weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed refractory neuron model is biologically inspired and brings interpretability to SNNs for long-sequence tasks. The experiments are well-structured, with results on a wide range of tasks (LRA benchmarks and WikiText-103) showing some promise." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents SPikE-SSM, a framework that aims to integrate spiking neural networks (SNNs) with state space models (SSMs) to address the challenges of long-sequence learning in an energy-efficient and sparse manner. The authors introduce several innovations such as the Parallel Max-Min Boundary Compression (PMBC) strategy for accelerating SNN inference, a refractory neuron model for temporal dynamics, and the integration of trainable thresholds for improved sparsity-accuracy trade-offs. Extensive experiments on LRA benchmarks and WikiText-103 demonstrate improved performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The core contribution—integrating SNNs with SSMs—is somewhat incremental. There are previous works, such as SpikingSSM, that already explore similar ideas. The novelty of the proposed solutions does not clearly differentiate it from existing methods, especially since the claimed improvements (like PMBC) are marginally better in terms of accuracy. The paper proposes several assertions and claims without sufficient theoretical backing. The proofs in the appendices are not rigorous enough and seem more heuristic than formal. This weakens the impact of the PMBC algorithm, which is a major part of the contribution. While the authors test their approach on multiple benchmarks, the results do not show significant improvements over previous methods. In many cases, the accuracy gains are minimal (sometimes less than 1%) compared to competing models, particularly when considering the energy-efficiency trade-offs. Furthermore, the benchmarks chosen (such as sMNIST) are relatively simple, and the performance on more challenging real-world tasks could have been more extensively evaluated.\n\nOne of the key selling points of SNNs is their energy efficiency. However, while the authors present some energy consumption estimates, these are based on theoretical assumptions (e.g., number of operations) rather than real-world implementations on neuromorphic hardware. Without real hardware measurements, the claimed energy benefits are speculative and diminish the practical relevance of the paper. The paper is dense and difficult to follow, particularly in sections where the PMBC algorithm and the refractory neuron model are introduced. The text often lacks clarity, and it is not always clear how the various components fit together. This makes the methodology hard to replicate." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "The work propsed a sparse, precise and efficient spiking state space model for long sequence learning, using parallel boundary compression and refractory spiking neurons." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024spikessm,\ntitle={{SP}ikE-{SSM}: A Sparse, Precise, and Efficient Spiking State Space Model for Long Sequences Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4ILqqOJFkS},\nnote={under review}\n}" }, "abstract": { "value": "Spiking neural networks (SNNs) provide a low-power, energy-efficient solution by utilizing the spike-based and sparse nature of biological systems. Since the advent of Transformers, SNNs have struggled to compete with artificial networks on long sequential tasks, until the recent emergence of state space models (SSMs), which offer superior computational efficiency and modeling capability. However, applying the highly capable SSMs to SNNs for long sequences learning poses three major challenges: ❶ The membrane potential is determined by the past spiking history of the neuron, leading to reduced efficiency for sequence modeling in parallel computing scenarios. ❷ Complex dynamics of biological spiking neurons are crucial for functionality but challenging to simulate and exploit effectively in large networks. ❸ It is arduous to maintain high sparsity while achieving high accuracy for spiking neurons without resorting to dense computing, as utilized in artificial neuron-based SSMs. To address these challenges, we propose a sparse, precise and efficient spiking SSM framework, termed SPikE-SSM. For ❶, we propose a boundary compression strategy (PMBC) to accelerate the inference of the spiking neuron model, enabling parallel processing for long sequence learning. For ❷, we propose a novel and concise neuron model incorporating reset-refractory mechanism to leverage the inherent temporal dimension for dynamic computing with biological interpretability. For ❸, we hierarchically integrate the proposed neuron model to the original SSM block, and enhance the dynamics of SPikE-SSM by incorporating trainable thresholds and refractory magnitudes to balance accuracy and sparsity. Extensive experiments illustrate the effectiveness and robustness of SPikE-SSM on the long range arena benchmarks and large language dataset WikiText-103, showing the potential of dynamic spiking neurons in efficient long sequence learning. The code will be publicly available." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "state space models", "spiking neural network", "long sequence modeling", "language modeling" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/9db2f9cc7f1d58a3162949512fe50d38467c2c6e.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "SPikE-SSM: A Sparse, Precise, and Efficient Spiking State Space Model for Long Sequences Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Regarding the proposed effective geometric measures to explain capacity changes, are there standard reference lengths compared to the radius? A simple manifold radius magnitude may not accurately capture the problem's complexity when the differences between manifold means scale identically.\n- Additionally, the definitions and implications of axes alignment, center alignment, and center-axes alignment are less discussed compared to radius and dimensionality (e.g., in Figures 5b, 6c).\n- When using geometric measures, which layer(s) should be analyzed? Are the results consistent across different layers?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The novel application of manifold capacity and other effective geometric measures to investigate the lazy-vs-rich dichotomy is intriguing. It shows better alignment with the degree of feature learning compared to other metrics, such as weight changes or NTK-label alignments.\n- Most of the derivations seem correct, though I could not verify every detail." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a geometric framework to examine the manifold of neural network representations, providing insights into the distinctions between lazy and rich training regimes in feature learning. Specifically, it revisits the manifold capacity concept introduced in Chung et al. (2018), theoretically demonstrating that manifold capacity can serve as an indicator of the underlying richness in feature learning. Based on empirical studies using synthetic data and two-layer neural networks, observations are made regarding the relationship between manifold capacity and the degree of feature learning, as well as the stages of feature evolution. The proposed geometric measures are further applied to neural networks in neuroscience and out-of-distribution generalization tasks to explore the broader implications of this approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The theoretical derivation relies on a one-step gradient argument, but the fact is not mentioned in the manuscript. Moreover, the link between Theorem 2’s results and the increase in feature learning degree is not entirely clear, and additional commentary could enhance clarity.\n- It would be helpful to discuss the generalizability of the observations, such as those in Figure 4. Various hyperparameters (e.g., the choice of optimization algorithms, weight initialization methods, batch size, learning rate, and scheduling) could influence implicit biases in the algorithm, affecting neural representations, geometric metrics, and even the stages of learning." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "-\tThere seems to be no explanation for Figure 2c. I don’t know what does those operations mean.\n-\tI’m a bit confused about eq (1), the definition of model capacity. Are the manifold $\\{\\mathcal{M}_i\\}$ predefined or they are changing when $N$ is increasing. If they are changing, how do you choose them? Also, what values should $y_i$ take?\n-\tWhen looking at Appendix C for the way to compute manifold capacity, several notions seem to be not defined, such as $T$, $\\lambda$.\n\n-\tI believe the way of computing manifold capacity should be mentioned in the main text, or at least mentioned under what conditions are those values computed (I believe they cannot be exactly computed unless some conditions are assumed).\n\n-\tFor Theorem 1, when looking at the actual statement in appendix, these results are proved only under the setting that with one gradient step update. Though it is understandable that going beyond this is technically difficult, I feel the statement in the main text gives the reader a wrong impression.\n\n-\tIn section 3.2/figure 3b, it’s not clear to me what the definition of wealthy and poor regime are and how they are related to the input dimension. Also, I’m not sure why at initialization there will be task-relevant features (and I don’t know what are task-relevant features in this setting). Moreover, the purple line in capacity in figure 3b changes only a little bit throughout the training, I’m not sure how to interpret this (feature learning only changes a little?).\n\n-\tIt’s a bit hard for me to follow section 4, presumably because the definition of these metric (e.g., dimension, radius,…) are missing.\n\n\nTypo:\n\n-\tLine 173, $o_N(1) \\to 1$ -> $o_N(1) \\to 0$" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "-\tUnderstanding feature learning in deep learning is an important and interesting research problem.\n-\tThe paper proposes a metric called manifold capacity to measure the feature learning progress, which seems to be new.\n-\tSeveral experiments in different domains are presented to support the claim." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors focus on the feature learning in deep learning. They use manifold capacity as a metric to quantify the degree of richness of feature learning. Experiment results show that such capacity can reveal different learning stages in different settings. They also apply this to problems in neuroscience and out-of-distribution tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "See questions section below." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In Section 2.1, what does \"i-th input category\" mean? In eq. (1), $P$ is the variable of the $\\max$ operator, and for any $i \\in [P]$, $\\mathcal M_i$ is defined, and $\\mathcal M_i$ is defined by $\\mathcal X_i$. This somehow implies that $\\mathcal X_i$ is also a part of the variable of the $\\max$ operator, instead of pre-defined. Is it true?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The idea of catching the richness by manifold capacity is potentially useful.\n- The experiments verifies that manifold capacity captures different stages of learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed that manifold capacity is a better way to identify the process of feature learning than commenly used rich-lazy dichotomy. This claim is theoretically proved in a limited setting, and empirically verified." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The presentation is very confusing. For example, in the experiments, there are several critical measures, such as effective radius, dimension, center alignment etc. but they are not defined (at least not in the main paper).\n- The main paper claimed that the Theorem 1 is proved for a 2-layer NN. However, the NN used in the proof is actually different from what people would expect to be a \"2-layer NN\", since based on Assumption 1: 1) the second layer is not trained but set to random; and 2) the choice of activation function is very limited as there is a rather strong condition on the activation function." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "What does \"untangling\" specifically mean? \n\nWhat are the implications of the 2-layer results on deeper neural networks?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "I think the use and introduction of manifold capacity as a measure is interesting for the ML community as a metric for representation learning. They relate manifold capacity to test accuracy in 2-layer networks and deeper networks empirically. They also relate to other measures such as weight changes and alignment, though this is mostly in the appendix." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper advocates the use of manifold capacity as a way of classifying as a measure of the regime of neural networks. Manifold capacity is a measure from neuroscience literature that quantifies the performance of a linear classifier over features for classification as a measure of the richness of classifiers. The authors advocate this measure over other measures such as \"lazy regimes\" during which features show limited change in learning. They use this to analyze 2-layer RELU networks" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper is haphazardly written. The difference between previous work and their problem statement is not well delineated. Their goals arent succinctly mentioned. The main theorem statement being nearly vacuous in the main paper. \"In 2-layer neural networks trained with gradient descent in the rich regime the changes in capacity track the underlying degree of richness in feature learning.\" Since richness itself is defined as capacity, this is either circular or vacuous. \n\nA deeper read of the result in the appendix shows that they approximate the result of gradient descent and use a gaussian model to approximate a 2-layer network. This is perhaps the most interesting result in the paper but is not at all covered in the main paper. The authors should state that they prove that using SGD, models with high capacity converge to high accuracy or something less vacuous and offer the proof. \n\nOther features such as manifold radius and \"alignment\" are also not measured. The connection to \"untangling\" is also not mentioned. Weight changes are used mainly as a strawman to invoke connections to NTK and need not be done. \n\nFigure 2 is very confusing with a nonsensical caption, \"Higher capacity means that a higher number of\nmanifolds per neuron can be packed in the neural state space.\" What does it mean to \"pack manifolds into neurons.\" This kind of shoddy language obfuscates the message to the reader. \n\nAs the current writing stands, this seems to contribute no more (and likely less by way of confusion) to the work in 2018 by Chung et al. which describes the full intuition of manifold capacity and outlines its use both in neuronal and neural networks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Could the authors provide an example of the lazy regime? Specifically, how can a neural network be trained without actually modifying its internal features?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors present both theoretical analysis and empirical evidence demonstrating that manifold capacity effectively quantifies the degree of feature learning. Comparisons with other metrics, such as accuracy and weight changes, further underscore the efficacy of manifold capacity in this context.\n2. The authors offer insightful empirical findings on the relationship between manifold geometry and learning stages, with applications in neuroscience and OOD detection. Robust experiments substantiate these findings, providing a solid foundation for their conclusions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel framework for understanding neural network feature learning beyond the “lazy” versus “rich” dichotomy by focusing on task-relevant manifold untangling using representational geometry. The authors introduce manifold capacity as a key metric to quantify feature learning richness, and explore how manifold geometric measures reveal distinct learning stages and strategies. This framework is applied in contexts ranging from standard machine learning tasks to neuroscience, offering insights into structural inductive biases and challenges in out-of-distribution (OOD) generalization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The structure needs refinement, as it currently feels like the authors have packed too much content into the main paper, resulting in diminished clarity. The main paper covers a wide range of topics—from manifold capacity to manifold geometry, from theory to experiments—and the relationship between manifold capacity and various manifold geometry measures is weakly explained, relying primarily on Figure 2c with minimal analysis. I suggest that the authors avoid treating manifold geometry as a separate section, even though some interesting findings are presented. An alternative approach would be to frame geometry as an extension of capacity (or to present capacity itself as a facet of manifold geometry) and to provide necessary analysis (I notice there is analysis in the appendix. It's better to consolidate relevant analysis from the appendix into a concise summary in the main paper).\n2. The algorithm is not clearly presented, which also appears to be a side effect of the structural issue noted above. Given that the paper aims to provide practical insights into feature learning, with applications in neuroscience and machine learning, it is important to ensure readability for readers unfamiliar with the manifold background. Thus, a clear algorithm outlining the process for computing capacity during training would be essential." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Quantifying feature learning beyond lazy versus rich through manifold capacity and representational geometry." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024beyond,\ntitle={Beyond the Lazy versus Rich Dichotomy: Geometry Insights in Feature Learning from Task-Relevant Manifold Untangling},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4IRYGvyevW},\nnote={under review}\n}" }, "abstract": { "value": "The ability to integrate task-relevant information into neural representations is a fundamental aspect of both human and machine intelligence. Recent studies have explored the transition of neural networks from the *lazy* training regime (where the trained network is equivalent to a linear model of initial random features) to the *rich* feature learning regime (where the network learns task-relevant features). However, most approaches focus on weight matrices or neural tangent kernels, limiting their relevance for neuroscience due to the lack of representation-based methods to study feature learning. Furthermore, the simple lazy-versus-rich dichotomy overlooks the potential for richer subtypes of feature learning driven by variations in learning algorithms, network architectures, and data properties.\n\nIn this work, we present a framework based on representational geometry to study feature learning. The key idea is to use the untangling of task-relevant neural manifolds as a signature of rich learning. We employ manifold capacity—a representation-based measure—to quantify this untangling, along with geometric metrics to uncover structural differences in feature learning. Our contributions are threefold: First, we show both theoretically and empirically that task-relevant manifolds untangle during rich learning, and that manifold capacity quantifies the degree of richness. Second, we use manifold geometric measures to reveal distinct learning stages and strategies driven by network and data properties, demonstrating that feature learning is richer than the lazy-versus-rich dichotomy. Finally, we apply our method to problems in neuroscience and machine learning, providing geometric insights into structural inductive biases and out-of-distribution generalization. Our work introduces a novel perspective for understanding and quantifying feature learning through the lens of representational geometry." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Computational neuroscience", "storage capacity", "neural manifolds", "representational geometry", "rich and lazy learning", "training dynamics", "feature learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/50738ce430af44d4b48fd2f40874dc6db9646568.pdf" }, "presentation": null, "primary_area": { "value": "applications to neuroscience & cognitive science" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Beyond the Lazy versus Rich Dichotomy: Geometry Insights in Feature Learning from Task-Relevant Manifold Untangling" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In Figure 3, why did the color of the part of the image outside the trigger also change?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Extensive experiments demonstrating the effectiveness of REFINE across different datasets.\n\n- Thorough ablation studies validating each component's contribution" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose an inversion-free backdoor defense method based on model reprogramming, REFINE. REFINE consists of an input transformation module and an output mapping module. The key idea is to transform both input and output domains to break the trade-off between model utility and defense effectiveness faced by existing pre-processing based defenses." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The connection between motivation and the proposed methods is not very close. However, the analysis of BTI-based defenses does not involve this domain-based perspective. The paper didn’t discuss whether and how the purification process in BTI-based defenses alter the image domain.\n\n- The theoretical analysis through Theorem 1 effectively explains the limitations of transformation-based defenses by quantifying how domain transformations affect defense performance.\n\n- Experiments mainly focus on ResNet. Conducting experiments under different network structures is beneficial for verifying the effectiveness of the proposed method, for example, VGG, DenseNet etc.\n\n- Requiring 20% clean data may be impractical in many real-world scenarios.\n\n- The proposed method requires training a U-Net, which involves a significant computational cost." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The two sentences (i.e., \"by first obtaining ... outcomes\" and \"the internet ... poisoned samples)\") in the Section 1 contradict each other. What is the difference between prior information and prior knowledge?\n2. This paper focuses on the dirty-image backdoor attacks. But considering the recent clean-image backdoor attacks that do not have trigger features, can the proposed defense method work?\n3. In Section 3.2, the \"Pad Size\" and \"Mask Ratio\" are not defined before. it is necessary to clarify.\n4. In this paper, there are many long sentences, which makes it hard to understand. Such as \"Additionally, we treat ... the transformation\" on Page 4.\n5. In Section 3.2, the authors did not show the dataset used for limitation analysis. Please clarify it.\n6. In Section 4.2.1, the authors present that traditional model reprogramming methods are insufficient to remove triggers. I suggest that the authors explain why traditional methods are insufficient in detail.\n7. The discussion about the feature distance changes in the proposed input transformation module is lacking. It is better to add it to highlight the function of this module." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. A theoretical analysis demonstrates that the effect of backdoor defenses is bounded by the distance of the output features before and after the preprocessing. Therefore, existing methods can not break the trade-off between the model utility and the defense effectiveness.\n2. The proposed method is novel and interesting. By integrating model reprogramming techniques, they only need to change the model input without changing the model parameters to achieve backdoor elimination, and it does not affect the original performance of the model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper analyzes existing pre-processing-based backdoor defense methods and their limitations. Then, a simple yet effective method is proposed. This method utilizes model reprogramming techniques rather than model retraining, which not only eliminates backdoor attacks but also maintains the model's performance. The evaluation of different backdoor attacks also demonstrates the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors discuss the pre-processing defense methods, i.e., input-transformation defenses and BTI-based defenses, and analyze their limitations in details. However, the proposed methods actually belong to the input transformation-based method. This paper also spend a large amount of time to analyze and compare BTI methods with the proposed method, which makes it hard to read.\n2. This paper assumes that they have access to an unlabeled dataset that is independent and identically distributed to the training dataset of the pre-trained model. The authors need to clarify how to acquire this dataset. In addition, since the performance of model reprogramming methods is related to the distribution of input data, if the author cannot obtain data with the same distribution, will it affect the performance of the model?\n3. The experiment is pretty insufficient. REFINE is only evaluated with the ResNet-18 model. The evaluation under more complex models (such as ResNet-50) or other architecture models (such as InceptionV3 or VGG16) is lack." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Please improve the soundness of the limitation study.\n2. Please discuss the reasonability of the categories of the mentioned defenses.\n3. Please explain the different experimental results." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Revisit the pre-processing defenses against backdoor attacks and reveal their limitations. The pre-processing-based defense is important to protect model security while not changing model structure or weights.\n2. Propose a pre-processing defense against backdoor attacks, which seems to be simple but effective.\n3. Conduct extensive experiments to demonstrate the effectiveness of the proposed defense." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes REFINE, an inversion-free backdoor defense method based on model reprogramming. Based on the research, the authors revisit existing pre-processing-based backdoor defenses and reveal their limitations. The proposed defense, REFINE, introduces trainable input transformation and output mapping modules for reprogramming and incorporates cross-entropy and supervised contrastive losses to enhance defense performance. Extensive experiments are conducted to validate the effectiveness of REFINE." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The claim of the limitations of prior works are subjective and confusing. For the first limitation, the authors think \"transformation-based backdoor defenses methods face a trade-off between utility and effectiveness\". So, can the proposed defense overcome this limitation? From the design and experimental results, REFINE also suffer from the same problem. Otherwise, the BA with REFINE should be same with original model. Moreover, the authors try to utilize experiments to validate their claims. However, the experimental setting is not fair enough. For example, the authors mention ShrinkPad as the baseline. However, in that paper, ShrinkPad is not the best defense method. Also, for the second limitation, the authors do not analyse the SOTA work, and the shown experimental results are different from the original paper, e.g., the experiments with BADNET shown in Fig.3.\n2. The categories of the mentioned defenses are not sound enough. According to the manuscript, the BTI-based defense (e.g., NC) reverses the backdoor trigger to eliminate the trigger in the inputs. However, such a kind of defense does not only work in this way. Thus, the comparison seems to be not fair enough.\n3. The experiments do not involve some SOTA works. e.g, [1, 2]. Moreover, the experimental results differ from the original paper a lot, e.g., BTI-DBF. Please explain why.\n[1] Neural Polarizer: A Lightweight and Effective Backdoor Defense via Purifying Poisoned Features\n[2] Black-box Backdoor Defense via Zero-shot Image Purification" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "See weakness part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The idea may seem bold and imaginative, but it is worth exploring in depth. I believe both the threat model setup and the proposed solution are good.\n2. Overall, the paper is well-written, making it relatively easy for someone unfamiliar with model reprogramming, like myself, to grasp the design concepts.\n3. The method operates effectively with a small amount of unlabeled data, which not only cleverly avoids direct comparison with fine-tuning-based defenses but also demonstrates its practicality through ablation experiments that confirm its minimal data requirements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the challenge of backdoor attacks in deep neural networks, where current defenses struggle to balance accuracy and trigger removal. The proposed REFINE method combines input transformation and output remapping to effectively neutralize backdoors without relying on trigger inversion. Experiments show REFINE's strong defense capabilities while maintaining model performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.Regarding the theoretical explanation in the paper (Formula 1), I feel it’s missing a key component and has only achieved about 50% of its purpose. It aims to explain how amplifying the distributional differences in output features changes the variation in prediction results. As the authors emphasize, model reprogramming significantly amplifies these distributional differences, and trigger patterns can be randomized. I agree with this point, but at this stage, the performance guarantee for clean samples doesn’t tie back to this theory; I see it as being empirically supported by the two losses. In fact, even with model reprogramming, Formula 1 still holds, so it may not fundamentally support the authors’ viewpoint (since Formula 1 is a rule that must be met across the overall distribution; I believe focusing on local distributions could be a potential solution).\n\n2.The whole experiments are only conducted on a ResNet-18 model. Furthermore, the discussion of the black-box part in the experiments seems counterproductive to me; I don’t understand the rationale behind setting up absurdly similar black-box and substitute models for exploration.\n\n3.The adaptive attack setup is unreasonable. Based on the general consensus in the community, in adaptive attacks, the attacker is assumed to know the defender’s hard-coded choices. Disallowing the defender from making real-time adjustments (which is a strategy directly aimed at adaptive attacks) is therefore a logical error. Additionally, the authors haven’t clearly explained why a conflict exists between adaptive loss and backdoor loss, resulting in a decrease in BA, and I don’t clearly see a trade-off here.\n\n4.Why is BTI-DBF unable to reverse even the most basic BadNet trigger? I seriously question the validity of this experiment. In the original BTI-DBF paper, it’s evident that it can reconstruct the triggers of certain fixed samples quite effectively.\n\n5.Additionally, I would like to see more examples of transformed samples. Do transformed samples from the same class still share some visual similarities, or do they contain almost no human-recognizable information? What is the underlying significance of these transformations? I am curious if the authors could offer some more unique, high-level insights." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024refine,\ntitle={{REFINE}: Inversion-Free Backdoor Defense via Model Reprogramming},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4IYdCws9fc},\nnote={under review}\n}" }, "abstract": { "value": "Backdoor attacks on deep neural networks (DNNs) have emerged as a significant security threat, allowing adversaries to implant hidden malicious behaviors during the model training phase. Pre-processing-based defense, which is one of the most important defense paradigms, typically focuses on input transformations or backdoor trigger inversion (BTI) to deactivate or eliminate embedded backdoor triggers during the inference process. However, these methods suffer from inherent limitations: transformation-based defenses often struggle to balance the intensity of transformations with preserving the model's accuracy, while BTI-based defenses require accurate reconstruction of the trigger patterns, which is rarely achievable without prior knowledge. In this paper, we propose REFINE, an inversion-free backdoor defense method based on model reprogramming. REFINE consists of two key components: (1) an input transformation module that disrupts both benign and backdoor patterns, generating new benign features; and (2) an output remapping module that redefines the model's output domain to guide the input transformations effectively. By further integrating supervised contrastive loss, REFINE enhances the defense capabilities while maintaining model utility. Extensive experiments on various benchmark datasets demonstrate the effectiveness of our REFINE and its resistance to potential adaptive attacks." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Backdoor Defense", "Model Reprogramming", "Backdoor Attack", "AI Security" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/69b0fb1497a8c6ace495f0e0afe647cdf0926876.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/0971dc6452f2bca11056898fa40c9009bcec8d05.zip" }, "title": { "value": "REFINE: Inversion-Free Backdoor Defense via Model Reprogramming" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "### Minor comments\n\nFigure 3, page 5: \nThe color bar is labeled \"Gzip Alignment\" instead of\n\"ZIP-FIT-Alignment\" from Algorithm 1; it may be confusing to readers.\n\nFigure 3, page 5, line 231: \nPlease mention also in the figure caption that the test loss is\ncalculated on ProofNet data." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper is concise, sound, well written, and the experimental section shows promise for the method, especially with regard to other embedding-free methods.\n\nThe conceptual simplicity combined with the empirical results of the method is an especially strong point of the work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new data selection mechanism based on text\ncompression distances. The concept of using compression methods for\ndeep learning follows several modern practical results and theoretical\nmotivations that language modeling is fundamentally based in text\ncompression. The method's conceptual simplicity combined with strong\nempirical results make it stand out as a modern way for filtering for\naligned data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Ideally, it would be shown how the size of $n$ (i.e., number of\nsamples from the target domain $p$) influences the performance of the\nmethod. If it is possible to pick $n$ just sufficiently large enough,\nit would greatly improve the computational efficiency of the method\nfor large target datasets.\n\nExperiments in other domains would be really nice to better\ndemonstrate the generalization capabilities of the method. Possibly\nthere is data that is not well-suited to compression and accordingly\nZIP-FIT, or where the data's compression factor varies too much\nbetween samples?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could you provide further insights into how ZIP-FIT might perform with data that have higher variability and diverse syntactic structures, such as conversational datasets?\n2. Can you clarify the theoretical basis for using gzip compression over other compression methods that might exploit redundancy differently? Would alternative compression algorithms affect the performance of ZIP-FIT?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. ZIP-FIT’s embedding-free approach is a refreshing deviation from common embedding-based methods, offering a novel solution by leveraging gzip compression. The concept of using normalized compression distance (NCD) as an alignment metric is insightful and could inspire future research in embedding-free methodologies for various data selection tasks.\n2. The empirical results support the claims, showing that ZIP-FIT achieves faster convergence and better performance than established methods. The experiments were conducted on both AutoFormalization and code generation tasks, demonstrating ZIP-FIT's versatility across different domains.\n3. The paper is well-structured, with a clear exposition of the algorithm, experimental setup, and results. The figures effectively illustrate the performance benefits of ZIP-FIT.\n4. ZIP-FIT could represent a significant advancement in data selection for machine learning, particularly in computationally constrained environments. Its potential to optimize model fine-tuning with minimal resource requirements makes it highly applicable for real-world use cases, especially in domain-specific and low-resource applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces ZIP-FIT, an embedding-free data selection method leveraging gzip compression to measure the alignment between training and target domains. Unlike existing approaches that rely on neural embeddings, ZIP-FIT uses a computationally efficient compression-based alignment metric, enabling faster data selection while maintaining high relevance to the target task. Empirical evaluations demonstrate ZIP-FIT’s superiority over baselines DSIR and D4 in AutoFormalization and code generation tasks, achieving significantly faster convergence and lower cross-entropy loss with reduced computational costs. ZIP-FIT’s promise lies in its scalability and effectiveness, particularly in low-resource settings, where traditional embedding-based methods may be impractical." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While ZIP-FIT achieves excellent results on the tasks tested, its reliance on gzip compression may limit its effectiveness in complex semantic domains where relationships are nuanced and less compressible. Embedding-free approaches, while efficient, may not be ideal for tasks that require deep semantic understanding or complex syntactic relationships." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "As specified in the \"Weaknesses\" section:\n- What is the score of the fine-tuned LLM using ZIP-FIT on benchmarks like HumanEval and PubMedQA compare to LLMs fine-tuned without using ZIP-FIT?\n- How does ZIP-FIT compare to prior method like https://arxiv.org/abs/2405.00705 in terms of both running time and final model score?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This paper is well-presented and well-motivated.\n- Studying computation-efficient methods for data selection in LLM instruction fine-tuning is a promising research direction.\n- The proposed ZIP-FIT is intuitive and easy to follow.\n- The proposed approach bypasses the need for LLM forward computation to obtain embeddings, making it computationally efficient.\n- The presented experimental results seem promising." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes ZIP-FIT, an efficient, embedding-free method for selecting high-quality, domain-specific fine-tuning data for language models (LMs). Prior methods often rely on computationally expensive neural embeddings or classifiers to filter aligned datasets, while those based on N-gram similarity may lack the structural depth needed for complex tasks like code generation. In contrast, ZIP-FIT leverages gzip compression to evaluate data alignment with target domains, based on the idea that compression algorithms encode information in a way similar to neural networks. The ZIP-FIT approach eliminates the need for LM forward passes to obtain embeddings, making it efficient and particularly suitable for low-resource environments. Experimental results show that ZIP-FIT outperforms prior data selection methods, such as DSIR and D4, as measured by test loss." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- [Major] The proposed method seems very simple and straightforward; using a gzip-style method to embed data appears to be a relatively standard approach.\n- [Major] All experimental results are based on test loss, which may not be very reliable. It would be essential to conduct evaluations on some standard benchmarks, such as HumanEval and MBPP for code evaluation, to demonstrate the scores the model can achieve.\n- It is unclear how the proposed ZIP-FIT compares to prior, more complex data selection methods in terms of both running speed and final model quality (e.g., [1]), aside from deduplication approaches like D4.\n- [Minor] The paper seems to be written somewhat in rush, the figure quality of Figure 2 does not seem to be very high.\n\n[1] https://arxiv.org/abs/2405.00705" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "(1) Could the authors provide additional evidence to support the claim that gzip is effective in capturing syntactic and structural relationships in textual sequences?\n\n(2) Would the authors be able to demonstrate the effectiveness of their approach using evaluation metrics beyond cross-entropy test loss, and compare it to relevant baselines, such as those mentioned earlier?\n\n(3) Could you provide more insight into why D4 was excluded from the code generation experiments, and specifically how it affected model performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "(1) Problem Significance: The author tackles a crucial problem in low-resource settings, addressing the challenge of fine-tune data selection without relying on GPU-intensive and embedding-based methods. This is a highly relevant and impactful research direction.\n\n(2) Innovative Filtering Criterion: The authors' inspiration from gzip compression methods has led to the proposal of a novel and intriguing selection criterion. This approach is not only interesting but also demonstrates out-of-the-box thinking, making it a notable contribution to the field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an innovative, embedding-free data selection method for efficient fine-tuning of large language models. Drawing inspiration from gzip compression techniques, the authors propose utilizing Normalized Compression Distance as a metric to filter and prune fine-tuning datasets. The authors conduct a comparative analysis with prior embedding-free methods, originally designed for filtering pre-training datasets, on Autoformalization and Python coding tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) **Inadequate Baselines**: The authors propose a data selection method for model alignment, but only compare it with prior works such as DSIR and D4, which were primarily designed for data selection during the pre-training phase. A more comprehensive literature review on data pruning methods for model alignment is lacking, including embedding-based methods [1], LLM model response metrics [2], Gradient-based metrics [3],Quality metrics judged by LLMs [4], inference loss on evaluation sets [5].\n\n(2) **Evaluation Metrics**: The authors primarily use test data cross-entropy loss as the evaluation metric, results are thus not surprising given that the data selection method uses the test data to anchor the selection criteria. However, the authors do not compare their results with widely accepted metrics in the research community for the studied downstream tasks, such as:\n\n(a). Autoformalization: proof success rates on miniF2F [6,7]\n\n(b). Python coding: functionality pass rates (pass@k on HumanEval) based on unit-tests [8,9]\n\n\n(3) **Clarifications on Motivation**: In Section 2.3, the authors argue that n-grams fail to capture syntactic or structural relationships within the data, while hypothesizing that gzip does. However, this hypothesis is not supported by theoretical or empirical evidence, weakening the motivation for the proposed approach. It is also not compared on if the proposed approach is better or worse than high-resource methods, such as embedding-based methods.\n\nReferences: \n\n[1] DEFT-UCS: Data Efficient Fine-Tuning for Pre-Trained Language Models via Unsupervised Core-Set Selection\n\n[2] From Quantity to Quality: Boosting LLM Performance with Self-Guided Data Selection for Instruction Tuning\n\n[3] LESS: Selecting Influential Data for Targeted Instruction Tuning\n\n[4] Alpagasus: Training a better alpaca with fewer data\n\n[5] Instruction Mining: Instruction Data Selection for Tuning Large Language Models\n\n[6] Autoformalization with Large Language Models\n\n[7] LEGO-Prover: Neural Theorem Proving with Growing Libraries\n\n[8] Evaluating Large Language Models Trained on Code\n\n[9] Is Your Code Generated by ChatGPT Really Correct? Rigorous Evaluation of Large Language Models for Code Generation" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "use gzip to select optimal data for code and autoformalization" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024zipfit,\ntitle={{ZIP}-{FIT}: Embedding-Free Data Selection via Compression-Based Alignment},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4JBEpP6eRS},\nnote={under review}\n}" }, "abstract": { "value": "Selecting high-quality, aligned fine-tuning data is crucial for improving the downstream performance of language models (LMs). Automatic data selection in these scenarios is challenging and often inefficient due to previous approaches relying on neural embeddings or limited n-gram representations to identify aligned datasets. In addition, traditional data selection methods often focus on increasing the size of the training data, making them computationally expensive to use and data inefficient. In this work, we introduce ZIP-FIT, an embedding-free, data-efficient selection framework that leverages gzip compression to measure the alignment between training data and target domains. We show that ZIP-FIT significantly outperforms two leading baselines, DSIR and D4, in selecting high-quality data for ProofNet, a formal mathematical dataset, and HumanEval, a benchmark for code generation tasks. Specifically, ZIP-FIT demonstrates a computational speed advantage, performing data selection up to 65.8\\% faster than DSIR and achieving its lowest cross-entropy loss up to 85.1\\% faster. Our findings suggest that ZIP-FIT offers a scalable and adaptable approach for data selection, enabling more precise fine-tuning for code generation domains. By demonstrating that embedding-free data selection can outperform established methods like DSIR and D4, our research opens new avenues for optimizing model training, thereby enhancing the effectiveness and efficiency of machine learning workflows." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "data centric machine learning", "autoformalization", "large language models", "reasoning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/a706bb5d7465dc4f64fab987f6d76be8adb5a408.pdf" }, "presentation": null, "primary_area": { "value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "ZIP-FIT: Embedding-Free Data Selection via Compression-Based Alignment" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The authors assume that the observation model p(y|x0) is linear gaussian. Is there any other real world problems can be expressed in this form? \n- Continue with weakness section: Can you provide more experimental results? \nFor example, does the performance changes when using more Heun steps? does the improvement also happen to other type of discretization methods, e.g. Euler? Also there are many other ODE samplers and SDE samplers, if you change the ODE samplers with SDE ones, does the introduced stochasticity make any difference when using your covariance estimate method?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Novel method, proposed a covariance estimate method via separately updating along time and position/space. Also provide a practical implementation for the proposed method. \n- The paper provides theoretical insights into why accurate covariance estimation is crucial for unbiased conditional generation, which is supported by experimental results on inverse problems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce a new covariance estimation method which makes use of covariance information in denoiser and the trajectory curvature without introducing significant additional compute. Moreover, to make the approach suitable for high-dimensional data (e.g. covariance matrices storage issue), the low-rank updates is proposed. The authors validate their approach on linear inverse problems, demonstrating its effectiveness compared to baselines under four metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experiment section is not as strong as methodology section: only one dataset and one ODE denoiser is tested under low NFE setting. The design of the experiment and the corresponding results analysis can be improved. \n- Only one ODE sampler (Heun) with steps=15, 30 is tested. While indeed results show the importance of accurate covariance estimate, it’s hard to conclude that this is applicable to all standard diffusion models." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In the image data results, adding the online update steps to either identity covariance or the FH covariance does not always improve the performance. Is there any explanation or further investigation on why it could happen? \n2. Table 1 shows that in most cases, FH+online outperforms FH. But in Figure 7, only the denoised image from FH is demonstrated. Is there any vision difference between FH and FH+online in terms of the images in Figure 7?\n3. Notational /exposition issues:\n 1. In Eq. (1), the \"noising forward process\" should be defined in the joint distribution of the stochastic paths or a stochastic differential equation. Eq. (1) gives the margin distribution of the state at time t, which cannot formally \"define\" a process.\n 2. In Eqs. (2) and (3), $\\dot\\sigma$ and $\\omega_t$ are undefined. \n 3. Above Eq. (7), $p_{0|t}$ is undefined. \n 4. In Eq. (11), add a bracket to make the inverse clear. I.e., $[\\nabla_{\\mathbf{x}}^2\\log p(\\mathbf{x},t)]^{-1}$.\n 5. The last few sentences in Sec. 3.1. confused me. The integral $\\int p(\\mathbf{x}')\\mathcal N(\\mathbf{x}'\\mid \\mathbf{0}, \\sigma^2\\mathbf{I})d\\mathbf{x}'$ is not a convolution. \n 6. In Eq. (18), should the right-hand side be $\\left[\\mathbf{\\Sigma}_{0\\mid t}(\\mathbf{x}+\\Delta\\mathbf{x})\\right]\\Delta\\mathbf{x}$.\n 7. In Eq. (23), use either $\\nabla_{\\mathbf{x}}$ or $\\nabla_{\\mathbf{x_t}}$ consistently throughout the paper.\n 8. In Eqs. (23) and (24), $\\mathbf{A}$ is undefined.\n 9. Above Eqs. (25), $\\mu_{0\\mid t}$ should be in boldface. (and many other occurrences throughout the paper).\n 10. In Sec. 4.2., \"we noticed that the guidance scale is overestimated the larger the dimensionality is, leading to overconfidence.\" does not read well --- the authors probably left over some words from a previous edition. \n 11. Throughout the paper and the appendix, please only label the equations you refer to later." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "Originality and significance: The proposed method is relevant to the diffusion models and is novel in that it provides a fast and efficient algorithm for estimating a broader class of covariance structures. \n\nQuality and clarity: The paper clearly connects the proposed method to existing literature on similar problems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors proposed a new methodological framework for estimating covariance in the generative methods of diffusion models. The estimation of the covariance in the reverse diffusion process is purely based on the existing samples through a two-step updating scheme. The examples show the proposed method outperforms other existing methods in denoising." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The exposition can be improved as the current manuscript contains many inconsistent notations, undefined variables, and many typos. I would say it is a lovely work with appealing results, but unfortunately, the manuscript should be significantly revised and proofread to be easier to follow." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Since the data covariance is estimated in the DCT basis where the covariance is approximately diagonal, have you considered training a model to directly estimate the denoiser covariance in this basis, or with a diagonal + low-rank representation?\n\n2. The numerical results in Table 1 seems to be worse than that in the DDNM paper, is this because the number of sampling steps is limited?\n\n3. What is the computational overhead for estimating the data covariance and doing the low-rank updates, compared to the cost of evaluating the neural network?\n\n4. What is the variance exploding case first mentioned in Line 213? This doesn't seem to be explicitly defined in the paper.\n\n5. What is the difference between $\\mu_{0\\mid t}(x)$ and $\\mu_{0 \\mid t}(x_t)$, as well as $\\sigma(t)$ and $\\sigma_t$?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. It is difficult to efficiently estimate the covariance of the denoiser due to its high-dimensionality, and methods that attempt to estimate it has have to use reduced-complexity representations. Using the DCT-diagonal basis to estimate the initial data covariance and a low-rank update inspired by quasi-Newton methods is a novel approach.\n\n2. Because this method does not require training an additional neural network, it can be applied to existing pre-trained diffusion models to be used for solving inverse problems.\n\n3. The experimental results seem promising and improves on previous methods on both qualitative examples and quantitative metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new method to estimate the covariance during the denoising process in diffusion models. The key idea is using Tweedie's formula to relate the covariance of the denoiser with the Hessian of $\\log(p(x_t))$, then using the gradient of $\\log(p(x_t))$ (i.e. the score function) evaluated during each denoising step to estimate this Hessian. Starting with an estimate of the covariance of the data distribution, the algorithm performs low-rank updates to this covariance matrix, inspired by quasi-Newton methods, using the score function estimates given by the neural network. This method is then applied to fast sampling for solving inverse problems, showing improvements over other methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There needs to be more theoretical or empirical justification for some of the algorithmic choices made, such as the choice of quasi-Newton method or the DCT basis used to approximate the data covariance.\n\n2. The notation in the paper is inconsistent and the definition of some quantities are not clear, for example are $p(x, t)$, $p(x_t)$, $p_t(x)$ referring to the same quantity? Also, it is not clear what exactly are the full algorithms used for solving linear inverse problems; Algorithms 1 and 2 in the paper only updates the covariance estimate.\n\n3. The experimental improvements for solving linear inverse problems seems to be only for the few-sample-step regime. It is unclear if this covariance estimation method can be helpful in other applications such as unconditional sampling, or with non-linear guidance terms." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. refer to weaknesses.\n2. in Figure 6, these setting of y is what? inpating transform or deblurring? Does this guidance meaningful across different y|x_0 and dataset?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. It is novel and reasonable to design low-dim expression and update rules for covariance approximation.\n2. There are abundant math derivations to support the methods" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new framework for estimating covariance in diffusion models, addressing issues in current methods such as high test-time computation and heavy approximations. The proposed method utilizes readily available covariance information from training data and the curvature of the generative trajectory. The authors also present a method for transferring covariance estimates across noise levels." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. To handle linear inverse problems using diffusion models, there are a lot of classes of methods [1]. The experiments in this work only compare some of them, arthors should at least mension why take them out of comparison.\n2. There lack of a experiment to back up whether the low-dim covariance of the proposed method can indeed match the underlying true one. I understand this cannot be done under high dimensional case, mayba a 2D or 10D suffice.\n3. There is other applications of diffusion models would need covariance like causual reasoning [2], likelihood-evaluation [3][4] and adjoint guidance[5], maybe you could add a discussion to these.\n\n\n[1] Daras, Giannis, et al. \"A survey on diffusion models for inverse problems.\" arXiv preprint arXiv:2410.00083 (2024).\n\n[2] Sanchez P, Liu X, O'Neil A Q, et al. Diffusion models for causal discovery via topological ordering[J]. arXiv preprint arXiv:2210.06201, 2022.\n\n[3] Lu C, Zheng K, Bao F, et al. Maximum likelihood training for score-based diffusion odes by high order denoising score matching[C]//International Conference on Machine Learning. PMLR, 2022: 14429-14460.\n\n[4] Anonymous. Gradient-Free Analytical Fisher Information of Diffused Distributions.\n\n[5] Song K, Lai H. Fisher Information Improved Training-Free Conditional Diffusion Model[J]. arXiv preprint arXiv:2404.18252, 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "see Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. It is novel to leverage L-BFGS way to approximate the covariance in a low-dim manner in DMs\n2. the mathematical derivation is rigorous, though it is quite similar to the math of BFGS optimization\n3. The design of initial covariance to be the data covariance is reasonable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduce a L-BFGS-like method to maintain the covariance information of denoiser along the inference path to better solve the linear inverse problems via diffusion models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The presentation of the paper could be improved for clarity. The term 'conditional generation' is often associated with 'text-guided' or 'label-guided' generation. However, in this paper, it refers to conditions based on partially corrupted samples. While this interpretation is not incorrect and indeed falls under the classifier-guided framework, it might be confusing for readers at first glance. \n\n2. There are typographical errors present in the paper, for instance, in equation (98). These need to be corrected for accuracy and better understanding.\n\n3. The paper lacks a theoretical bound on the approximation error of the covariance using the proposed method. Both Equation 6 and the low-rank approximation involve approximation errors." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a new, efficient method for denoiser covariance estimation in diffusion models, which can be used for conditional generation and inverse problems" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024free,\ntitle={Free Hunch: Denoiser Covariance Estimation for Diffusion Models Without Extra Costs},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4JK2XMGUc8},\nnote={under review}\n}" }, "abstract": { "value": "The covariance for clean data given a noisy observation is an important quantity in many conditional generation methods for diffusion models. Current methods require heavy test-time computation, altering the standard diffusion training process or denoiser architecture, or making heavy approximations. We propose a new framework that sidesteps these issues by using covariance information that is available for free from training data and the curvature of the generative trajectory, which is linked to the covariance through the second-order Tweedie's formula. We integrate these sources of information using (i) a novel method to transfer covariance estimates across noise levels and (ii) low-rank updates in a given noise level. We validate the method on linear inverse problems, where it outperforms recent baselines, especially with fewer diffusion steps." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "diffusion model", "conditional generation", "inverse problems", "denoiser covariance estimation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/ee3b68d8cb17efc6aa3b1d0a5d36edc4e6fc7585.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/5c042e4d069a5c3eaccb3fadad08ed86db001629.zip" }, "title": { "value": "Free Hunch: Denoiser Covariance Estimation for Diffusion Models Without Extra Costs" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Why the proposed method is only applied to RNA design? Can the method be extended to any other biological sequence (DNA, protein, peptide) design? If yes, why restrict its applicability/generalization?\n\n2. As also stated in the first weakness, what is the main source of benefit of using an RL method, instead of performing collective design for combinatorial optimization as in [1]?\n\n3. How does CocoRNA deal with sparse and/or delayed rewards?\n\n4. How does CocoRNA generalize across different target structures?\n\n5. How is SAER related to/inspired by HER? How the goals are sampled in SAER? How do you gradually reduce the SAER operation, this is not clear in the paper.\n\n6. Can you provide some insight/discussion on how embedding SAER into the training process would not yield a suboptimal solution?\n\n7. In Section 5.3 (and Figure 4), the two ablation studies are done on which dataset? Or is it averaged over replications or datasets-- what do error bars represent? It could have been clarified in the figure caption." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper presents an efficient RNA secondary structure design method using a collective design approach, with empirical evaluations provided on two RNA design benchmark datasets.\n- The paper demonstrates the potential of cooperative MARL approaches to RNA design tasks.\n- The proposed approach addresses a real-world problem, validated through experiments on real-world datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the challenge of efficient and scalable RNA secondary structure design. Designing RNA sequences that reliably fold into specified structures is difficult due to the complexity of the combinatorial search space. The paper proposes a collective RNA design approach called CocoRNA which uses cooperative multi-agent reinforcement learning. CocoRNA designs RNA sequences by decomposing the design task into subtasks assigned to multiple agents." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main contribution of the proposed method is presented as designing RNA secondary structure as a \"collective design\" problem. However, the contribution is not novel as the recent work [1] already introduces such a collective design idea, that is, efficiently designing biological sequences using cooperative design framed as a cooperative game between players (here it is called agents), and [1] should be cited in the relevant work section. Based on this, it should be further clarified that what is the main source of benefit of using a MARL method, instead of performing collective design directly using combinatorial optimization as in [1]? What are the key differences and advantages of your method over [1]?\n\n2. The paper lacks a discussion or a theoretical analysis of the convergence of the distributed policies to the global optimum. In the abstract and also in lines 126-129, it is stated that CocoRNA enables such a convergence, however, there is no guarantee or analysis that supports this claim. I would suggest that either a theoretical analysis or some clear discussion/intuition should be provided in the paper or the claims should be modified.\n\n3. It would support the empirical performance of CocoRNA better if an analysis of the performance of CocoRNA under sparse and delayed rewards is presented. Instead of using the reward function in equation (12), how would the performance change under sparse reward (e.g., $R_t$ = {$0$ if $H_t > 0$; $C$ otherwise})?\n\n4. While motivating CocoRNA in the Introduction, the authors state that RL-based methods do not exploit learning to generalize across different target structures (line 62). CocoRNA is stated to mitigate problems of (1) the curse of dimensionality and (2) sparse rewards (although not supported enough for (2)). However, the paper does not present how CocoRNA generalizes across (3) different target structures.\n\n5. I think the related work section (2.1) should be restructured. There should be a section for RL-based methods that are used to design biological sequences such as [4,5]. These are potential SOTA baselines for CocoRNA. In addition, a discussion on why MARL is needed over these RL methods would clarify the contribution within RL & biological sequence design context. Instead of providing general MARL works in detail, this section would have provided the flow from RL to MARL within the problem context.\n\t\n6. An ablation study on the grouping of players would be helpful in showing the effectiveness of CocoRNA regarding the interdependence of nucleotides at different positions. The paper presents only agent per position (n many agents) analysis. How does the performance of CocoRNA change with respect to the agent size?\n\n7. Different than the point above, an ablation study considering the decomposition scheme such as position + structure assigned to an agent would have shown the proposed method's flexibility regarding decomposition choices. This is a more specific decomposition than only structure-based decomposition.\n\n8. Regarding Section 4.4, in Hindsight Experience Replay (HER), additional goals are used to store additional episodes in the replay buffer to deal with sparse reward environments. Hence, the goal influences actions, but not the environment dynamics. In Search-augmented Experience Replay (SAER), how are the goals defined? How are additional goals for the replay sampled? It is not clear to relate SAER to HER and not provide clear explanations. I think to better motivate CocoRNA's sample efficiency and to build a better connection with HER, SAER should have experimented on a sparse reward environment. \n\n9. The empirical evaluation is done against a limited amount (and type) of baselines. The proposed method is compared only against RL-based methods. The paper only cites antaRNA [1] and MCTS-RNA [2] approaches, however, these search-based fundamental approaches should be included as baselines; which would support CocoRNA's performance against a diverse set of baselines. Furthermore, another valid baseline from literature that combines RL with local search [3] could have been considered.\n\n10. It would have supported the generalizability of the proposed approach better if an empirical analysis on any other biological sequence design domain such as protein design (which has some benchmark datasets e.g. GB1) was presented.\n\n11. For a thorough analysis, training performance could have been provided. Further, an ablation over the trained policy (e.g. under different training steps or hyperparameters) could have demonstrated the effectiveness of CocoRNA under various training settings.\n\n12. There is no clear discussion on the limitations and potential drawbacks of the method.\n\n**Minor:**\n\n- In line 114, the reference (Eastman et al…) is doubled. \\cite{} should be corrected to avoid repetition.\n\n- Regarding notations, instead of using the notation k for several places, another symbol could be used. Specifically, it is used both in equation (5) to denote the subsequence of nucleotides and in equations (7), (9) to denote cumulative future rewards.\n\n> [1] Bal, M. I., Sessa, P. G., Mutny, M., & Krause, A. (2023). Optimistic Games for Combinatorial Bayesian Optimization with Applications to Protein Design. NeurIPS 2023 Workshop on Adaptive Experimental Design and Active Learning in the Real World. https://openreview.net/forum?id=ScOvmGz4xH (or alternatively: http://arxiv.org/abs/2409.18582)\n\n> [2] Kleinkauf, R., Houwaart, T., Backofen, R., & Mann, M. (2015). antaRNA–Multi-objective inverse folding of pseudoknot RNA using ant-colony optimization. BMC bioinformatics, 16, 1-7.\n\n> [3] Yang, X., Yoshizoe, K., Taneda, A., & Tsuda, K. (2017). RNA inverse folding using Monte Carlo tree search. BMC bioinformatics, 18, 1-12.\n\n> [4] Eastman, P., Shi, J., Ramsundar, B., & Pande, V. S. (2018). Solving the RNA design problem with reinforcement learning. PLoS computational biology, 14(6), e1006176.\n\t\n> [5] Angermueller, C., Dohan, D., Belanger, D., Deshpande, R., Murphy, K., & Colwell, L. (2019). Model-based reinforcement learning for biological sequence design. In International conference on learning representations. (ICLR)\n\n> [6] Feng, L., Nouri, P., Muni, A., Bengio, Y., & Bacon, P. L. (2022). Designing biological sequences via meta-reinforcement learning and bayesian optimization. arXiv preprint arXiv:2209.06259." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1.\tIt is not clear to the reviewer what is the output of the actor network. Is it the nucleotide type for 1 position or for all positions associated with that agent?\n2.\tIs code available for the proposed algorithm at an anonymous link?\n3.\tThe reviewer suggests adding the part highlighting the difference from how the experience replay is applied from Appendix A to the manuscript.\n4.\tThe authors mention two possible decompositions: (i) position-based; (ii) structure-based. In the experiments section, it seems that only position-based is mentioned. Are structure-based decomposition results also shown in the manuscript?\n5.\tThe proposed method is compared with other RL-based baselines. It would be interesting to compare the proposed method to the other generative-based model baseline (Patil et al, 2024) having a small comparison regarding success rates and discussing which sequences can't be predicted by the generative-based model baseline because of their sequence length.\n6.\tFrom the reviewer's understanding, with n=4, four individual policies are trained. In this scenario, would not be the case to use four shared policies during training (with shared weights). It would be interesting to also have this ablation study.\n7.\tFor the experiments, a random split of the datasets was performed. Similarly to protein-related tasks, a split based on hamming distances of structures to check the generalization capabilities of the policies would be desired in the reviewer’s opinion.\n8.\tIt would be interesting to discuss the trade-offs between the proposed methodology and other generative-based alternatives such as graph neural networks using graph-based representations. As other parts of the structure might also affect the design of the sequence, having partial observations might not give enough information even with a centralized critic. For this, using a GNN architecture for decoding or a GNN-based policy for the RL methodology could alleviate this issue." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tA novel and interesting MARL-based methodology is proposed for RNA secondary structure design.\n2.\tA search-based augmented experience replay technique is proposed inspired by HER.\n3.\tThe use of multiple policies improves sample efficiency issues and, given enough computational resources, also improves computational efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper tackles the RNA secondary structure design problem proposing a novel approach using cooperative multi-agent RL. Multiple policies are jointly trained to design the sequence for parts of the RNA structure and a centralized critic is used to maximize the reward of the entire final sequence. The authors show that this methodology improves sample efficiency and computational efficiency while improving over other traditional RL-based baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe framework is still model-based and relies on repeatedly applying folding algorithms for reward calculation.\n2.\tUsing the position-based decomposition, it is hard to justify having different local policies and not a shared policy. At least an ablation seems needed for that.\n3.\tGiven the current manuscript, it is hard for readers to reproduce it, code is not available, there are some details that need additional information, and the random dataset splits might leak data." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "please see weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) **Significance of the Problem**: The paper tackles the complex challenge of RNA secondary structure design, which holds significant implications in fields like synthetic biology and gene regulation. By addressing the combinatorial nature of RNA design and the need for efficient, scalable methods, the authors provide a contribution to computational biology.\n\n2) **Clear and Well-Structured Presentation**: The paper is well-organized, with each section logically progressing from the problem statement to methodology and experimental evaluation. The clear exposition of the multi-agent reinforcement learning framework, algorithmic details, and ablation studies makes it easy to understand both the innovation and the practical execution of the proposed approach.\n\n3) **Robust Experimental Results**: The experimental results presented on the Rfam and Eterna100-v2 datasets are comprehensive and demonstrate COCORNA’s superior performance over existing methods in terms of success rate and design time. The ablation studies further substantiate the model's robustness, providing evidence that each component of the method contributes to the overall improvement.\n\n4) **Intuitive and Effective Approach**: The proposed multi-agent reinforcement learning framework is well-suited to the complex, distributed nature of RNA design tasks. The decomposition of the design problem and the centralized training with decentralized execution approach provide a practical and computationally feasible solution. The inclusion of the SAER method to improve initial data quality and learning efficiency is a thoughtful addition that strengthens the model’s effectiveness in this challenging domain." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new approach for RNA secondary structure design by leveraging cooperative MARL. The proposed method, COCORNA, breaks down the RNA design task into multiple sub-tasks managed by individual agents. Through a centralized Critic and decentralized Actor architecture, COCORNA enables these agents to cooperate, aiming to address the combinatorial complexity of RNA sequence folding.\n\nThe model was trained on RNA design tasks using a novel Search-Augmented Experience Replay (SAER) mechanism, which improves initial learning efficiency. Experimental results on the Rfam and Eterna100-v2 datasets demonstrate that COCORNA outperforms existing methods in both design time and success rate, highlighting its potential for efficient and scalable RNA sequence design without further task-specific optimization. This study showcases COCORNA as a promising tool for addressing complex biological sequence design challenges through MARL." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) **Lack of Visual Aids for Method Explanation**: The paper lacks visual illustrations of the proposed method, which is a drawback given the complexity of multi-agent reinforcement learning and RNA secondary structure design. Effective diagrams and flowcharts could have greatly enhanced the readability and accessibility of the methodology, particularly for readers unfamiliar with MARL frameworks in biological sequence design. Including such visuals would improve the reader's ability to grasp the significance and innovative aspects of this work.\n\n2) **Limited Novelty**: A significant concern lies in the limited novelty of the approach. The paper formulates RNA design as an MARL problem and primarily applies established MARL methods to solve it. The novelty of specific components, such as the reward function and the Search-Augmented Experience Replay (SAER) module, also seems limited. It would strengthen the work if the authors provided more innovative, task-specific adaptations or insights that build on existing methods in a unique way.\n\n3) **Need for Deeper Conceptual Insights**: The paper would benefit from more in-depth conceptual insights into the unique challenges and opportunities specific to RNA design in the context of MARL. For instance, an analysis of how different decomposition approaches (e.g., position-based vs. structure-type-based) impact learning, or a discussion on task-specific challenges in RNA sequence alignment, would offer valuable perspectives. Such insights could highlight the authors’ deep understanding of the problem and provide a stronger foundation for the applicability and potential extensions of COCORNA." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the weaknessnes above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The authors efficiently decompose the complex RNA design problem into mutiple sub-tasks. These tasks are allocated to cooperative agents to solve collaboratively. They introduce a search-augmented experience replay method to improve learning\nefficiency, which improves the efficiency of RNA design. The proposed method significantly outperforms existing methods in terms of both design time and success rate." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel collective RNA design method based on cooperative MARL to solve the RNA secondary structure design problem. Empirical results demonstrate the outperformance of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper is more like an application of MARL in RNA design. The contributions to MARL method to solve the specific issues when applying MARL in RNA design should be stated clearly.\n2. Besides the design of observations and reward functions, the authors should provide more explanations on how the agents cooperatively to sovle the RNA design task. \n3. The authors do not discuss the limitations of the proposed method." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose CocoRNA, a cooperative multi-agent reinforcement learning framework for RNA secondary structure design, which achieves better solving efficiency and higher success rates." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024cocorna,\ntitle={Coco{RNA}: Collective {RNA} Design with Cooperative Multi-agent Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4JZ56UVJYf},\nnote={under review}\n}" }, "abstract": { "value": "Ribonucleic acid (RNA) plays a crucial role in various biological functions, and designing sequences that reliably fold into specified structures remains a significant challenge in computational biology. Existing methods often struggle with efficiency and scalability, as they require extensive search or optimization to tackle this complex combinatorial problem. In this paper, we propose CocoRNA, a collective RNA design method using cooperative multi-agent reinforcement learning, for the RNA secondary structure design problem. CocoRNA decomposes the RNA design task into multiple sub-tasks, which are assigned to multiple agents to solve collaboratively, alleviating the challenges of the curse of dimensionality as well as the issues of sparse and delayed rewards. By employing a centralized Critic network and leveraging global information during training, we promote cooperation among agents, enabling the distributed policies to converge toward the global optimum and resulting in a high-quality collective RNA design policy. The trained model is capable of completing RNA secondary structure design with less time and fewer steps, without requiring further training or search on new tasks. We evaluate CocoRNA on the Rfam dataset and the Eterna100 benchmark. Experimental results demonstrate that CocoRNA outperforms existing algorithms in terms of design time and success rate, highlighting its practicality and effectiveness." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multi-agent", "reinforcement learning", "RNA design" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/698348cff75e900a5ee741095e8c02700001e3c7.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "CocoRNA: Collective RNA Design with Cooperative Multi-agent Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper is clearly written and easy to follow, with a well-articulated description of the proposed approach. The visualizations are also clear and supportive of the narrative.\n* The proposed method demonstrates superior performance compared to baseline methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces CR-Planner, a method combining critic-guided planning with retrieval augmentation to tackle reasoning-heavy tasks. The primary contribution of this work is the use of fine-tuned critic models to assist in action selection during inference, with training data for the critic models derived from offline Monte Carlo Tree Search (MCTS) data collection. Experimental results indicate that the proposed method outperforms baseline approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* While the performance improvement over baselines is evident, it is not entirely surprising. The critic essentially distills the MCTS policy from GPT-4 and is expected to perform better than the original policy, given the presence of a reliable external verifier. Moreover, the comparison between CR-Planner and other baselines may not be fair, as CR-Planner likely incurs significantly higher computational costs due to the extensive computations required for MCTS offline data collection.\n* The training prompts for the critic model appear to be identical to the test tasks, which raises concerns about the generalizability of the learned critic. It would be beneficial to demonstrate the performance of the learned critic on different datasets to establish its ability to generalize beyond the training data. Otherwise, the model may merely approximate the offline MCTS performance.\n* The authors should also include the performance metrics of the vanilla MCTS policy to provide a baseline comparison. This would help quantify the extent of performance degradation when using the learned neural approximator (i.e., the critic).\n* To this matter, the technical contributions presented in this work seem to be limited.\n\nOverall Assessment:\nWhile the paper presents a well-written and clear description of CR-Planner and shows promising performance improvements, the evaluation lacks important details regarding computational fairness and generalizability, as well as its technical contribution limitation. Consequently, I recommend rejection to this submission." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "uestion 1: Although decomposing the planning process into three stages—Reason, GenQuery, and Retrieve—seems intuitively reasonable, it appears that there are no experimental results provided to validate this approach. For example, is there a need to fine-tune an additional critic for evaluating and executing the sampling answers for the \"Standard\" method in baselines? Can the authors provide such results and compare them with those of the CR-Planner?\n\nQuestion 2: Is there indeed a necessity to fine-tune a critic for Sub-Goal Selection? It seems that there is a certain temporal relationship among the stages of Reason, GenQuery, and Retrieve: the system must first reason, then generate a query based on the reasoning result, and finally retrieve the result based on the generated query. Are there scenarios where this temporal sequence does not apply?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Strength 1: The paper aims to solve challenging reasoning problems, such as those found in programming and theorem-driven mathematical reasoning, which is a meaningful endeavor. Additionally, the experimental performance that exceeds baselines by an average of 10.06% is striking.\nStrength 2: The motivation to use fine-tuned critics for evaluating planning answers is intuitively reasonable. Furthermore, identifying the source of planning errors—whether in reasoning or retrieving—and proposing an approach that breaks down the planning process into Reason, GenQuery, and Retrieve is both impressive and rational.\nStrength 3: Overall, the article is well-written, with clear articulation and methodological figures that effectively convey the content." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel planning framework called Critic-guided planning with Retrieval-augmentation (CR-Planner). Due to frequent reasoning errors and irrelevant knowledge retrieval in planning, existing LLM-based methods that leverage techniques like Chain-of-Thought (CoT) and Retrieval-Augmented Generation (RAG) often struggle with reasoning-heavy and domain-knowledge-intensive tasks, such as competitive programming and mathematics. To address this, CR-Planner fine-tunes critic models using data collected from Monte Carlo Tree Search (MCTS) to evaluate and execute answers at each stage of planning, thereby enhancing the reasoning capabilities of planning systems. Experiments conducted on challenging tasks that require domain knowledge and reasoning, including competitive programming, theorem-driven mathematical reasoning, and complex domain retrieval problems, demonstrate that CR-Planner outperforms the baseline by an average of 10.06%." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weakness 1: The training data used for fine-tuning the Critic is collected via MCTS, which imposes a heavy computational load. Moreover, the nodes generated during MCTS are sampled from LLM. Consequently, the process of labeling data by MCTS and LLM is costly.\nWeakness 2: Intuitively, there is a temporal relationship among the three sub-goals: Reason, GenQuery, and Retrieve. The system must first engage in reasoning, then generate a query based on the reasoning result, and finally retrieve the result based on the generated query. Therefore, the idea of fine-tuning a critic for sub-goal selection seems somewhat unnecessary." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. Could the authors clarify the volume and scope of data required for effective critic model training?\n - Is the data collection process designed to be online or offline?\n - Are there distinct types of data required for training each critic model?\n - To what extent is the trained critic model generalizable across different tasks or domains?\n\n2. The paper (lines 149-151) notes that MCTS is used to gather training data that captures global reward information. However, in large language models (LLMs) where the action space is effectively infinite, selecting optimal actions solely based on a value function may be impractical, as an optimal action cannot be deduced simply by comparing all possibilities.\n - How do the authors address the challenges posed by this expansive action space?\n - Given that value functions alone may not guide optimal action selection in LLM settings, what alternative strategies or adaptations are employed?\n\n3. The paper mentions temperature sampling for the REASON and GENQUERY actions and top-k candidates for the RETRIEVE action, yet it is unclear how these discrete and continuous spaces are integrated.\n - Could the authors provide a clearer description of the combined action space used in execution selection, beyond the reference to Appendix C and Table 6?\n\n4. In Equation 1, the policy selects actions based on the reward function rather than the value or action-value function, which may not yield the optimal trajectory.\n - Could the authors explain the rationale for using the reward function over other potential selection criteria that might better optimize the full trajectory?\n\n5. The section \"Collecting data via MCTS\" outlines MCTS with some modified definitions, yet lacks specifics.\n - How does the approach balance exploration within unknown search spaces?\n - Is data collected in a single batch or in increments, and what is the distribution among different data types collected?\n\n6. The experiment section does not address other relevant work on MCTS in reasoning contexts.\n - Could the authors discuss their approach in relation to similar works, such as:\n - *Q*: Improving Multi-step Reasoning for LLMs with Deliberative Planning. http://arxiv.org/abs/2406.14283\n - Quiet-STaR: Language Models Can Teach Themselves to Think Before Speaking. http://arxiv.org/abs/2403.09629\n - Monte Carlo Tree Search Boosts Reasoning via Iterative Preference Learning. http://arxiv.org/abs/2405.00451" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The paper evaluates CR-Planner on diverse, challenging domains, demonstrating substantial improvements in both reasoning-intensive and retrieval-intensive tasks.\n2. By utilizing MCTS for data generation, CR-Planner effectively addresses data scarcity in training critic models, which is a significant practical contribution." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents CR-Planner (Critic-guided Planning with Retrieval-Augmentation), a framework designed to enhance reasoning and factual correctness in large language models (LLMs) by employing critic models for sub-goal selection and execution guidance. Traditional methods like chain-of-thought (CoT) reasoning and retrieval-augmented generation (RAG) have proven effective in structured reasoning tasks, but struggle with complex challenges in areas such as competitive programming and theorem-driven mathematics. CR-Planner addresses these challenges by using fine-tuned critic models to guide reasoning and retrieval, enabling a systematic selection and execution of sub-goals. Monte Carlo Tree Search (MCTS) is utilized to train data collection. Experiments are across competitive programming, math problems, and domain retrieval." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. A key concern with the proposed approach is the soundness of using a critic model for action selection. While the critic model is trained to optimize the reward function, this function alone does not constitute an actionable policy, potentially limiting the framework's effectiveness. This limitation is further complicated by the mixed action space, which combines both continuous and discrete elements in execution.\n\n2. The paper would benefit from substantial revisions to provide clearer explanations of each component. In particular, detailed descriptions are needed for the problem formulation, the critic model’s training methods, and the use of pairwise ranking loss, along with a comprehensive equation and explanation. Additionally, the data collection process for each critic model requires clarification to enhance transparency and reproducibility.\n\n3. The novelty of the proposed approach is limited, as the use of MCTS for reasoning is well-established in existing literature. The lack of comparisons with related works on MCTS-based reasoning approaches makes it difficult to assess the unique contributions of this paper within the field." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "What were the domain-specific tasks used for training the critics?\n\nGiven the need for training domain-specific critics, why not simply fine-tune a domain-specific generator? Perhaps the explicit planning, following the structure of CR-Planner gives a bigger performance boost, but empirically verifying this is important.\n\nWhat was used to motivate the separate fine-tuning of four domain-specific critics? Did you try creating one dataset for all of the critic tasks and fine-tuning a single critic? Results on this would be useful." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The use of stepwise critics seem to address an important problem: guiding the generation of long plans and reasoning traces without compounding errors or other degenerative behaviour. This is especially challenging given the inclusion of RAG, which can help simultaneously address knowledge-intensive aspects of tasks.\n\nThe method is thoroughly described and overall very well-communicated.\n\nThe results show performance gains over relevant benchmarks, although it is unclear whether the comparisons are fair given domain-specific fine-tuning budgets (see 'Weaknesses')." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper structures response planning for reasoning-intensive tasks as a series of reasoning steps and knowledge retrieval steps. To guide this complex plan formation, reasoning, query generation and retrieval critics are used to score the execution of sub-goals. A separate sub-goal selection critic is used for informing sub-goal selection. These critics are obtained by fine-tuning an open-source reward model on data collected via MCTS. Evaluations are performed on a competitive programming benchmark, a theorem-driven math benchmark and two reasoning-heavy domain retrieval benchmarks. Further analysis and comparisons are performed on the programming benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Nowhere is it stated what tasks are used for training the critic models. I can therefore only assume that training subsets of the benchmarks used for evaluation are used for training the critics. If so, this makes comparison to the current baselines unfair. \n\nThe reliance on domain-specific fine-tuning of several critic models is a notable limitation, especially given that the paper does not compare to simply fine-tuning the generating model with the same budget.\n\nThere is no ablation of the individual critics. The paper shows the importance of fine-tuned, domain-specific critics, but does not ablate the inclusion of a critic for each of the sub-goal execution types and sub-goal selection. \n\nThere are no ablations or experiments showing the impact, in terms of sample efficiency, of MCTS on critic fine-tuning." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present critic-guided planning with retrieval-augmentation (CR-Planner), a novel framework that leverages fine-tuned critic models to guide both reasoning and retrieval processes through planning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024can,\ntitle={Can We Further Elicit Reasoning in {LLM}s? Critic-Guided Planning with Retrieval-Augmentation for Solving Challenging Tasks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4JfFW7d1gu},\nnote={under review}\n}" }, "abstract": { "value": "State-of-the-art large language models (LLMs) exhibit impressive problem-solving capabilities but may struggle with complex reasoning and factual correctness. Existing methods harness the strengths of chain-of-thought (CoT) and retrieval-augmented generation (RAG) to decompose a complex problem into simpler steps and apply retrieval to improve factual correctness. These methods work well on straightforward reasoning tasks but often falter on challenging tasks such as competitive programming and mathematics, due to frequent reasoning errors and irrelevant knowledge retrieval. To address this, we introduce Critic-guided planning with Retrieval-augmentation, CR-Planner, a novel framework that leverages fine-tuned critic models to guide both reasoning and retrieval processes through planning.\nCR-Planner solves a problem by iteratively selecting and executing sub-goals. Initially, it identifies the most promising sub-goal from reasoning, query generation, and retrieval, guided by rewards given by a critic model named sub-goal critic. It then executes this sub-goal through sampling and selecting the optimal output based on evaluations from another critic model named execution critic.\nThis iterative process, informed by retrieved information and critic models, enables CR-Planner to effectively navigate the solution space towards the final answer.\nWe employ Monte Carlo Tree Search (MCTS) to collect the data for training the critic models, allowing for a systematic exploration of action sequences and their long-term impacts.\nWe validate CR-Planner on challenging domain-knowledge-intensive and reasoning-heavy tasks, including competitive programming, theorem-driven math reasoning, and complex domain retrieval problems. Our experiments demonstrate that CR-Planner significantly outperforms baselines, highlighting its effectiveness in addressing challenging problems by improving both reasoning and retrieval." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "reasoning", "planning", "retrieval-augmented generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/32a92a4c5644af406f32a31b8cb65fb2d468d85d.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Can We Further Elicit Reasoning in LLMs? Critic-Guided Planning with Retrieval-Augmentation for Solving Challenging Tasks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I hope to confirm with the authors that if you claim supremacy in any metric of the proposed method compared to traditional FEM solvers?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The writing is clear and detailed.\n- The experiments are rich in problem types, specific difficulties, and baseline comparisons." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose training neural PDE solvers by variable separation and random sampling of neural network weights. The neural network ansatz is utilized for the spatial domain, and the system evolving in time is solved by classical ODE solvers. Extreme learning machines and adaptive sampling techniques (SWIM) are applied for better training efficiency. An SVD layer is introduced to improve the condition number of the associated ODE. It is claimed that the proposed method outperforms PINN by 1 to 5 orders of magnitude in time efficiency and accuracy, for PDEs including Advection, Euler-Bernoulli, Nonlinear diffusion, and Burgers'." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Meaning no offense, but I think researchers in AI4PDE with more AI background will think of this work as a huge step backward. The essence of deep neural networks is their surprisingly good performance in approximating high-dimensional functions, and the efficiency of backpropagation in implementing neural networks with huge amounts of parameters. Surely there are still issues even if we can obtain the gradients cheaply, but zeroth-order optimization, according to my personal judgment, cannot be the solution because it will only scale poorly.\n- For the experiments, the spatial dimension is 1 or 2, and small in range. It would be interesting to see some results for problems huge in space." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "What is the n-width of the problems considered (as given by the spectral decay of the snapshot matrix)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The presentation is clear and the literature review is thorough and provides a good introduction.\n- The method shows strong performance on the chosen benchmarks" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present a method of solving PDEs by parameterizing solutions fields with neural networks whose parameters depend on time. The integration scheme solves for the last layer cofficents. The basis functions, induced by the inner parameters, are generated via a data driven or data agnostic way." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The motivation for the method is not totally clear. Introducing neural networks to solve PDEs where the parameterization evolves nonlinearly in time is motivated by breaking the kolmogorov n-width as in [1,2,3,4]. In this work the parameterization still evolves linearly in time. The neural network is only used to choose a basis. So it is unclear why one would pick this method over a traditional solver, which are extremely well understood in terms of convergence properties and accuracy. It seems to me the only reason would be to deal with complicated geometries? If so currently the paper does not devote enough attention to arguing and demonstrating this advantage. Additionally for these reason the comparison with PINNs is ill-chosen. The most appropriate comparison would be to traditional methods which also evolve linearly in time. While comparison is made to a finite-element method, this is not the best choice for some of the problems present. For the data-agnostic a reasonable spectral method should also be compared to and for the data-dependent method POD should be compared to.\n\nIt would be helpful to:\n\n- make more explicit the advantages over traditional, show this advantages clearly in the experiments.\n- add a comparison to a spectral methods for the data-agnostic case.\n- add a comparison to POD for the data-dependent case.\n\n\n[1] Evolutional Deep Neural Networks. Du et al.\n[2] Randomized sparse neural galerkin schemes for solving evoluation equations with deep networks. Berman et al.\n[3] Positional embeddings for solving PDEs with evolutional deep neural networks. Kast et al.\n[4] Breaking the Kolmogorov Barrier with Nonlinear Model Reduction. B Peherstorfer." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Could the authors clarify the absence of experiments involving higher-dimensional PDEs? Given the introduction’s emphasis on the limitations of mesh-based methods—particularly their impracticality in complex domains and high-dimensional spaces—it would be valuable to see examples where the proposed method effectively addresses these challenges. Higher-dimensional cases are particularly relevant to machine learning applications, where scalability in complex domains is critical." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors propose a backpropagation-free method that leverages random sampling techniques like Extreme Learning Machines (ELM) and Sampling Where It Matters (SWIM) to address the inefficiencies of traditional backpropagation, especially for complex time-dependent PDEs.\n\n2. The paper reports significant speed gains in training time, with improvements of up to 5 orders of magnitude over standard PINN approaches.\n\n3. Specialized handling of boundary conditions and separation of variables for time-dependent PDEs are some of the contributions that could impact future neural PDE solvers.\n\n4. The authors demonstrate extensive benchmarking across a range of PDEs with different challenges, showing superior performance in terms of speed and accuracy.\n\n5. The paper is well-written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a method for training neural PDE solvers without backpropagation, which aims to improve efficiency in solving time-dependent partial differential equations (PDEs). The authors integrate two main ideas: separating space and time variables and randomly sampling weights and biases in hidden layers. By reformulating the PDE as an ordinary differential equation (ODE) using neural networks for spatial components, they leverage traditional ODE solvers for time evolution. The approach is benchmarked against standard backpropagation-based Physics-Informed Neural Networks (PINNs). It shows improvements in accuracy and speed on complex PDEs involving non-linearities, high-frequency temporal dynamics, and shocks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors have mentioned the limitations of their method and share possible directions to follow in future work." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Experiments.\n- The boundary conditions are approximated \busing a boundary-compliant layer. For instance, in the case of periodic BC, the authors approximate $\\sin(kx)$ and $\\cos(kx)$ by applying a linear transformation to the basis function. However, this raises the question: what advantage does the proposed method offer compared to just using $\\sin(kx)$ and $\\cos(kx)$ as basis functions, or P1, P2 basis functions in FEM? A numerical comparison in this scenario would be helpful.\n- It appears that $C(t)$ is calculated by multiplying the pseudo inverse of feature matrix $[\\Phi(X),1]$, where $X$ contains all the collocation points. In cases of high dimensionality $d>>1$ where $N>>1$ to cover the entire domain, there may be significant computational demands. Further discussion and experiments on the computational cost in high-dimensional settings would be needed.\n\n2. Theoretical contributions\n- Does ELM possess a universal approximation property? If so, can this be generalized to the neural PDE solver setting?\n\n3. Limited applications\n- As the authors mention, the method cannot be applied to grey-box or inverse problem settings. Given this, what advantage does the mesh-free nature provide?\n- If the pseudo-inverse calculation for $[\\Phi(X),1]$ becomes computationally expensive, especially in high-dimensional problems, what practical benefit does mesh-free implementation offer?\n- Overall, what advantages does the proposed method offer over mesh-based approaches? In many cases presented in the paper, mesh-based methods achieve superior test accuracy with shorter training(computing) times." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Significantly lower relative error compared to PINNs\n- Substantially faster training speed than PINNs\n- Achieves both improvements without backpropagation while retaining a mesh-free approach" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a backpropagation-free training algorithm for a neural partial differential equation solver, utilizing the Extreme Learning Machine (ELM) framework. The method reformulates the partial differential equation (PDE) as an ordinary differential equation (ODE) problem through the separation of variables, which is then solved using classical ODE solvers. Numerical experiments show that the proposed method outperforms traditional PINNs in both test accuracy and training speed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experiments are insufficient to fully support the authors' claims.\n- The paper lacks theoretical contributions.\n- The proposed method has a limited range of applications, which restricts its overall contribution." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Is the proposed method able to handle PDEs with higher-order time derivatives?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The proposed method is novel and provides a distinct method to solve time-dependent PDEs other than classical numerical methods and PINNs.\n- The experiment results show that te proposed method outperforms PINNs by orders of magnitude of accuracy; the accuracy is even comparable to classical numerical solvers.\n- The authors also provide techniques to satisfy boundary conditions and improve the condition number of the associated ODE." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to use a hybrid framework consisting of a neural network ansatz and a classical ODE solver to solve typical time-dependent PDEs. Specifically, the neural network ansatz features separation of spatial and temporal variables, and the parameters of this network is randomly sampled rather than trained with back propagation. Numerical experiments are conducted to verify the high accuracy and reduced training time of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper should consider add more backgrounds about the random-sampling methods of neural network weights. Without back-propagation, how does this random-sampling of weights influence the final solution of the proposed method? As can be seen in Table 2, the standard deviations of your proposed method is relatively larger than PINNs, although the accuracy is significantly better. \n- The paper should add some ablation studies to provide more insight about each component of the proposed method. For example, the necessity of the SVD layer, the influence of number of hidden neurons.\n- It would add more practicabillity of the proposed method by providing more detailed comparisons between ELM-ODE and SWIM-ODE. Is one strategy better than another, or one should choose between these two strategies based on the PDE to tackle?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a backpropagation-free algorithm to train neural PDE solvers for time-dependent problems." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024backpropagationfree,\ntitle={Backpropagation-free training of neural {PDE} solvers for time-dependent problems},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4KKqHIb4iG},\nnote={under review}\n}" }, "abstract": { "value": "Approximating solutions to time-dependent Partial Differential Equations (PDEs) is one of the most important problems in computational science. Neural PDE solvers have shown promise recently because they are mesh-free and easy to implement. However, backpropagation-based training often leads to poor approximation accuracy and long training time. In particular, capturing high-frequency temporal dynamics and solving over long time spans pose significant challenges. To address these, we present an approach to training neural PDE solvers without back-propagation by integrating two key ideas: separation of space and time variables and random sampling of weights and biases of the hidden layers. We reformulate the PDE as an ordinary differential equation using a neural network ansatz, construct neural basis functions only in the spatial domain, and solve the ODE leveraging classical ODE solvers from scientific computing. We demonstrate that our backpropagation-free algorithm outperforms the iterative, gradient-based optimization of physics-informed neural networks with respect to training time and accuracy, often by several orders of magnitude using different complicated PDEs characterized by high-frequency temporal dynamics, long time span, complex spatial domain, non-linearities, and shocks." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "neural PDE solvers", "time-dependent partial differential equations", "random feature networks", "backpropagation-free training" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/00d58aa9d9ea0d94648b58effd93cbe51a0c2d31.pdf" }, "presentation": null, "primary_area": { "value": "learning on time series and dynamical systems" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/acf7faa5e4b7bdb1629658074d00b0b05b59ef3f.zip" }, "title": { "value": "Backpropagation-free training of neural PDE solvers for time-dependent problems" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. Could the authors clarify the computational cost associated with generating the dynamic mask at each iteration and how it compares to MeZO in practice? such as a wall-clock time wise comparison similar to Figure 1\n\n2. How sensitive is Sparse-MeZO’s performance to the choice of layer-wise sparsity threshold, and what considerations guided the threshold selection?\n\n3. Can you provide detailed threshold hyperparameters for reproducibility?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Incorporating sparsity and MeZO is an interesting direction for performance improvement. \n\n2. The paper offers good empirical validation, showing clear improvements over baseline methods across a range of fine-tuning tasks.\n\n3. The paper is structured logically, making it easy to follow the motivation and methodology" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes Sparse-MeZO, a memory-efficient zeroth-order optimization (ZO) technique for fine-tuning LLM by selectively optimizing a subset of parameters, rather than all. Based on MeZO, Sparse-MeZO introduces a sparse mask that targets smaller, noise-resilient weights, thereby mitigating gradient estimation noise, improving convergence speed, and reducing memory usage. Key results demonstrate Sparse-MeZO’s efficiency, achieving faster convergence and improved performance over MeZO, with the ability to fine-tune models like LLaMA-30b on limited hardware resources (e.g., a single A100 GPU)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper lacks details on how thresholds for selecting small-magnitude weights are determined, how they vary across tasks, and whether they require re-tuning for different settings.\n\n2. It’s not clear if masking and selecting small-magnitude weights specifically benefit zeroth-order optimization more than generic fine-tuning (first-order methods). Since subset selection often improves generalization, it would be needed to evaluate this effect.\n\n3. There is no specific numbers regarding the computational overhead introduced by dynamic masking and threshold calculation.\n\n4. The motivation for why zeroth-order optimization particularly benefits from small-magnitude weights lacks enough theoretical or empirical support. An ablation study showing this effect more clearly would strengthen the argument." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Mask Updates per Iteration: Are the masks updated in each iteration? The concept of dynamic masking appears similar to Dynamic Sparse Training (DST), where the frequency of mask updates is a critical factor affecting performance [1,2]\n\n- Can Sparse MeZO be integrated with LoRA, similar to how MeZO has been combined with LoRA in previous work?\n\n- When claiming the 3.5× speed-up in Figure 1, are both methods using the same learning rate, or does the speed-up depend on differing learning rates?\n\n- Determining the masks is a key component of this work, and the authors use a simple magnitude-based method. Do you believe that employing more advanced methods like Wanda [3] or SparseGPT [4] could further enhance performance?\n\n[1] Do we actually need dense over-parameterization? in-time over-parameterization in sparse training, ICML 2021\n\n[2] Rigging the Lottery: Making All Tickets Winners， ICML 2020\n\n[3] A Simple and Effective Pruning Approach for Large Language Models, ICLR2024\n\n[4] SparseGPT: Massive Language Models Can Be Accurately Pruned in One-Shot" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Interesting Idea: Proposes Sparse MeZO, selectively applying zeroth-order optimization to improve memory efficiency.- \n- Good Performance: Achieves significant gains in accuracy and convergence speed over existing methods.\n- Simple but Efficient Algorithm: Offers a straightforward yet effective approach for fine-tuning large models on limited hardware." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Sparse MeZO, a memory-efficient optimization method for fine-tuning large language models. By selectively applying zeroth-order updates to a carefully chosen subset of parameters, it reduces estimation errors inherent in zeroth-order methods. This approach improves both performance and convergence speed without increasing memory usage, enabling efficient fine-tuning of large models like LLaMA-30b on limited hardware resources." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Some details need to be clarify, as explained in Questions" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* Given a fixed training budget does SparseMeZO outperform fine-tuning with adapters (+ gradient accumulation)?\n\n* I think it would be insightful to add a comparison between the gradient updates with standard SGD optimization and SparseMeZO on a smaller scale model (say of size ~1B parameters)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The fact that instability is caused by the noisy optimization of weights with large magnitude appears to be a useful practical insight. \n\n* SparseMeZO consistently improves upon dense MeZO optimization in different setups. \n\n* The proposed memory-efficient implementation of S-MeZO incurs almost zero memory overheads relative to original MeZO and allows tuning large models on a single GPU." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes zeroth order memory-efficient optimization with sparse updates of model parameters for fine-tuning. Sparse updates are shown to facilitate better convergence than vanilla MeZO. The introduced method is evaluated on several tasks from SuperGLUE and several model architectures - Llama-1, Mistral, OPT." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The overall contribution is incremental as the work adds additional sparsification steps on top of the MeZO optimization algorithm. \n\n* Evaluation setup is outdated (as for Fall 2024). Llama-1 and Mistral-v0.1 were released more than a year ago, and current models (Llama-3, Llama-3.1, Qwen-2.5, gemma-2) have advanced dramatically in terms of quality since then. In addition, I would suggest testing the approach on a more challenging and representative fine-tuning setup, such as instruction tuning on Alpaca/OASST or any other instruction fine-tuning dataset to fully appreciate the efficacy of SparseMeZO. \n\n* Some important baselines are not taken into consideration. PEFT adapters (LoRA, DoRA) allow for significant memory reduction on optimizer states. Memory footprint on activations and gradients is not very large for relatively short sequences (which is the case for SuperGLUE tasks) and small batches. In addition, it can be further decreased by gradient accumulation at the cost of additional compute. I would suggest adding comparison with these options and reporting memory usage in Table 3." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "See above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is well motivated, the approach is novel and the results are convincing. I particularly enjoyed the writing style and clear structure of the paper, it was mostly easy to follow, enjoyable to read and the reader was not left with many questions. The proposed method is interesting, however there are open questions, which I will discuss below. Overall, I think this is a good paper that however needs to resolve several questions before being suited for publication." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper builds on the Memory-efficient Zeroth-order optimizer (MeZO) and introduces a new method, called Sparse MeZO. While MeZO is successful in training a neural network only with forward passes and by estimating the batch gradient using a finite difference scheme, there is still a significant gap between MeZO and common first-order methods. The authors aim to narrow this gap by updating only a sparse subset of the model parameters in each iteration. In particular, they propose to only update the smallest weights in each layer. They provide efficient implementations of doing so and show that Sparse MeZO performs better than multiple baselines, including MeZO." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I appreciate research in this important field, I have several concerns regarding soundness, clarity and contribution of the work, which I explain in detail below. I hope that these remarks are helpful in improving the work and I am happy to discuss my evaluation.\n\n### Soundness\n- In Figure 2b, you compare the potential loss increase on a different batch than the one used by MeZO, you then conclude that \"zeroth-order gradient estimates suffer from overfitting or noise\" (Line 189). I think to really conclude this, you would need to compare the MeZO gradient on a batch with the actual gradient on the same batch. The very same statement could be true for e.g. SGD, especially if the batch size is small and the variance of the gradient is high. I would greatly appreciate to see this figure instead or in addition to Figure 2b (potentially in appendix).\n- In Line 197, you write that \"all parameters share the same value\". What is meant by this? The formula you say that \"is used to estimate the gradient\" only gives the magnitude of the multiplier, the actual gradient vector is given by $z$, where definitely not all elements are the same. Apart from that being a false premise, you then conclude that not all parameters are optimized in the true gradient direction, which I found a surprising statement. As far as I understand it, MeZO or variants randomly sample a direction along which the gradient is estimated, constrained to that particular direction. It will with very high probability not be anywhere close to the true gradient direction, by virtue of the fact that you are random sampling the directional derivative. Please clarify this, maybe I am mistaken here. \n- In the same paragraph, you somehow derive that \"small weights are less impacted by noise corruption and can generalize better\" - it is not clear to my why you derive this conclusion, is it just by the fact that you empirically find this to work better? What is the mathematical intuition or even justification for this statement? I think to resolve this soundness issue, you have to answer to very important questions: First of all, why does it make sense to use less weights in the first place? For First-order Optimization this would be absurd, what is the difference here? Why should this work better than using all weights? And secondly, why would exactly the smallest weights have the highest impact? I could potentially follow this argument if you were actually sparsifying the computed gradient, then one could argue with Taylor approximation or the like, but you are not doing this, you are instead sparsifying the weights. Why?\n- Figure 4 highlights the effect of sparsity on the final performance. It is not clear why this curve looks like this and why that would make sense. The authors should clarify why they think that sparsity in the weights during gradient estimation helps (apart from providing empirical evidence that is performs better). What sparsity 70% performs bad, then sparsity 75% performs pretty good, and sparsity 80% performs worse again seems very arbitrary and not intuitive. In the worst case, this is just a noise artifact.\n\n\n### Clarity and Experimental Validation\n- Line 123: What is \"SPSA\"? Is this defined anywhere in the paper?\n- In Line 260, you explain how you determine the mask. Why don't you set the threshold dynamically in each iteration? It's efficient and seems like the obvious thing to do.\n- In Table 1, you report values for MeZO-LoRA. While I might have a vague idea of what that could be, this seems to be nowhere explained nor defined. Is this known from the literature or was this just missed?\n- Are the numbers in the tables reported with respect to the same overall runtime or number of iterations? If MeZO variants are faster than e.g. FT, it would be nice to see how they compare on the same runtime.\n- In the appendix (e.g. F), you refer to Lemma 3.2. What exactly is this, where is this defined? The same holds for the following appendix sections, I think these are referring to theorems/lemmata that do not exist or are not defined with the same numbering style. Please clarify.\n\n### Contribution\n- As outlined above, I think the contribution of the work is a major issue here. I fully acknowledge that Sparse MeZO achieves better results and is in that sense a meaningful contribution. However, the derivation of the method seems to be at least somewhat vague, it lacks justification apart from achieving better results. There is not much insight to gain since the paper lacks mathematical justifications, or at least intuitions. The derivation from gradient noise seems to be not very rigorous, at least to me. I hope that the authors can convince me otherwise.\n\n\n### Minor Remarks\n- Section 4.1 uses \\citet everywhere where I think \\citep is intended, this hinders readability.\n- Line 483: There is a typo, I guess it should be \"on a single A100 GPU\"." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024sparse,\ntitle={Sparse Me{ZO}: Less Parameters for Better Performance in Zeroth-Order {LLM} Fine-Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4Kw4KAoVnx},\nnote={under review}\n}" }, "abstract": { "value": "While fine-tuning large language models (LLMs) for specific tasks often yields impressive results, it comes at the cost of memory inefficiency due to back-propagation in gradient-based training. Memory-efficient Zeroth-order (MeZO) optimizers, recently proposed to address this issue, only require forward passes during training, making them more memory-friendly. However, compared with exact gradients, ZO-based gradients usually exhibit an estimation error, which can significantly hurt the optimization process, leading to slower convergence and suboptimal solutions. In addition, we find that the estimation error will hurt more when adding to large weights instead of small weights. Based on this observation, this paper introduces Sparse MeZO, a novel memory-efficient zeroth-order optimization approach that applies ZO only to a carefully chosen subset of parameters. We propose a simple yet effective parameter selection scheme that yields significant performance gains with Sparse-MeZO. Additionally, we develop a memory-optimized implementation for sparse masking, ensuring the algorithm requires only inference-level memory consumption, allowing Sparse-MeZO to fine-tune LLaMA-30b on a single A100 GPU. Experimental results illustrate that Sparse-MeZO consistently improves both performance and convergence speed over MeZO without any overhead. For example, it achieves a 9% absolute accuracy improvement and 3.5x speedup over MeZO." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Zeroth-Order Optimization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/6ec020f892bf140449f916b445c31289cd76ef71.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Sparse MeZO: Less Parameters for Better Performance in Zeroth-Order LLM Fine-Tuning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "There is no potential violation of the CoE." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses part." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The paper addresses a challenging TTA settings where only a single, unlabeled sample is given for adaptation during test time.\n- The paper adopts an interesting approach of ensemble learning to dynamically optimize a group of learners (pre-trained models), showing improved TTA performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces an intra-model ensemble learning method for single sample test-time adaption. It minimizes the cross-entropy losses between the output with the highest confidence score and all other classifier's outputs. It optimizes all trainable parameters (except for BN layers) and requires only a single sample for TTA. It achieves improved performance on corrupted datasets including CIFAR10-C, CIFAR100-C and ImageNet-C." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed algorithm offers no substantial improvement over existing ensemble learning methods. It simply combines 1) selecting the most confident prediction and 2) cross-entropy minimization of ensemble models. Technical contributions to both ensemble learning and single-sample TTA remain limited.\n- The paper lacks sufficient experimentation to demonstrate the proposed method’s effectiveness. It only compares results across different backbone architectures, without considering other baseline methods suitable for single-sample TTA, such as NOTE [1] and REALM [2]. Additionally, it does not explore alternative TTA settings, such as continual TTA, where incoming domains continuously change.\n- The experiment section (Sec. 4) requires more careful writing and clarification. For instance, it should include a clear definition and detailed description of the tuning set samples, as well as more comprehensive experiments, including ablation studies, to examine the correlation between the number of ensemble models and TTA performance.\n- In Section 4.1, the authors state that no catastrophic forgetting was observed on the ImageNet-C dataset. However, this is unlikely to be accurate since only 7,000 samples per corruption type from ImageNet-C were used for evaluation. More rigorous experiments and substantiated claims are needed.\n- As noted in the limitations, the proposed method requires significant computational resources to utilize multiple pre-trained networks. However, the paper does not provide any empirical analysis or comparison of computational cost or adaptation latency.\n\n[1] Gong et al., Note: Robust continual test-time adaptation against temporal correlation. NeurIPS'22. \\\n[2] Seto et al., REALM: Robust Entropy Adaptive Loss Minimization for Improved Single-Sample Test-Time Adaptation. WACV'24." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weaknesses above that are directly actionable. No further Qs." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The technique is simple and elegant. It generalises the typically used concept of a student/teacher setup which typically uses a single model, and is straightforward to scale in practice (by increasing the size of the ensemble)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a test-time adaptation (TTA) technique based on model ensembling. A set of models is simultaneously trained, and for each sample, the model with highest confidence is used as a teacher for all student models. Updates happen via standard cross-entropy. The authors show improvements over a non-adapted baseline model across CIFAR10/100-C and ImageNet-C." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Summary**\n\nThe presentation and investigation done in the paper is well below the bar for ICLR. There are no baselines, and the results are not well presented and contextualised. The introduction to the paper is lengthy, and should be made more crisp. The contribution statement is not accurate and needs to be adapted to what is actually shown in the paper. A lot of important controls and analysis experiments are missing to back up the claims. While I think that the general idea has potential, it needs a much better investigation to be considered for acceptance. A list of actionable points is given below. \n\nFor full transparency, while I would be happy to increase the score if my points are addressed, I doubt this is doable within the rebuttal period. Depending on the other reviews, I would already suggest that the authors consider a full revision of the paper and submission of their paper to the next conference. That being said, I think this could become a very nice paper if more work is spent on the experiments, and I would be happy to iterated with the authors during the discussion period.\n\n**Major**\n\n**W1** There is a very lengthy introduction. The method section only starts on page 5. Before, a fair amount of related work is cited (great), but without considering any of that later as methods for comparisons.\n\n**W2** A naiive baseline is omitted: What happens if a state-of-the-art adaptation technique like EATA, or even older/simpler techniques like TENT are used, but adapting $n$ models in parallel, and then ensembling of the outputs?\n\n**W3** Section 3 needs a rewrite and improvements to clarity. For example, basic metrics like the cross-entropy loss are abbreviated with an extra symbol $\\delta$, it would be better to clearly define the loss formulation used instead.\n\n**W4** The contributions reference “continual learning” (l. 130, l. 134), but there is no experiment backing this up. The reference should be removed.\n\n**W5** Claim 3 in the contributions (ll. 135-136) states that TTA requires a full batch. This is misleading or even wrong. There are certainly techniques that measure improvements when only a single sample is available (= model is adapted on that sample, and performance is measured) before the domain changes (e.g. Batch norm adaptation, or MEMO). However, in the setting considered here, model updates still accumulate on the same domain. In that setting, any TTA technique, like TENT, can be made suitable for the discussed setting by only updating every N steps (where N is the batch size), collecting the samples in the meantime.\n\n**W6** Claim 4 in the contributions is not at all corroborated by results in the paper and should be dropped.\n\n**W7** The paper needs to add recent methods to compare to. The current tables provide a lot of irrelevant details, and should be compressed. Instead of listing all different domains, it would be better to run a sufficient number of baseline models on the considered datasets to contrast the results. When doing so, it could be interesting to investigate whether the ensembling mechanism proposed is *orthogonal* to other adaptation methods: E.g., if we consider the EATA loss function for ensembling, does this improve over EATA?\n\n**W8** Analysis should be added: What happens if the number of models in the ensemble varies? How robust is this technique to common problem in cross-entropy based adaptation (model collapse) when training over long time intervals?\n\n**W9** In case the authors decide to keep the continual learning claim: How does the model perform in the continual adaptation settings used in CoTTA or EATA, and how does the model perform on long-term adaptation on CCC (Press et al., 2023)?\n\n**Minor**\n\n- Related work in l. 178: Batch norm adaptation (Schneider et al., 2020) re-estimates batch norm statistics, and this was shown to also work on single samples. Test time training (Sun et al., 2019) also considers single samples. TENT, in contrast, *additionally* performs entropy minimisation akin to the cross-entropy loss discussed in the study.\n- Figure 1 misses a legend. The color code could be adapted, and e.g. corruptions of the same type could get the same color with different marker colours. That would make the plot more interpretable than the current 15 random colours." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No ethics review needed." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* What are the key technical differences of the proposed method to simply applying gradient-based self-training to an ensemble classifier, one sample at a time? \n\n* What are the key technical differences to prior work on ensemble self-training, such at [Ghosh, 2021]?\n\n* Please clarify these claims: 1) page 3: \"we minimize the diversity of the ensemble [...] in a way that facilitates generalization of the source training domain to the new testing domain.\" What does \"facilitates generalization\" mean here? Just higher test accuracy? In what way is lower *diversity* of the ensemble an important factor for that? 2) In the conclusion: \"member models are optimized and improved like in knowledge distillation and create an upward spiral effect\". What do you mean by upward spiral effect? E.g. from Figure 3 it appears that model accuracy goes up and down over epochs. \n\n* How were the method's hyperparameters chosen, given that standard model selection is not possible in the test-time adaptation setting?\n\n* Which variant/subset of ImageNet was used? Is the \"test\" data the actual \"test\" set or the \"validation\" part?\n\n* The manuscript emphasizes a batchsize of 1, but multiple *epochs* are run on the adaptation data. Does this mean that your method must buffer all test data, or at least see it repeatedly, such that statistics can be collected? Wouldn't batch-based TTA techniques be also applicable then?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Strengths:\n* the studied problem is realistic and relevant\n* the propose method makes sense and is simple enough for practical use\n* experiments show improved accuracy by the proposed test-time-adaptation" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The submission describes a method for adapting an ensemble of classifiers to a new data distribution using only unlabeled data. Specifically, the classifiers are trained sequentially soft pseudo-labels, which are the output of the model that has the highest predicted probability for the majority class across classifiers. Experiments are performed on standard datasets (CIFAR10, CIFAR100, ImageNet) with synthetic distribution shifts for an ensemble consisting of 5 pre-trained image classification models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Unfortunately, the submission has a number of weaknesses. \n\n* incomplete method\n\nThe method works iteratively but the submission does not provide a termination condition. This is a not minor issue, as the manuscript makes it sound: without a termination condition, the method cannot be run in practice, and because no labeled data is available at test time, standard techniques, such as checking for a drop of validation accuracy, cannot be applied, either. \n\n* lack of clarity in contribution and scientific comparison to prior work\n\nThe submission almost exclusively showcases the proposed method itself, but it does not put the method into context and competition with prior or alternative techniques. This leaves the novelty and relevance of the contribution unclear. Specifically, for a publication at a scientific venue, I expect that the novel aspect of the proposed method is clearly described, contrasting it to ideas and techniques that already existed in previous works. The manuscript does not do a good enough job doing so. As related work it concentrates almost exclusively on recent papers from the application domain of test-time adaptation with deep networks. However, ensemble methods have been studied extensively in machine learning, e.g. [Alaydin, \"Introduction to Machine Learning\", 2014; Chapter 17], and self-training methods for learning from unlabeled data also have a long history, going back at least to [Fralick, \"Learning to Recognize Patterns without a Teacher\", 1965], and having emerged many times afters, e.g. in the context of semi-supervised learning (e.g. Chapelle \"Semi-Supervised Learning\", 2006) and also test-time adaptation, e.g. [Royer et al, \"Classifier adaptation at prediction time\", 2015] . Even for adapting ensemble methods self-training was proposed before, e.g. [Ghosh et al, \"Self Training with Ensemble of Teacher Models\", 2021]. In light of extensive prior work, the manuscript should make clear what the technical novelty of the proposed method is. \n\n* lack of baselines in experiments \n\nThe experiments evaluation presents results for the proposed method, but it does not contrast them against prior works. Consequently, the reader only learns that using proposed method often (but not always) works better than not doing test-time adaptation, but if the method is better than already existing methods for test-time-adaptation. It would also be important to see at least a comparison to obvious baselines, such simply self-training each network individually. For the specific choices made, e.g. using the majority vote prediction as target label but the softmax scores of the most confidence classifier, an ablation study would be needed if this is actually useful, or if maybe using hard labels or the ensemble average confidence would work equally well (or better).\n\n* unsubstantiated claims\n\nThe manuscript contains factual claims, e.g. about design choices that are not self-evident but also not substantiated with evidence. Some of these appear misleading and/or are not credible, see \"Questions\". The counterexample on page 4 seems to be a misunderstanding of prior work. Claims in the literature are not that diversity is *necessary* for a good ensemble. Indeed, as the submission states, an ensemble consisting of many copies of a perfect classifier are still perfect. But rather, diversity of prediction *mistakes* between models in an ensemble can provably beneficial accuracy. Without any errors, that notion is not defined. But if mistakes do occur, the variance of predictions is decreased if errors are negative correlated (e.g. (17.5) in [Alpaydin, 2014]). \n\n* shortcomings in the experimental protocol \n\nSeveral aspects about the experimental evaluation are unclear or misleading (see questions below). \n- The reported accuracy value in Tables 1-3 are \"highest accuracy improvements\" over all epochs.\nThat means they are not unbiased estimates, but potentially overconfident. \n- The description of ensemble elements is not clear from the text, it currently seems only provided in the Table headers. \n- The specified regularization constant $\\alpha=10e^{-11}$ (which should probably be $\\alpha=10^{-11}$) is so small that no regularizing effect is mathematically possible even after hundreds of thousands of steps. I would expect $\\alpha=0$ to work equally well. \n- The exact variant of \"ImageNet\" dataset used is not clear. Best provide a source how the data was obtained and prepared. \n- The results table lists differences in accuracy, but not absolute accuracy numbers, which would be useful to judge e.g. if the base classifiers are suitable for the task. \n- It is not specified how the method's hyper-parameters were selected. \n- Given the mixed positive and negative results, a test of significance should be performed if the reported results are not equally well explained by random chance (e.g. Wilcoxon signed rank). \n\nFurther comments:\n- I found the analogies with human learning or research (for example the top of page 3) rather superficial, and I would recommend to remove those. \n- The reference section should be corrected. Many papers are listed as arXiv or without publication venue, which actually have been published, e.g. [Arazo et al, IJCNN 2020], [Bucila etal, KDD 2006], ..." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Is there any theoretical or empirical evidence that a minimizing the diversity of the ensemble is beneficial for TTA? \n\n2. In Figure 2, it shows that predictions with lower IEL loss have lower entropy. However, previous study [3], also cited in the paper (Lines 101–105, 173–176), claimed that reducing entropy of output during training can be problematic for TTA. This raises doubts about whether lower IEL loss actually leads to higher TTA performance, and I am curious whether there is any evidence to verify the relationship between IEL and TTA performance. \n\n3. Are there any experimental results that address the weaknesses mentioned (e.g., comparisons with previous studies, ablation study on the number of models, computational cost comparisons, and performance comparisons across various TTA scenarios)? \n\n4. If there are multiple source-trained models, what advantages does IEL offer over an approach that applies conventional TTA methods to each model individually and then performs an ensemble?\n\n[3] Entropy is not enough for test-time adaptation: From the perspective of disentangled factors" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. This paper is well written and easy to follow along with their reasoning.\n2. The idea of utilizing ensemble methods for TTA is simple, intuitive and easy to adapt.\n3. Experiments were conducted on a single-sample setting, which is one of the challenging tasks in TTA." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces a new method for Test-Time Adaptation (TTA), Intra-model Ensemble Learning (IEL), that optimizes multiple models for ensembled inference. In the IEL framework, the output of the model with the highest probability for the class that received the most votes is set as the pseudo-label, and all other models are optimized to output this pseudo-label via minimizing the cross-entropy. This process minimizes the diversity within the ensemble and aligns the outputs of the models to ensure mutual agreement. Experimental results show the effectiveness of the proposed framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is insufficient justification for minimizing the ensemble diversity. Personally, I believe that as diversity increases, performance can be improved in general including TTA since the information is also increases. For example, in [1] which is also referenced in this manuscript, it was shown that increasing diversity can lead to a lower loss. Additionally, the counterexample of ensemble of models with 100% performance (Lines 161–166) is unrealistic and therefore inappropriate for supporting the claim. If there is a large distribution shift and the source-trained models perform poorly on target dataset, reducing diversity may actually have an adverse effect. In conclusion, it remains unclear how reducing ensemble diversity benefits TTA.\n\n2. Due to multiple assumptions, the scope of application of this research is limited. The authors assume there are multiple source-trained models (Lines 230–232), but it is questionable whether this assumption is easily met in practice. Furthermore, the assumption of stationary distribution shifts (Line 265-266) raises concerns about whether IEL would be effective in other TTA scenarios such as online imbalanced label distribution shifts or mixed distribution shifts [2].\n\n3. The experiments conducted do not sufficiently demonstrate the effectiveness of IEL. For example, the authors should include 1) performance comparisons with various previous works on TTA, 2) ablation study on the number of ensemble models, and 3) comparisons of computational costs associated with using multiple models. As mentioned earlier, including experiments across diverse TTA scenarios would also provide a more comprehensive understanding of IEL’s effectiveness.\n\n[1] A unified theory of diversity in ensemble learning\n[2] Towards stable test-time adaptation in dynamic wild world" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Ensemble and Entropy-Optimization based algorithm for adapting sets of pre-trained models to distribution shifted data." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024iel,\ntitle={{IEL}: Intra-Model Ensemble Learning For Single Sample Test-Time Adaptation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4LiegvCeQD},\nnote={under review}\n}" }, "abstract": { "value": "Test-Time Adaptation (TTA) problems involve adapting pre-trained models to new data distributions in testing time, with access to only model weights and a stream of unlabeled data. In this work, we present IEL, a method for adapting sets of independently pre-trained classifiers to distribution shifted data one sample at a time without labels. We minimize the cross-entropy between the classifier output that has the highest predicted probability for the majority voted class (a high confidence softmax) and all other models in a set of classifiers. The majority voted model that all others learn from may change from sample to sample, allowing the group to collectively learn from each other. Our method uniquely optimizes all trainable parameters in each model and needs only a single sample for adaptation. Using sets of independently pre-trained base classifiers with distinct architectures, we show that our approach can reduce generalization error for image classification tasks on corrupted CIFAR-10, CIFAR-100, and ImageNet while also minimizing the entropy of model outputs." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Test-Time Adaptation", "Ensemble Learning", "Entropy-Regularization", "Knowledge Distillation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/fa64481045e339fea8fd39930e5a287ab15dc720.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "IEL: Intra-Model Ensemble Learning For Single Sample Test-Time Adaptation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper is well-written and easy-to-follow.\n2. The paper makes a commendable observation concerning the entropy spectrum in deep neural network layers, which is a significant contribution to the field and forms the basis for the proposed defense mechanism.\n3. The efficiency of the proposed democratic training method is noteworthy. It circumvents the need to generate UAPs during training, instead utilizing a limited number of epochs to identify low-entropy examples, which is a resourceful approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents an investigation into the defense against universal adversarial perturbations (UAPs), with a particular focus on targeted UAPs. The authors have made a notable observation regarding the entropy spectrum in hidden layers when UAPs are introduced, which serves as the cornerstone for their proposed 'democratic training' approach. This method aims to enhance the adversarial robustness of neural network models. The empirical results provided in the paper demonstrate the efficacy of the approach, as well as its ability to maintain the clean accuracy of the model.\n\nIn general, this paper is well-structured and presents a novel approach, which meets the basic criteria of ICLR conference. However, there are some aspects that could be improved or expanded upon to enhance the overall quality and impact of the paper." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The threat model employed in the experiments primarily utilizes gradient-based attack methods. These methods presuppose access to the model's parameters, aligning with white-box attack scenarios. This appears to be at odds with the assertion in Section 2.3 that adversarial knowledge does not extend to the internal parameters of the model. Clarification on this point would be beneficial.\n\n2. The comparison with adversarial training methods may require further refinement. Adversarial training aims to bolster adversarial accuracy by integrating adversarial examples with clean examples during training. Constraining the number of training epochs could result in an underfit model, which may not provide a fair benchmark. Additionally, it would be advantageous to include a comparison with the widely recognized TRADES method[1], which is absent from the current manuscript.\n\n3. The potential for adaptive attacks warrants consideration. If adversaries are aware of the defense strategy, they could tailor adversarial examples to bypass the defense. I know that in the threat model, no adaptive attacks are considered since the attackers do not know the internal details of the models. However, the chosen attack methods in the experiments inherently rely on gradient information. So I would suggest that the authors should consider the potential for adaptive attacks.\n\n4. The scope of the experiments is largely limited to datasets comprising natural images. It would be beneficial to extend the evaluation to smaller-scale datasets, such as CIFAR-10, to complement the findings and potentially leverage open-source robust models for further exploration of the neuron entropy spectrum concept.\n\n5. While the paper discusses various existing defensive methods against UAPs and includes experimental comparisons, a direct comparison with state-of-the-art methods is missing. It is recommended to condense the background section and incorporate a more thorough comparison with leading-edge techniques.\n\n6. Minor Issues\n (1) Please consider to reduce the margin between Figure 1 and the text.\n (2) Suggesting to add necessary notations (SR) from the main test to the Table 2 for better understanding.\n\n\n[1] Zhang, Hongyang, et al. \"Theoretically principled trade-off between robustness and accuracy.\" International conference on machine learning. PMLR, 2019." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "Yes, Privacy, security and safety" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How about performances of the method on ViT?\n2. What's the time cost of the method?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The use of entropy to reveal the dominance of UAPs and the concept of Democratic Training as a defense mechanism is innovative.\n2. The method was evaluated across various neural network architectures and benchmark datasets, which strengthens the claim of its general applicability.\n3. Unlike other defense methods, Democratic Training does not require architectural modifications, which makes it easy to integrate into existing systems" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel defense method called Democratic Training to mitigate the impact of Universal Adversarial Perturbations (UAPs) on deep neural networks. The authors observed that UAPs lead to an abnormal entropy spectrum in hidden layers, which shows the model's prediction is dominated by a small subset of features. Democratic Training mitigates this issue by increasing entropy to ensure model predictions rely on a wider range of features. The approach was evaluated on multiple models and datasets and was found to be effective in reducing attack success rates while preserving accuracy on clean data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The evaluation focused primarily on benchmark datasets and common UAP generation methods. It would be beneficial to see how this approach performs on more sophisticated and adaptive attacks, such as adversarial examples generated in dynamic environments.\n2. The proposed method mainly works well on CNN. Authors should validate it in more types of networks, such as transformers.\n3. The method requires access to a small set of clean data for entropy measurement and training, which might not always be practical" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. For RQ3: did you adversarially train a model from scratch or just fine-tune a pretrained model with an adversarial training objective?\n2. I'm not very convinced by the claim in Lines 239-243 that says middle-layer feature entropy suggests the model's classification confidence." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The experiments are comprehensive.\n2. The proposed defense is attack-agnostic which is more practical and efficient.\n3. The proposed defense largely reduced the targeted attack success rate.\n\nI tend to accept this pape. However, since I'm not familiar with UAP attack and defense baseline methods, I will listen to other reviewers and public comments and then decide." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To improve neural network robustness to targeted UAPs, this paper proposed an adversarial training-like method that fine-tunes a pretrained model to reduce middle-layer entropy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. UAP attacks evaluated in the paper were published in 2018,2019,2020 and seem out-of-date.\n2. After democracy training, there is still a gap between ``AAcc.'' and clean accuracy. I wonder about the effectiveness of democracy training against non-targeted UAPs.\n3. Average results in Table 4\\&5 are ambiguous since there can be a large bias among different networks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "no" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1: The author obtained the idea of Democratic Training by studying the performance of UAP between different layers in network. I want to know why author said 'Democratic Training does not rely on generating UAPs and are thus not limited to specific UAP attacks.', As mentioned in the article, the UAP analyzed by the author is produced based on FA-UAP, so the properties analyzed should mainly for such kind of UAP, and the algorithm guided by this should also mainly target on FA-UAP. If it cannot be argued that all UAPs have such properties, then this statement seems unreasonable?\n\n2: The author did not provide a detailed explanation for $H(i)$ in algorithm 2 line 2, according to the equation (3), I think that the author is trying to say $H(i)$ is an entropy loss. Less strictly speaking, maximizing H(x) means that the various components of x are directly averaged as much as possible. So, is the goal of SG(SampleGenerator, Algorithm 2) equivalent to finding an adversarial sample with each component average? Is this a form of weakened PGD? If so, why is finding weaker adversarial samples beneficial for improving robustness? Why not directly target finding UAP for SG?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1, The setting is reasonable, author want to resist universal adversarial samples through small cost, this may be useful in some situation.\n2, Good experiment results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The author aims to improve the robustness of universarial adversarial by fine-tuning with a small amount of data, mainly by performing entropy based data augmentation to suppress the influence of general adversarial perturbations in a given model. Then the experiment was presented." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1, The method provided in the article is not novel, overall, it is still based on the 'min-max' methods. And in my opinion, it seems to be a weakened version of adversarial training.\n\n2, The symbols are somewhat confused. For example, in euqation (4), author use $L_{cce}$, I think this means cross entropy loss, but author did not introduce what is $L_{cce}$ before the euqation (4); in algorithm line 3, author define $I_b^{en}$ which is got by SampleGenerator, but do not use it in the following of algorithm, or maybe it have been wirtten as $i_{en}$ in line 4?\n\n3, I try to test the author's algorithm on CIFAR10(VGG16, budget 8/255), but I didn't get such a good result shown in table2, the SR has only decreased by about 10%. Author did not submit the code, so I hope the author can provide a detailed introduction to the settings for Algorithms 1 and 2 (For example, how to select the epoch, learning rate, hyperparameters), and it's best to provide a detailed explanation of each symbol and steps in the algorithm." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A novel method to mitigate the effect of UAPs via democratic training." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024democratic,\ntitle={Democratic Training Against Universal Adversarial Perturbations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4M0BRyGMnJ},\nnote={under review}\n}" }, "abstract": { "value": "Despite their advances and success, real-world deep neural networks are known to be vulnerable to adversarial attacks. Universal adversarial perturbation, an input-agnostic attack, poses a serious threat for them to be deployed in security-sensitive systems. In this case, a single universal adversarial perturbation deceives the model on a range of clean inputs without requiring input-specific optimization, which makes it particularly threatening. In this work, we observe that universal adversarial perturbations usually lead to abnormal entropy spectrum in hidden layers, which suggests that the prediction is dominated by a small number of ``feature'' in such cases (rather than democratically by many features). Inspired by this, we propose an efficient yet effective defense method for mitigating UAPs called Democratic Training by performing entropy-based model enhancement to suppress the effect of the universal adversarial perturbations in a given model. \\emph{Democratic Training} is evaluated with 6 neural networks trained on 4 benchmark datasets and 4 types of state-of-the-art universal adversarial attack methods. The results show that it effectively reduces the attack success rate, improves model robustness and preserves the model accuracy on clean samples." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Neural network adversarial attack; Universal adversarial perturbation; Adversarial attack defense" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/e095a7d7577f3fbd41df74b2a759cc1c02fe072d.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Democratic Training Against Universal Adversarial Perturbations" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Q1: what are the backgrounds of the developers participating in the evaluation and data annotation, and how their potential biases may have affected the soundness of the approach and the evaluation? \n\nQ2: the larger LLMs have clear edges over the CodFavor improved small models with much lower costs than human evaluators. Why would they not be a better option than using CodeFavor to improve the smaller models?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper presents a significant advancement in the field of code preference learning by introducing the CODEFAVOR framework, which innovatively utilizes synthetic evolution data to train models that predict meaningful code preferences. The novelty lies in its dual focus on aligning human and model preferences with verifiable code properties, addressing a critical gap in existing research. Key contributions include the development of CODEPREFBENCH, a comprehensive benchmark with 1364 curated tasks that evaluate code based on correctness, efficiency, and security, thus providing a robust evaluation framework for future studies. The results demonstrate that CODEFAVOR can enhance model accuracy by up to 28.8% while being more cost-effective than larger models, highlighting its practical significance in improving code generation assessments. Additionally, the paper sheds light on the limitations of human-based assessments, emphasizing the need for model-based approaches in evaluating non-functional code properties, which further underscores the importance of the research findings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the challenge of assessing code generation based on well-formed properties and aligning it with developer preferences, which has proven difficult in the context of Large Language Models (LLMs). To tackle this issue, the authors propose CODEFAVOR, a framework designed to train pairwise code preference models using synthetic evolution data, including code commits and critiques. Additionally, they introduce CODEPREFBENCH, a benchmark consisting of 1364 curated code preference tasks that evaluate three key properties: correctness, efficiency, and security, alongside human preferences. The main results indicate that CODEFAVOR significantly enhances the accuracy of model-based code preferences by up to 28.8%, while also demonstrating that these models can perform comparably to those with 6 to 9 times more parameters, all while being 34 times more cost-effective. Furthermore, the study highlights the limitations of human-based code preference assessments, revealing that a substantial percentage of tasks remain unsolved despite considerable time investment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The problem formulation/setting can be improved in terms of clarity, motivation, and realism. The framework is proposed to serve code assessment purposes, i.e., judging automatically which version of code generated by a model from a prompt is preferred (i.e. more correct/secure/efficient) between a pair of two versions. The questions are (1) in what scenarios would these two versions be available, and (2) how realistic it is that there are such strong and discriminative contrasts between the two versions (i.e., correct versus wrong, fast versus slow, secure versus vulnerable). In a typical use scenario of LLMs for code generation, developers may feed the LLM with a prompt and get a response. Should they always ask the model for two responses? If so, the cost would double. More importantly, it is probably unlikely that there is a such contrast between the two versions of code generated from the same prompt---e.g., the two versions could be similarly good or bad. Learning preferences in a strongly differentiable pair of two responses does not seem to be realistic. If so, the paper may want to provide supporting evidence that this is the case. Or the problem itself is not motivated convincingly. \n\nAnother primary weakness is the heavy reliance on synthetic evolution data for training the CODEFAVOR framework. While synthetic data can be useful, it may not fully capture the complexities and nuances of real-world coding scenarios. This limitation raises concerns about the generalizability of the model's performance in practical applications, as the evaluation may not reflect actual developer preferences or code behavior in diverse environments.\n\nThe paper acknowledges the prohibitive costs and limitations of human-based code preference assessments, noting that despite significant time investment, a substantial percentage of tasks remain unsolved (15.1% to 40.3%) . This suggests that human evaluators may struggle with certain tasks, which could undermine the reliability of the human preference data used for comparison. The paper could benefit from a more in-depth exploration of these limitations and their implications for the overall findings.\n\nThe paper mentions that the reliability of using large language models (LLMs) as evaluators often hinges on their reasoning capabilities, which can be subject to inherent biases . This raises questions about the objectivity of the model-based preferences derived from LLMs, as biases could skew the results and affect the alignment with human preferences. A more thorough examination of potential biases and their impact on the findings would strengthen the paper's arguments.\n\nThe evaluation framework, CODEPREFBENCH, focuses on three specific properties: correctness, efficiency, and security. While these are important aspects, the paper does not justify the choices, among various other quality aspects of code, such as maintainability or readability. Also, it seems that each of these chosen properties is separately considered, yet in real-world scenarios developers need to balance multiple factors at the same time when choosing which code to adopt (e.g., code that is both secure and correct). The interplay among these potentially competing factors is not considered in the approach nor in the evaluation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "(1) In Table 1, to validate the effectiveness of the developed training framework, might it be helpful to add some baseline training approaches which also train the LLMs using the same training data used by CODEFAVOR?\n\n(2) In Table 1, might it be helpful to understand whether CODEFAVOR can further improve these Open-Weight Models with larger sizes?\n\n(3) Would the developed approach also be effective, if the developed approach is applied to some other LLMs for code related tasks (e.g., Code Llama, Lemur)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(1) The paper is well written and easy to follow.\n\n(2) The paper introduces a benchmark which can potentially be used by future papers.\n\n(3) The developed approach is evaluated using multiple LLMs, showing that the developed approach is generally effective.\n\n(4) The developed approach has good intuitions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes CODEFAVOR, a framework for training pairwise code preference models from synthetic evolution data, including code commits and code critiques. To evaluate code preferences, the paper introduces CODEPREFBENCH, a benchmark comprising 1364 rigorously curated code preference tasks to cover three verifiable properties - correctness, efficiency, and security - along with human preference. The evaluation shows that CODEFAVOR holistically improves the accuracy of model-based code preferences." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) In Table 1, it seems that the approaches in the rows are either LLMs, or LLMs with the training framework developed in this paper. To validate the effectiveness of the developed training framework, might it be helpful to add some baseline training approaches which also train the LLMs using the same training data used by CODEFAVOR?\n\n(2) In Table 1, considering that there is still a gap between the Open-Weight Models and Our Models and Baselines (i.e., LLMs used with CODEFAVOR), might it be helpful to understand whether CODEFAVOR can further improve these Open-Weight Models with larger sizes?\n\n(3) It might be helpful if the paper can show the effectiveness of the developed approach when the approach is applied to some other LLMs for code related tasks (e.g., Code Llama, Lemur)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Do you have concerns that the synthetic data generation methods, Commit-Instruct and Critic-Evol, may not fully ensure code correctness? If not, this raises another question: the quality of synthetic data is highly dependent on the LLM used to generate it. How do the experiments demonstrate that CODEFAVOR’s performance gains are due to the framework itself rather than simply distilling knowledge from a stronger LLM (Llama-3-70B-Instruct)?\n- Could you provide more details on the inference process for the evaluation results in Table 3? Specifically, how many samples were created for each problem, what temperature was used, and are the results statistically significant?\n- Could you elaborate on any aspects that emphasize the novelty of your work compared to previous studies?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* This paper contributes two code preference synthetic dataset and a CODEPREFBENCH, a collection of 1,364 carefully curated preference tasks, To evaluate code preferences labeled by various approaches.\n* This paper comprehensively quantify and conduct case studies on code preferences derived from human developers and LLMs.\n* CODEFAVOR models can match the preference accuracy of models that are larger by 6∼9×, while being cheaper by 34×" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes CODEFAVOR, a framework for training pairwise code preference models using synthetic evolution data generated from code commits and LLM critiques. This approach addresses the challenge of aligning code generation with developer preferences, focusing on correctness, efficiency, and security through a benchmark called CODEPREFBENCH, which includes 1,364 preference tasks. CODEFAVOR models achieve comparable performance to much larger models while being more cost-effective, and experiments reveal that human preferences often fall short in non-functional objectives like efficiency and security. The study provides insights into balancing model and human preferences, highlighting the potential limitations and strengths of each approach​." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The approach to synthetic data generation lacks originality, as creating datasets from git commits [1,6] and evolving from sampled code[2,3] are common practices in the field.\n- The pairwise modeling approach is also not particularly novel; using pairwise prompts, criterion-based prompting, and classification or generation labels [4,5,7] have been previously explored in other studies.\n- Additionally, there is concern that synthetic data generation may not fully ensure code correctness, as it heavily depends on the LLM used for critique and generation. The chosen model, Llama3-70B-Instruct, is relatively weak compared to state-of-the-art models and limited to only this single model.\n- Finally, it is challenging to determine whether the performance gains following CODEFAVOR training are due to the distillation of knowledge from stronger LLMs used in data generation or from the CODEFAVOR training itself.\n\n\n1. Jimenez, Carlos E., et al. \"Swe-bench: Can language models resolve real-world github issues?.\" arXiv preprint arXiv:2310.06770 (2023).\n2. Luo, Ziyang, et al. \"Wizardcoder: Empowering code large language models with evol-instruct.\" arXiv preprint arXiv:2306.08568 (2023).\n3. Wei, Yuxiang, et al. \"Magicoder: Empowering code generation with oss-instruct.\" Forty-first International Conference on Machine Learning. 2024.\n4. Dong, Yi, et al. \"Steerlm: Attribute conditioned sft as an (user-steerable) alternative to rlhf.\" arXiv preprint arXiv:2310.05344 (2023).\n5. Wang, Zhilin, et al. \"Helpsteer: Multi-attribute helpfulness dataset for steerlm.\" arXiv preprint arXiv:2311.09528 (2023).\n6. Ding, Yangruibo, et al. \"Crosscodeeval: A diverse and multilingual benchmark for cross-file code completion.\" Advances in Neural Information Processing Systems 36 (2024).\n7. Qin, Zhen, et al. \"Large Language Models are Effective Text Rankers with Pairwise Ranking Prompting.\" Findings of the Association for Computational Linguistics: NAACL 2024. 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. The Security score of human developers in Table 3 is only 59.7. Does this indicate that humans are not proficient at judging code security, even similar to random selection? \n\n2. Could you further explain “Scores within 1 percentage point of the highest” in Table 3, as well as the detailed measurement method for \"uncertain responses\"?\n\n3. The authors discovered that code comments may negatively affect model preferences, which is a bit strange and may be harmful to real-world applications. Is it possible to result from the class imbalance in comments (e.g., a higher proportion of comments in positive examples)? Could you provide the number of comments in positive and negative examples in the training and testing sets?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The synthetic dataset construction method (commits and critiques) is technically sound and novel to me.\n\n- The authors conducted a comprehensive evaluation of the method. In addition to correctness, which is the focus of many traditional code generation studies, the authors also assess efficiency, security, and human developer preference.\n\n- The authors put significant effort into the formatting of images and tables, which enhances the readability of the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to enable LLMs to better assess the quality of two code snippets by constructing a synthetic pairwise code preference dataset. The dataset is built using code commits (with pre- and post-commit versions as contrastive samples) and code critiques (the code snippet improved by a superior LLM as a contrastive sample). The authors have built benchmarks in Correctness, Efficiency, Security, and Human Preference to test the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My main concern is that the authors overlook many works on code ranking and do not provide any experimental comparison. Many statements, such as \"learning code preferences has been *largely under-explored*\", \"the *first* open recipe to train pairwise code preference models\", and \"understudied code generation domain\", appear to overclaim. To name a few: \n\n- A basic, training-free baseline is to compare the mean log probability of two code snippets and select the one with the highest probability, as in [1]. Furthermore, [2] also uses model likelihood for code selection. \n- Some research also explores training classifiers to choose the better code, as in [3].\n- The authors did not compare their work with the dataset in [4] mentioned in the Related Work section.\n- In addition, but less importantly, to better contextualize the paper, some words about recent advances in execution-based code selection [5,6,7,8] would be appreciated. Particularly, [8] also employs a generation task similar to this paper. Considering that the work is recent, this comparison is not necessarily required.\n\nSince the authors only reported the performance of the backbone LLMs and lacked empirical comparisons with advanced code selection methods, it is difficult to determine the relative improvement level of this work within related studies.\n\n2. Some training details in the paper require further clarification. For instance, does the classifier task operate on the next token in Equation (1)? If so, considering that the label tokens (\"A\" or \"B\") and the criterion $c$ are tightly connected without a delimiter or explicit prompt, how does the LLM recognize where the criterion ends to output label tokens?\n\n3. Since the authors collected both the training set and testing benchmarks, it's unclear whether they took decontamination steps to prevent test set leakage. If no decontamination was performed, analyzing the potential overlap between the training and test sets would be beneficial.\n\n**Minor comments**\n\n- The caption for Listing 1 is missing a period at the end.\n- It would be better to place Equation (2) right after line 141.\n\n**References**\n\n[1] Evaluating Large Language Models Trained on Code, https://arxiv.org/abs/2107.03374\n\n[2] Coder Reviewer Reranking for Code Generation, ICML 2023.\n\n[3] Fault-Aware Neural Code Rankers, NeurIPS 2022.\n\n[4] CodeUltraFeedback: An LLM-as-a-Judge Dataset for Aligning Large Language Models to Coding Preferences.\n\n[5] CodeT: Code Generation with Generated Tests, ICLR 2023.\n\n[6] Natural Language to Code Translation with Execution, EMNLP 2022.\n\n[7] B4: Towards Optimal Assessment of Plausible Code Solutions with Plausible Tests, ASE 2024.\n\n[8] Sifting through the Chaff: On Utilizing Execution Feedback for Ranking the Generated Code Candidates, ASE 2024." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We train models and build benchmarks to predict code preference towards verifiable properties and human preference." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning Code Preference via Synthetic Evolution},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4MWUdp6deL},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) have recently demonstrated remarkable coding capabilities.\nHowever, assessing code generation based on well-formed properties and aligning it with developer preferences remains challenging.\nIn this paper, we explore two key questions under the new challenge of code preference learning:\n(i) How do we train models to predict meaningful preferences for code? and\n(ii) How do human and LLM preferences align with verifiable code properties and developer code tastes?\nTo this end, we propose CodeFavor,\na framework for training pairwise code preference models from synthetic evolution data,\nincluding code commits and code critiques.\nTo evaluate code preferences,\nwe introduce CodePrefBench, a benchmark comprising 1364 rigorously curated code preference tasks to cover three verifiable properties—correctness, efficiency, and security—along with human preference.\nOur evaluation shows that CodeFavor holistically improves the accuracy of model-based code preferences by up to $28.8$%.\nMeanwhile, CodeFavor models can match the performance of models with $6\\sim 9\\times$ more parameters\nwhile being $34\\times$ more cost-effective.\nWe also rigorously validate the design choices in CodeFavor via a comprehensive set of controlled experiments.\nFurthermore, we discover the prohibitive costs and limitations of human-based code preference:\ndespite spending 23.4 person-minutes on each task, $15.1\\sim 40.3$% of tasks remain unsolved.\nCompared to model-based preference,\nhuman preference tends to be more accurate under the objective of code correctness,\nwhile being sub-optimal for non-functional objectives." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Code Generation", "Large Language Model", "Preference Learning", "Evaluation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/565586e8dffe117bed8212374deddd4773516d99.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Learning Code Preference via Synthetic Evolution" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weaknesses above. A few additional questions are below:\n\n* Could a discussion on certain edge cases where the theory does not hold be provided? More precisely, would it be possible to find situations where the assumptions made in theory do not hold well, resulting in a breakdown of expected results?\n\n* Additionally, could results for ViT be provided in situations where the number of parameters are the same as parameters in FourierFT, for a more intuitive comparison?\n\n* I would also like to see results in the Natural Language Generation task - particularly for the GPT-2 training performed by FourierFT, as it would indicate the effectiveness of the method when ported to a Conv1D based implementation of the MLP." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ The paper takes an existing idea - learning PEFT parameters in the frequency domain and reconstructing the weight matrix from those learned parameters - and performs an in-depth analysis of the approach. Based on the obtained insights, it presents a new method that is theoretically better than existing approaches in approximating the ideal fine-tuning matrix and shows quantitative improvements on the base method, FourierFT, in most cases. \n+ The paper includes intuitive theoretical statements, backed by mathematical proofs, which is good to see in a PEFT paper since existing methods often lack theoretical insights and are heuristic-based. \n+ The paper is also well written and intuitive to follow, with rigorous experiments. \n+ I especially liked the approach presented for learning discrete locations to be optimized, which I believe is a novel contribution." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel low-rank adaptation approach for fine-tuning transformer-based pretrained models by deriving weights from parameters learned in the frequency domain using the iDCT operation. Compared to the similar existing method FourierFT, the approach, in theory, promises better reconstruction of the oracle update matrix. Empirically, the results do improve over the baseline FourierFT approach in most cases, indicating effectiveness of the approach. The paper also provides a method to learn locations in the frequency domain where coefficients are required, which is novel and interesting in the context of PEFT." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper starts with some terminology that are not elaborated: “optimization space” and “flexible optimization”. These terms are not defined precisely anywhere, nor is their link to the theory or empirical results clear. It would be better to ground the explanation to well-defined terms that are used in the analysis.\n\n- According to the reference (Yohai & Maronna, 1979), the initial assumption, the equation in Sec 2, L99, is true only under certain conditions. \nThis variant of the equation holds true only when $\\psi$ is monotone and X to have full rank. However, this issue is not addressed anywhere in the paper. For eg., the matrix X according to the paper is the input matrix. In practical implementation, the dimension X is m*n where (m<n), i.e. for X to be full rank, we need rank(X) = m. This is not stated or supported in the paper.\n\n- The theory presented is highly domain-specific: it does not translate to more general PEFT methods such as VeRA or DoRA, and requires significant theoretical adaptations to allow for comparisons with arbitrary low-rank methods. \n\n- The theory also does not *always* agree with practice - there are certain cases where LoRA and FourierFT perform better. This indicates come confounding factors, yet no discussion on this has been included. I do not see this as a reason to reject however, as this is commonly seen in this area, but would appreciate a discussion on the same.\n\n- In the case of ViT, I would have liked to see comparisons of the proposed approach with FourierFT having the same number of trainable parameters, as done for NLU and IFT tasks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1.Why are the accuracy rates of the baseline methods on the Stanford Cars and FGVC datasets more than 5% higher than those reported in related papers? I mainly compared the experimental results from the FourierFT paper(https://arxiv.org/pdf/2405.03003) and yours, and found that the differences are small on other datasets, but the results on the Stanford Cars and FGVC datasets are significantly beyond normal error margins. I am unsure whether this is due to errors caused by carelessness in the experimental process, or if you used different ViT models compared to theirs. Specifically, the experimental results on the Stanford Cars and FGVC datasets are emphasized in your work, and it is crucial to ensure the precision of these results.\n\n2.Why are there so few ablation experiments for FourierFT fine-tuning on ViT? As the most competitive counterpart, additional experimental results for FourierFT 239K and FourierFT 480K after fine-tuning on ViT could be included. After all, LoCA presents results for two different parameter budgets, while FourierFT only provides results for the smallest parameter budget for comparison, which does not meet the fairness requirements of an ablation study.\n\n3.What are the differences between LoCA and other methods in terms of Memory Cost and Training Time? You may use charts to illustrate these differences explicitly.\n\n4.Why does LoCA not show a significant advantage over FourierFT , on fine-tuning various foundation models such as RoBERTa, LLaMA, and ViT, in terms of reducing parameter budget and improving accuracy? Does this suggest that, while your work is strongly interpretable, it may have limited practical value?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.As emphasized by the authors, their iDFT-based variants has managed to outperform the expressivity of previous low-rank-based methods.\n\n2.Overall, the presentation is clear, supported by rigorous mathematical derivations and extensive experimental results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Location-aware Cosine Adaptation (LoCA), a novel frequency-domain parameter-efficient fine-tuning method for pre-trained LLMs. By leveraging the inverse Discrete Cosine Transform (iDCT) as well as selectively learning components in the frequency domain, LoCA addresses the constraints of the naive low-rank adaptation (LoRA) method. \nIn a word, LoCA enhances expressiveness while maintaining computational efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.Some baseline experimental results differ significantly from those in related papers, which may indicate carelessness in the experimental process. Also, more ablation experiments are needed to increase confidence.\n\n2.For most datasets, LoCA doesn't show a clear advantage over FourierFT in terms of reducing parameter budget and improving accuracy.\n\nPlease see the questions section for more details." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper provides a rigorous theoretical comparison between frequency-domain and low-rank adaptation methods, filling a gap in the literature on expressivity and optimization constraints.\n2. LoCA’s use of iDCT with dynamic selection of frequency components represents a creative improvement over conventional low-rank methods, particularly for parameter efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel parameter-efficient fine-tuning method, Location-Aware Cosine Adaptation (LoCA), that leverages inverse Discrete Cosine Transform (iDCT) for selectively optimizing frequency-domain components in pre-trained language and vision models. LoCA aims to surpass traditional low-rank adaptations by dynamically choosing informative frequency components, thus balancing parameter efficiency and performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall it's a good paper, and I will raise my score if the authors could address my concerns.\n1. LoCA introduces a computationally complex process with alternating optimization steps and central difference approximation, which could pose practical challenges.\n2. How does LoCA handle potential noise in frequency component selection, and are there measures in place to stabilize the optimization process?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1.The paper primarily compares LoCA with LoRA-related fine-tuning techniques. Has consideration been given to performance comparisons with other fine-tuning methods such as prompt learning and adapter tuning?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. LoCA introduces a novel approach for parameter-efficient fine-tuning in the frequency domain through the inverse Discrete Cosine Transform (iDCT) and selective learning of frequency components. This method demonstrates the potential to surpass traditional low-rank decomposition techniques both theoretically and empirically, which is of significant value for resource-constrained environments and the deployment of large-scale models. Furthermore, your work provides a comprehensive theoretical analysis comparing frequency domain methods with low-rank decomposition approaches, which is meaningful.\n2. The methodology section of the paper is rigorous, and the experiments cover multiple domains, including natural language processing and computer vision. The paper offers comparisons with existing techniques, such as LoRA and FourierFT, which help readers understand the performance and efficiency of LoCA. Additionally, the in-depth theoretical analysis provides a solid foundation for frequency domain parameter-efficient fine-tuning methods.\n3. The writing of the paper is clear and logically structured, with a coherent flow from the introduction to the methodology, experimental results, and conclusions. In particular, the detailed explanation of how LoCA operates, including the application of the inverse Discrete Cosine Transform and the alternating optimization strategy, enhances the reader's understanding of the relevant work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper titled \"LoCA: Location-Aware Cosine Adaptation for Parameter-Efficient Fine-Tuning\" introduces a novel parameter-efficient fine-tuning (PEFT) method called Location-Aware Cosine Adaptation (LoCA). This method is designed to adapt pre-trained large language models (LLMs) to downstream tasks with improved optimization flexibility and parameter efficiency. LoCA is based on the inverse Discrete Cosine Transform (iDCT) and selectively tunes learnable components at specific locations" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper contains a limited amount of content related to RELATED WORK, with insufficient coverage of the existing literature in the field.\n2. While the experimental results are convincing, the paper could further expand the experimental section to include the verification of LoCA's performance on more datasets. Additionally, a more in-depth analysis of LoCA's performance on different model scales and tasks of varying complexity would help to further demonstrate its applicability and robustness." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I will keep my positive score if the authors address Question 1. Other questions require much more experiment time and are quite minor to improve the paper.\n\n1. MT-bench is considered an unstable benchmark. It is strongly recommended that the authors utilize the MathInstruct Dataset instead, which is more stable and generally requires a higher level of expressive power.\n\n2. For fine-tuning Roberta, typical benchmarks include RTE, BoolQ, SST-2, WSC, WIC, MultiRC, SQuAD, CB, COPA, DROP, GSM8K, and ReCoRD. Could the authors consider adding any benchmarks that are currently missing?\n\n3. COLA, ReLoRA, and DoRA represent typical LoRA variants. It would be beneficial if the authors could include any of these variants that are not already covered.\n\n4. In Figure 3, it appears that the performance gain may continue to increase with a larger value of 'r.' Could the authors extend the range of 'r' to determine the optimal value that yields the best performance?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The concept of applying low-rank adaptation within the Fourier domain is intriguing, and it implicitly suggests a method of tuning that utilizes all available parameters.\n\n2. The theoretical results appear to be novel and have captured the interest of the reviewer.\n\n3. The proposed method delivers strong performance benefits while maintaining an exceptionally low parameter cost." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Location-aware Cosine Adaptation (LoCA), a novel method for fine-tuning large language models (LLMs) and vision models in a parameter-efficient manner. LoCA is based on the inverse Discrete Cosine Transform (iDCT) and optimizes the locations of learnable frequency components. It addresses limitations of previous low-rank adaptation methods by providing greater optimization flexibility and expressivity. Theoretical analysis and empirical observations support the superiority of LoCA over traditional low-rank methods and iDFT-based methods. LoCA dynamically selects the most informative frequency components during training, leading to enhanced parameter efficiency and computational feasibility. The method demonstrates state-of-the-art performance on diverse language and vision tasks with fewer parameters. The introduction of the paper contextualizes the need for parameter-efficient fine-tuning methods due to the prohibitive costs of fully fine-tuning increasingly large models, and LoCA is presented as a solution that maintains performance while reducing trainable parameters." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The reviewer, not being an expert in this area, has not identified any major weaknesses. However, with some background in empirically tuning LLMs and ViTs, the reviewer would like to inquire further about the experimental setup.\n\n1. There lack some benchmarks and baselines. \n\n2. Common advantages of the PEFT method include reduced computation and memory costs. The paper's contribution would be strengthened if the authors included these aspects in their analysis." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "● Q1: LoRA may not be the most parameter-efficient approach among spatial-domain PEFT methods. The work by Kopiczko et al. (2023) in \"VeRA: Vector-based Random Matrix Adaptation\" (arXiv:2310.11454) demonstrates a more parameter-efficient and lightweight alternative, focusing on diagonal matrix adaptations to achieve efficient adaptation without the need for frequency-based transformations. Could the authors clarify whether their proof and theoretical framework apply to VeRA? Additionally, this paper lacks a comparative analysis of VeRA, both theoretically and experimentally. Would the established proof also support an evaluation of DoRA’s expressivity?\n\n● Q2: For each layer in LoCA, both frequency component locations and coefficients are optimized individually. This approach appears to introduce a higher number of parameters compared to FourierFT, which selects $n$ locations randomly and shares these locations across all layers. Specifically, FourierFT’s parameter count is:\n\n$2n + n* L = n (L + 2)$\n\nwhere $L$ represents the number of layers in the pre-trained model.\n\nIn contrast, LoCA introduces $2n$ parameters for each layer’s and locations, and $n$ for each layer’s coefficients, resulting in a total parameter count of:\n\n$3n \\times L$\n\nThis yields a parameter ratio between LoCA and FourierFT of:\n\n$\\frac{3L}{L + 2}$\n\n\nFor example, with LLaMA-2 7B where $L = 32$, LoCA’s parameter count is approximately 2.82 times that of FourierFT. This raises concerns about parameter efficiency, especially in large models. To clarify whether the additional parameters in LoCA yield proportional benefits, could the authors provide empirical comparisons across various model sizes and tasks, measuring both fine-tuning performance and resource usage (e.g., memory and compute requirements)? Specific metrics, such as performance improvements relative to parameter increase and scaling efficiency on different benchmarks, would help assess whether gains in expressivity or accuracy justify the increased parameter cost.\n\n● Q3: In lines 191-199, the authors claim that randomly selecting frequencies for FourierFT yields the lowest expressivity, performing worse than LoRA; however, this claim lacks experimental support. For instance, Figure 3 in this paper shows mixed results for FourierFT on the FGVC task, whereas Figure 4 in Gao et al. (2024), *\"Parameter-Efficient Fine-Tuning with Discrete Fourier Transform\"* (arXiv:2405.03003), presents contrasting findings, particularly on the QQP task in GLUE. In Gao et al., FourierFT consistently outperforms LoRA across GLUE benchmarks, achieving higher accuracy with minimal random spectrum updates and fixed locations across layers. Furthermore, Section C.2 on Expressive Ability in the FourierFT paper’s supplementary material reinforces FourierFT’s superior expressivity over LoRA. Could the authors provide empirical comparisons to clarify these discrepancies, ideally across multiple model sizes and tasks, with metrics on fine-tuning performance and resource usage (e.g., memory and computational requirements)? Demonstrating whether the increased parameter count in LoCA yields proportional performance benefits would strengthen the case for its efficiency.\n\n● Q4: Additionally, the paper lacks empirical evaluations comparing selective FourierFT and LoCA, which would be valuable in validating the theoretical claims. For instance, in line 191, the statement that $W_F(3)$ can outperform \\$W_F(2)$ would benefit from empirical results to illustrate how these specific configurations impact performance. Further analyses using different selection strategies within FourierFT would also help substantiate the expressivity claims and clarify the mixed findings observed.\n\n● Q5: The proof assumes asymptotic normality of incremental weight updates, enabling statistical analysis of expressivity via the Central Limit Theorem and empirical spectral density. However, in LoRA, only a subset of weights is updated through low-rank reparameterization, while frequency-based methods like LoCA further restrict updates to high-amplitude frequency components. Given that these updates are gradient-driven and thus correlated, the i.i.d. assumption essential for CLT may not strictly hold. With limited, targeted updates, the cumulative effect lacks the \"sum of many independent adjustments\" necessary to ensure asymptotic normality. Could the authors provide further justification for assuming convergence to normality under selective updating, and clarify how potential deviations from i.i.d. behavior may impact expressivity comparisons? It would be helpful if the authors could conduct specific analyses or empirical tests, such as quantifying deviations from normality in the weight updates or performing sensitivity analyses to assess the impact of non-normality on expressivity. \n\n● Q6: In the alternating optimization strategy, the method first optimizes the coefficients of selected frequency components $\\alpha$, before refining their locations for $B_a$ steps. Then, with $\\alpha$ fixed, it optimizes the locations $l$ for $B_l$ steps, and finally, the procedure fixes the locations and optimizes only the coefficients $\\alpha$ until convergence.\n\nCould the authors clarify the rationale behind this specific order of coefficient-first optimization and its impact on stability and convergence? While this separate optimization approach might simplify the process, it may not fully capture the interactions between coefficients and locations, potentially limiting optimality. Have the authors explored an alternative order—optimizing locations first and then coefficients—and could they provide insights on how this might affect convergence and final performance?\n\nIn the ablation study (lines 480-485), the authors present several variant comparisons, yet they do not include an analysis of this alternative pipeline. Additionally, how are the parameters $B_l$ and $B_s$ selected—is their choice task-specific? From Table 4, it appears that the V5 variant achieves relatively better results, but this is not consistent with the description of the alternative policy in lines 284-292 and the algorithm in lines 945-963. Could the authors clarify these inconsistencies and provide further justification for the selected optimization order and parameter settings? \n\n\n● Q7: Could the authors clarify how LoCA ensures stable convergence given the dynamic selection of specific magnitude(e.g., high-magnitude) frequency components in $\\Delta W$ across epochs? Specifically, as top-ranked frequencies may shift due to gradient changes, how does LoCA maintain consistency in updates to avoid potential instability in the training process? Additionally, could the authors explain how the specific frequency components selected in LoCA—whether high or low frequencies—consistently contribute to model performance across tasks? Is there a risk that focusing solely on high-magnitude components could lead to loss of task-relevant information in lower-magnitude frequencies, which may carry finer-grained details?\n\n\n● Q8: Could the authors clarify the computational and memory overhead associated with estimating location gradients using finite-difference approximation? Specifically, does this approach increase CUDA memory requirements significantly, and if so, how does it impact the overall efficiency of LoCA? Additionally, an analysis of the trade-offs between accuracy and resource usage in this approximation method would be valuable to understand its practical feasibility." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper introduces LoCA, a frequency-based approach to parameter-efficient fine-tuning that selectively optimizes specific frequency components using the Discrete Cosine Transform. By focusing on significant frequencies within weight matrices, LoCA aims to reduce the parameter count needed for fine-tuning while maintaining model expressivity. This selective frequency adaptation presents a practical alternative to spatial-domain methods like LoRA, providing a new angle on efficient model tuning. The paper’s theoretical framework, including empirical spectral density and the Central Limit Theorem to analyze expressivity, helps ground LoCA's approach in established statistical methods, adding quality to the work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes LoCA, a method for fine-tuning pre-trained models using frequency-domain adaptation via the inverse Discrete Cosine Transform (iDCT). LoCA focuses on selecting key frequency components to improve the expressivity and efficiency of model adaptation. The theoretical analysis argues that iDCT-based adaptation can match or exceed the effectiveness of low-rank methods. However, the empirical gains over existing methods like LoRA are marginal, especially in vision tasks. LoCA’s added complexity, due to finite-difference approximations and alternating optimization, may not be fully justified by these modest improvements, potentially limiting its practical appeal." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "● This paper appears to provide inadequate empirical support for its theoretical claims. A central claim of the paper is that randomly selected frequencies for FourierFT yield lower expressivity than LoRA; however, this claim lacks direct experimental validation, which is critical to substantiate the theoretical conclusions. For instance, Figure 3 shows mixed results for FourierFT on the FGVC task, while Figure 4 in *Parameter-Efficient Fine-Tuning with Discrete Fourier Transform* by Gao et al. (2024) (arXiv:2405.03003) presents empirical evidence that contradicts this claim by showing that FourierFT achieves higher accuracy than LoRA across multiple GLUE benchmarks. Additionally, Section C.2 on Expressive Ability in the FourierFT paper’s supplementary material further supports FourierFT’s superior expressivity. The paper also lacks empirical evaluations for selective FourierFT and LoCA, which would further validate the claims made.\n\n● The paper omits a comparison with a highly representative spatial-domain PEFT method, VeRA, which focuses on lightweight adaptations in the spatial domain and would serve as a useful benchmark for LoCA's performance.\n\n● The design of LoCA introduces significant parameter overhead due to the individual optimization of frequency component locations and coefficients for each layer. For example, in a model with \\( L = 32 \\) layers (e.g., LLaMA-2 7B), LoCA’s parameter count is approximately 2.82 times that of FourierFT, raising concerns about the scalability and efficiency of LoCA for large-scale models.\n\n● The theoretical framework assumes asymptotic normality of weight updates, enabling the use of the Central Limit Theorem and empirical spectral density for analyzing expressivity. However, this assumption relies on i.i.d. updates, which may not hold in the context of gradient-driven, correlated weight adjustments inherent in LoRA and LoCA. Given the limited and targeted nature of LoCA’s updates, the cumulative adjustments may lack the “sum of many independent adjustments” necessary for CLT to apply reliably. This assumption weakens the robustness of the theoretical claims, as the actual distribution of weight updates is likely far from normal in practical implementations. \n\n● LoCA’s dynamic selection of high-magnitude frequency components across epochs may introduce instability during convergence, as the selection of significant frequencies may shift due to changing gradients. This could impact the model’s ability to achieve stable and consistent updates over time. Furthermore, by focusing solely on high-magnitude frequencies, LoCA risks omitting task-relevant information in lower-magnitude components, potentially limiting its adaptability in tasks requiring finer-grained details.\n\n● The method also relies on finite-difference approximation to estimate location gradients, which introduces additional computational and memory costs. This overhead may significantly increase CUDA memory requirements, particularly in high-dimensional models or when frequent updates are necessary." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024loca,\ntitle={Lo{CA}: Location-Aware Cosine Adaptation for Parameter-Efficient Fine-Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4NRjdISWby},\nnote={under review}\n}" }, "abstract": { "value": "Low-rank adaptation (LoRA) has become a prevalent method for adapting pre-trained large language models to downstream tasks. However, the simple low-rank decomposition form may constrain the optimization flexibility. To address this limitation, we introduce Location-aware Cosine Adaptation (LoCA), a novel frequency-domain parameter-efficient fine-tuning method based on inverse Discrete Cosine Transform (iDCT) with selective locations of learnable components. We begin with a comprehensive theoretical comparison between frequency-domain and low-rank decompositions for fine-tuning pre-trained large models. Our analysis reveals that frequency-domain approximation with carefully selected frequency components can surpass the expressivity of traditional low-rank-based methods. Furthermore, we demonstrate that iDCT offers a more efficient implementation compared to inverse Discrete Fourier Transform (iDFT), allowing for better selection and tuning of frequency components while maintaining equivalent expressivity to the optimal iDFT-based adaptation. By employing finite-difference approximation to estimate gradients for discrete locations of learnable coefficients on the DCT spectrum, LoCA dynamically selects the most informative frequency components during training. Experiments on diverse language and vision fine-tuning tasks demonstrate that LoCA offers enhanced parameter efficiency while maintains computational feasibility comparable to low-rank-based methods." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Parameter-efficient fine-tuning", "discrete cosine transform", "transfer learning", "adaptation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/4c737e648836f6f708508659aa0fae987f7d241a.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/4f5c88d595653e74bb8b81085101719dd5ade5aa.zip" }, "title": { "value": "LoCA: Location-Aware Cosine Adaptation for Parameter-Efficient Fine-Tuning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Could the authors clarify if they’ve noticed stability shifts in higher-dimensional or complex tasks and if adjustments might bolster robustness? Additionally, what drives the choice of a limited set of losses—are there theoretical or practical reasons for omitting other f-divergences, like Hellinger? Lastly, insights into each loss function’s hyperparameter sensitivity and effects on convergence guarantees would further clarify their resilience and adaptability." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper introduces a systematic framework for designing regression losses in GFlowNet training, linking each loss to specific divergence measures for targeted properties. Resulting in three new losses—Shifted-Cosh, Linex(1/2), and Linex(1)—that enhance exploration and exploitation balance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel framework for GFlowNet objective functions, unifying existing training algorithms and clarifying key components. By establishing a connection between objective functions and divergence measures, it offers valuable insights into designing effective training objectives. The authors investigate key regression properties—zero-forcing and zero-avoiding—and propose three new loss functions (Linex(1), Linex(1/2), and Shifted-Cosh) to balance exploration and exploitation. Extensive experiments on benchmarks, including hyper-grid, bit-sequence, and molecule generation, show that these losses outperform the common squared loss in convergence speed, diversity, quality, and robustness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Broader exploration of other potential divergence-based losses would offer a more comprehensive understanding of the effects of different divergence properties on GFlowNet training. Although the novelty lies in extending GFlowNet loss functions, there are similar attempts in reinforcement learning and generative models. Although the paper derives theoretical properties of zero-forcing and zero-avoiding, it lacks direct theoretical comparison with existing GFlowNet training algorithms." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Areas for improvement and suggestions:\n\n1. Making a literature connection with f-GAN, which also uses f-divergence in GANs, might be insightful to readers.\n\n\n2. Include a discussion connecting with off-policy exploration methods. Are your loss and off-policy search orthogonal? Which means, is your loss function combined with an off-policy method (e.g., local search) better than the TB loss combined with an off-policy method?\n\n\n3. It's good to see the categorization of prior GFlowNet works. Can you include this recent work [1] that uses genetic search as an off-policy method for training GFlowNets and provide some discussion?\n\n\n\n[1] Hyeonah Kim et al., \"Genetic-guided GFlowNets for Sample Efficient Molecular Optimization,\" NeurIPS 2024." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-written and easy to follow.\n\n\n2. The theories are insightful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel theoretical finding for GFlowNets regarding their objective function. Using f-divergence theories, they connect existing objectives of GFlowNets and show that they are special cases of the squared loss. They design a new loss structure that combines both properties together: (1) zero forcing (as considered in existing losses) and (2) zero avoiding, which compensates for exploration. Their new loss function seems to have empirical benefits." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The empirical results seem to be weak; they are only varied in synthetic tasks" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "See Weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "Contributions:\n- generalizing the objective function of gflownet\n- derive impact of loss function on the gradient\n- define zero-forcing (encourage epxloitation) and zero-avoiding (encourage exploration) as two key properties induced by certain loss functions\n- They create 3 new losses (alonside the existing quadratic loss) to tackle all 4 possible combinations (with/without zero-avoiding, with/without zero-forcing)\n- Linex(1) corresponds to the KL divergence\n- experiments on 3 datasets\n\t- Non-zero-forcing losses (Linex(1) and Linex(0.5)) converge faster on hyper-grid\n\t- Linex(1) obtains all the modes almost always the fastest, but spearman corr between train and test is highest for shifted-cos on bit-sequence\n\t- Linex(1) tends increase diversity while quadratic and shifted-cos give higher quality (high average rewards) on molecule generation\n\nPaper is well written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose to modify the loss function of Gflownet (which has been completely overlooked by prior work). They show that dinstinct losses lead to different divergences. They propose three new loss functions, evaluate them extensively on diverse benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- it would be nice to have a few more choices of losses, taking inspiration let say from f-divergences (Reverse-KL, JSD, etc.)\n- there may be more than just zero-forcing and zero-avoiding to the key properties of loss functions hence why studying more losses would be helpful\n- it would be nice to let say consider hybrid methods with some kind of annealing. For example, why not use Linex(1) for its fast convergence to a large number of nodes, before then transitioning to shifted-cos for higher rewards around those now-discovered modes.\n\nSo to me, the paper is great, but its kind of stopping too quickly, it feels like its only just tapping the surface. These kinds of ideas would be easy to test out and add to the papers. \n\nIf the authors add bit more meat to the paper, i.e. extra loss functions and hybrid-annealing (as discussed above), I would likely increase my score. Like I said, its just missing a little bit of filling to make it a great paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024beyond,\ntitle={Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4NTrco82W0},\nnote={under review}\n}" }, "abstract": { "value": "Generative Flow Networks (GFlowNets) are a novel class of generative models designed to sample from unnormalized distributions and have found applications in various important tasks, attracting great research interest in their training algorithms. In general, GFlowNets are trained by fitting the forward flow to the backward flow on sampled training objects. Prior work focused on the choice of training objects, parameterizations, sampling and resampling strategies, and backward policies, aiming to enhance credit assignment, exploration, or exploitation of the training process. However, the choice of regression loss, which can highly influence the exploration and exploitation behavior of the under-training policy, has been overlooked. Due to the lack of theoretical understanding for choosing an appropriate regression loss, most existing algorithms train the flow network by minimizing the squared error of the forward and backward flows in log-space, i.e., using the quadratic regression loss. In this work, we rigorously prove that distinct regression losses correspond to specific divergence measures, enabling us to design and analyze regression losses according to the desired properties of the corresponding divergence measures. Specifically, we examine two key properties: zero-forcing and zero-avoiding, where the former promotes exploitation and higher rewards, and the latter encourages exploration and enhances diversity. Based on our theoretical framework, we propose three novel regression losses, namely, Shifted-Cosh, Linex(1/2), and Linex(1). We evaluate them across three benchmarks: hyper-grid, bit-sequence generation, and molecule generation. Our proposed losses are compatible with most existing training algorithms, and significantly improve the performances of the algorithms concerning convergence speed, sample diversity, and robustness." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "GFlowNet", "Generative Models", "f-Divergence", "Loss Function" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/19af6e2a44f9fa1923f4536bbf114c1797ae8ae0.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Beyond Squared Error: Exploring Loss Design for Enhanced Training of Generative Flow Networks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Following are the questions for which I would highly appreciate an answer, these questions have not impacted my current recommendation for this paper, however, the response might have a significant impact on my final recommendations.\n\nQ1- **Unclear evaluation details for adversarial attacks used.** \nThe epsilon values used for attack are starting from 0.1, 0.2 going up to 1., here the attacks l-infinity norm bounded? If yes, then what is the valid image space? Is it [0, 1] or is it [0, 255], meaning when epsilon = 1, does this mean that the epsilon is actually 1/255 (meaning that the valid image space is [0, 255]), or is the value of epsilon actually 1, meaning the entire image is nothing but adversarial noise? In this case, the image would also look semantically different to the human eye meaning that it will no longer be a valid adversarial attack.\nAnd if the epsilon value is in fact 1/255, then the drop in performance is too significant for a very small epsilon value indicating the method is not truly robust to adversarial attacks. Could you also please comment on this?\n\nQ2- **The idea of using Evidential Learning for Pixel-Matching is not entirely novel.** \nWhile the exact downstream task in [3] is different from the one explored by this proposed work, the core ideas for both seem unusually very similar, the key difference being the distributions used, while [3] used a Normal Inverse-Gamma (NIG) distribution, this work uses a Dirichlet distribution. Would you please further highlight the key differences between the two other than some task-related implementation details?\n\n\n**References**\n\n[3] Chen Wang, Xiang Wang, Jiawei Zhang, Liang Zhang, Xiao Bai, Xin Ning, Jun Zhou, Edwin Hancock,\nUncertainty estimation for stereo matching based on evidential deep learning,\nPattern Recognition, Volume 124, 2022,108498, ISSN 0031-3203, https://doi.org/10.1016/j.patcog.2021.108498. (https://www.sciencedirect.com/science/article/pii/S0031320321006749)" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "If the idea of using evidential learning for feature matching is truly novel then that makes the work quite interesting and significant.\nApart from a couple of small typos, the paper is very well written. \nThe structure of the paper and the intended story are easy to follow.\nThe abstract of the paper is well written and to the point." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an interesting idea of evidential learning for certainty estimation for the dense pixel-level feature matching task. \nThe proposed method is supposedly more OOD and adversarial robust than the current SotA RoMa. \nIt is tested against 2D Common Corruptions variants of 2 commonly used datasets for this task i.e. MegaDepth-1500 and ScanNet-1500.\nIt is also tested against outdated adversarial attacks such as FGSM and PGD." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1- **A lot of implementation details are missing from the paper.**\nSimply mentioning that it is built on top of RoMa is insufficient information.\nIt is understandable to do so for the main paper to save space; however, the supplementary material should be used to provide such information, for example, the exact architecture, the training procedure, details about the datasets, HPC resources used, and other details important for reproducibility. \n\nW2- **Needs a stronger argument for why OOD and Adversarial Robustness is important.**\nThe argument made in the introduction to explain why OOD and Adversarial robustness are important for this task can be made significantly stronger. Unfortunately, a case has not been made for why this is interesting and important for the community. \n\nW3- **Out-dated evaluations for robustness.**\nIf the argument for OOD and Adversarial robustness is readiness for the real world, then the evaluations used do not hold up to the argument. Since the 3D Common Corruptions [1] are more real-world common corruptions than the 2D Common Corruptions used in the paper. Additionally, FGSM and PGD attacks were used for evaluating adversarial robustness, however [2] showed in their work that these attacks, originally proposed image classification are inadequate for pixel-wise prediction tasks, such as the one used in this proposed work. This is because FGSM and PGD optimize the attack by increasing the aggregate loss and not the per-pixel loss, this can cause the attack to be highly localized making a non-robust method appear very robust as the mean performance would still be quite well over the rest of the image space. Thus, specialized pixel-wise attacks such as CosPGD are essential for truly evaluating the adversarial robustness of pixel-wise prediction tasks. \n\nW4- **Using 2D Common Corruptions on other known datasets is not always a novel contribution.**\nIt is unclear if the contribution of the 2 supposed OOD Robustness evaluation datasets MegaDepth-1500-C and ScanNet-1500-C is merely using 2D Common Corruptions proposed for ImageNet-1k and CIFAR datasets but changing their resolutions and applying them to the respective iid datasets or if there is more to the story, for example, some unforeseen complications that needed to be handled? If not, then simply applying these corruptions to other datasets is not exactly a novel contribution, it is still an interesting study just not a \"new contribution\" as claimed in the bullet points in the introduction of the paper.\n\nW5- **Almost Redundant Presentation of Results.**\nIncluding both Table 1 and Figure 3 is redundant. I understand that Table 1 contains the mean values over the 5 severity levels while Figure 3 shows the values at each severity, however by using straight dashed lines of respective colors, with y = mean value for all x values the need for Table 1 is eliminated.\n\n\n\n\n**References**\n\n[1] Kar, Oğuzhan Fatih, et al. \"3d common corruptions and data augmentation.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.\n\n[2] Agnihotri, Shashank, Steffen Jung, and Margret Keuper. \"CosPGD: an efficient white-box adversarial attack for pixel-wise prediction tasks.\" Forty-first International Conference on Machine Learning. 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-motivated, and the main idea is clearly explained.\n\n2. Experiments are conducted across a wide range of benchmarks with various types of corruptions and adversarial attacks. The proposed method outperforms in most cases.\n\n3. The paper includes visualizations to analyze why the proposed method performs better than comparison method, particularly on corrupted data across different datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper applies evidential deep learning to feature matching tasks, introducing an evidential learning framework for certainty estimation in dense feature matching problems. The proposed method enhances robustness in dense matching against corruptions and adversarial attacks, with extensive experiments conducted and visualization presented to demonstrate its performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Several questions need to be addressed:\n\n1. This work employs a two-dimensional evidential deep learning (EDL) framework to certainty estimation in both coarse-scale and fine-scale losses. What would happen if EDL were applied to only one of these loss scales? Conducting an ablation study could provide insights into the effectiveness of EDL at each scale. It would be great if the authors could report performance results by applying EDL exclusively to coarse-scale or fine-scale losses, compared to using it on both losses. \n\n2. Experiments are conducted on two datasets, MegaDepth-1500 and ScanNet-1500. There are other datasets mentioned in RoMa paper such as the street-view IMC 2022 and the challenging WxBS Benchmark. Evaluating the proposed method on those different datasets could further demonstrate its generalizability across diverse scenarios. \n\n3. The proposed framework incorporates evidential deep learning into the training process. Could you provide details on how the proposed framework affects computational time, specifically in terms of training and inference times compared to the baseline RoMa method?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How much does the threshold used for balanced sampling matter for the robustness under image degradations? Both in the original RoMa and the new model.\n2. My reading is that computationally, the new certainty estimation is more or less as heavy as the old in terms of inference and training speed. Is this correct?\n3. Is the increased performance due to only better certainties? For example, is the performance the same if certainty scores from the new model are combined with matches from the original RoMa?\n4. Is there a way to make use of the built in uncertainty in the Dirichlet distribution as described in Weakness 3?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper proposes a simple tweak to the RoMa-method, which gives good results experimentally.\n2. Within the framework of the RoMa matcher, certainties are updated iteratively in each warp refinement step by a \"logit-offset\" in the original model. It seems more intuitive to let each refinement step produce positive evidence values for the two classes \"correct\" and \"incorrect\" that are summed over the steps, as is done in this paper. Perhaps, the authors could expand on this in the paper.\n3. In general, outputting good certainties is an underexplored part of deep learning for 3D vision. The application of evidence based learning to dense matchers is novel." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose modelling certainty of correspondences in dense matchers using an evidential deep learning approach. Instead of just estimating a certainty between 0 and 1, the model outputs the parameters of a Dirichlet distribution over probabilities of the two classes \"the predicted correspondence is reliable\" and \"the predicted correspondence is unreliable\". The certainty output is then the expected probability of the first class according to this Dirichlet distribution. The authors show experimentally by retraining the dense matcher RoMa that their approach leads to improved certainty scores, in particular leading to increased robustness to image corruptions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiments are limited in that only a single model is tested. Hence, it is difficult to say if the improvements generalize to other models.\n2. The explanation of evidential deep learning was difficult to understand, and I had to refer back to the original paper by Sensoy et al. I think this section could be improved.\n3. Evidential deep learning has a built-in uncertainty measure. In the context of the present paper, we get a Dirichlet distribution over the classes \"the predicted correspondence is reliable\" and \"the predicted correspondence is unreliable\", and the associated uncertainty describes how spread out this Dirichlet distribution is. This uncertainty is however not used in the present approach. This makes it a bit difficult to interpret the method. We get a Dirichlet distribution but only use its expected value over the first class. This expected value should signify correspondence reliability, but there is also an uncertainty of this prediction inherent in the Dirichlet distribution, which is not used. Since RoMa uses regression-by-classification in the coarse matching step, a more natural approach may be to reformulate the loss for that classification over $N\\times N$ image patches as evidential and use the uncertainty of the predicted Dirichlet distribution as an uncertainty score." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "If EDL is mainly used for certainty estimation, what are the differences or relationships of the proposed method compared to outlier filtering post-processing in feature matching (RANSAC)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The incorporation of EDL to dense feature matching is interesting, and has not been investigated before.\n2. The proposed method enjoys good performance in corrupted data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to unify evidential deep learning (EDL) and dense feature matching, achieving more robust matching results, especially for corrupted image pairs. The authors propose MegaDepth-1500-C and ScanNet-1500-C benchmarks to evaluate the robustness of the proposed method under common image corruptions. The proposed method enjoys superior results in both clean and corrupted data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the point of EDL is interesting, the usage of EDL for certainty estimation in dense feature matching is still questionable. From the introduction in Section 2.3, I think EDL's main advance is to detect out-of-distribution samples or mining pseudo-unknown objects. However, the certainty estimation in feature matching is just a binary classification task (matched or not matched). Why is EDL still effective? The authors did not provide a more insightful discussion about this key question.\n\n2. The overall contribution is limited. Because of lacking enough in-depth discussion about EDL and certainty estimation in feature matching, makes this work appear more as a mere combination of these two approaches rather than a convincing exploration.\n\n3. The introduction of EDL in Section 3.2 is insufficient, missing the necessary background/preliminary in related works.\n\n4. Experiments are not sufficient, missing discussion of visual localization (InLoc and AachenDay-Night) and homography estimation (HPatches). The proposed method achieves significant improvements in corrupted data, while the improvements based on clean data are limited. As a general certainty estimation, the usage of EDL should consistently improve the matching accuracy in all scenarios." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "An evidential deep learning framework to improve the robustness of dense feature matchers" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024evidential,\ntitle={Evidential Learning-based Certainty Estimation for Robust Dense Feature Matching},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4NWtrQciRH},\nnote={under review}\n}" }, "abstract": { "value": "Dense feature matching methods aim to estimate a dense correspondence field between images. Inaccurate correspondence can occur due to the presence of unmatchable region, highlighting the need for certainty measurement. This is typically addressed by training a binary classifier to decide whether each predicted correspondence is reliable. However, deep neural network-based classifiers can be vulnerable to image corruptions or perturbations, making it difficult to obtain reliable matching pairs in corrupted scenario. In this work, we propose an evidiential deep learning framework to enhance the robustness of dense matching against corruptions. We modify the certainty prediction branch in dense matching models to generate appropriate belief masses and compute the certainty score by taking expectation over the resulting Dirichlet distribution. We evaluate our method on a wide range of benchmarks and show that our method leads to improved robustness against common corruptions and adversarial attacks, achieving up to 10.1% improvement under severe corruptions." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Evidential Deep Learning", "Dense Feature Matching", "Pose Estimation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/790294d561a63d2bc2a422f266be24e2e4707c7e.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Evidential Learning-based Certainty Estimation for Robust Dense Feature Matching" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "None" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper has theoretical claims that connects DIET to CL.\n2. This paper has some empirical evidence that the proposed S-DIET can match the performance on benchmarks like CIFAR and ImageNet-100." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies DIET, which is a method that has rather few adoptions. This paper claims that DIET uses fewer memory than common contrastive learning approaches, and proved some theoretical results showing that DIET and spectral contrastive learning share the same solutions. Moreover, this paper proposes a new alternative S-DIET to further improve its performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The method DIET studied in this paper has a fatal limitation, which is that the labels are essentially the sample index. As the dataset size increases, the classification head will need to increase linearly as well, which makes it impractical to use in large scale dataset training. Even though the proposed S-DIET does not require the classification head to be always loaded into the memory, it is still unnecessary to store such a large head, especially when one is training on millions of data." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Please refer to the weakness." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper provides comprehensive and rigorous theoretical proofs.\n2. Addressing the high memory demand of DIET is a well-motivated objective with strong practical significance.\n3. The experimental results presented in Table 5 demonstrate promising improvements.\n4. The paper conducts a detailed and insightful ablation study." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents S-DIET, a memory-efficient modification of DIET for self-supervised contrastive learning. It proves that DIET with a linear encoder and MSE loss is theoretically equivalent to spectral contrastive loss, and proposes feature normalization and projection head use to enhance performance. S-DIET significantly reduces DIET's memory requirements and achieves state-of-the-art performance without extensive hyperparameter tuning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper aims to address three issues: why DIET can perform comparably to CL, DIET's failure to learn all features, and its high memory demand. However, these issues are addressed in a fragmented manner without clear logical connections between them, making it difficult for readers to grasp the paper's central thesis.\n2. The paper does not provide code or pseudocode, which hinders understanding of the proposed method and limits the ability to verify its effectiveness.\n3. It is well-known that MSE is not typically used as a loss function for classification tasks. In the original DIET paper, each sample is treated as a separate class in a classification problem using cross-entropy loss. Why, then, is MSE employed as the loss function in Section 4? Does Theorem 4.5 hold if cross-entropy loss is used instead?\n4. Due to the use of W1, it is unclear how Theorem 4.3 is related to the proposed method.\n5. The theoretical analysis in the paper relies entirely on linear assumptions, while the proposed method (Equation 5) is based on empirical assumptions. These assumptions raise concerns about the rigor of this work." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1) The review in the manuscript is clearly incomplete. Some important literature are missing. The author may want to have a more comprehensive literature review. \n\n2) The analysis in the Section 4 is not surprising due to the linear setting. Can the analysis be extended to the nonlinear case? The author may read the recent development under the nonlinear setting in Wang 2023. \n\n3)Assuming $W_H$ is an isometry seems a very strong assumption. Do the authors put a constraint for $W_H$ in the loss function to enforce such isometry? Otherwise, the authors may consider removing this assumption.\n\n4)The form of training example introduced in Section 5 is unrealistic. When does this assumption hold (at least approximately) in piratical example? Why are there two features, one is low noise and the other is high noise?\n\n5)Can we generalize the results in Theorem 5.1 to a more realistic setting?\n\n6) The idea in memory-efficient DIET is straightforward." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The connection between DIET and spectral contrastive loss is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The manuscript studied the property of a supervised representation learning, called DIET, and then propose an improved version of this method. Specifically, the authors show the equivalence between DIET and spectral contrastive loss proposed by HaoChen \\& Ma under a linear case. In addition, the improvement is motivated by the insight derived from the model introduced in the setting of Section 5.1. Although it looks interesting when all the strong assumptions are true, it is unclear if the results presented in the manuscript can provide guidance in the practical settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The analysis is built on some unrealistic assumptions. It is unclear whether the results and development in the manuscript can hold in practice. It could be more meaningful to make assumptions more carefully." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see **Weaknesses**." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This paper explore the limitation of DIET, an important method of CL, and obtain several conclusions: (i) for linear encoders, DIET with MSE loss is equivalent to spectral contrastive loss; (ii) DIET tends to learn features with less noise but may not capture all relevant aspects of the training data; (iii) feature normalization can help mitigate this issue, while incorporating a projection head can further enhance performance. \n\n- This work further introduces SCALED-DIET (S-DIET) to improve the model's linear probe accuracy, i.e., use batch cross entropy and the multistep update formula for AdamW.\n\n- Some experiments demonstrate the effectiveness of the proposed S-DIET." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study provides a theoretical analysis of DIET, a recently proposed supervised approach for self-supervised learning. DIET, as a menthod of CL, labels each example by its datum index and employs a supervised loss for training. This work obtains several conclusions, including (i) for linear encoders, DIET with MSE loss is equivalent to spectral contrastive loss; (ii) DIET tends to learn features with less noise but may not capture all relevant aspects of the training data; (iii) feature normalization can help mitigate this issue, while incorporating a projection head can further enhance performance. This work further introduces SCALED-DIET (S-DIET) to improve the model's linear probe accuracy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The motivation behind this paper is unclear. According to P2 in the Introduction, DIET's advantage lies in its ability to mitigate CL's reliance on large datasets, while requiring a smaller parameter dimension to balance with sample size. The claim that smaller encoder dimensions are a key drawback for handling large data is not well justified. Furthermore, the authors state, \"not clear whether DIET can capture the pairwise similarities between views...SSL,\" yet DIET, as a CL algorithm utilizing supervised loss, does not explicitly depend on pairwise similarities; this is merely an implementation choice rather than a fundamental mechanism. I fail to see how this motivation strongly connects DIET with contrastive loss, is pairwise similarity closely related to memory? While exploring CL from an efficiency perspective do be valuable, a thorough reading of the paper reveals a lack of such information. Perhaps I overlooked some details, and I hope the authors can clarify their insights.\n\n- The key conclusions in this work need further explanation in relation to the core ides, i.e., MEMORY-EFFICIENT. For instance, the relationships between memory and features, encoder parameter dimensions, and projection heads are not adequately described, leading to fragmented conclusions.\n\n- The choice to study DIET is justified by its independence from large training data, but there are other CL algorithms that also do not rely on large datasets, such as few-shot SSL. Additionally, DIET requires labeled information. Can the analyses in this work be applied to these other methods? If so, what differentiates this work? A broader exploration of algorithms and their mechanisms might enhance the reliability of this study.\n\n- A code that can reflect the idea of ​​algorithm implementation is encouraged, since it is currently only described in 6 lines. At the same time, the introduction of these modules will increase the computational overhead, and related experiments are also necessary, after all, the focus is on memory.\n\n- (Minor) The paper's template appears to differ from the one provided on the official website, such as in the line numbering. Please consider making further corrections.\n\n**I would be happy to reconsider my score if these concerns can be addressed.**" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "The margin difference is not noticeable without a ruler (all the style difference of this paper is due to the usage of a legacy template), and the paper will fit into 10 pages with the correct template. After more cases emerges and based on more discusses, we decided to lean on the lenient side. Please proceed with reviewing of this paper." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": { "value": "We approve the reversion of desk-rejected submission." }, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": { "value": "margin violation" }, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "Submission Desk Rejected by Program Chairs" }, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024memoryefficient,\ntitle={Memory-Efficient Self-Supervised Contrastive Learning with a Supervised Loss},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4NgxI6Z74n},\nnote={under review}\n}" }, "abstract": { "value": "Contrastive Learning (CL) is among the most popular methods for self-supervised representation learning. However, CL requires a large memory and sample size and careful hyperparameter tuning.\nThese factors make it difficult to\nlearn high-quality representations with limited amount of memory. In this work, we theoretically analyze a recently proposed \\textit{supervised} approach, DIET, for self-supervised representation learning. DIET labels every example by its datum index and trains on the labeled data with a supervised loss. DIET does not require a large sample size \nor hyperparameter tuning. However, it falls short when using smaller encoders and is memory intensive due to its massive classifier head.\nGiven its remarkable simplicity, it is not obvious whether DIET can match the performance of CL methods, which explicitly model pairwise interactions between augmented examples. We prove that, perhaps surprisingly, for a linear encoder DIET with MSE loss is equivalent to spectral contrastive loss. Then, we prove that DIET is prone to learning less-noisy features and may not learn all features from the training data. We show feature normalization can provably address this shortcoming and use of a projection head can further boost the performance. Finally, we address the scalability issue of DIET by reducing its memory footprint.\nThe modified approach, namely S-DIET, substantially improves on the linear probe accuracy of DIET across a variety of datasets and models and \noutperforms other SSL methods,\nall with limited memory and without extensive hyperparameter tuning. This makes S-DIET a promising alternative for simple, effective, and memory-efficient representation learning." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "contrastive learning", "self-supervised learning", "representation learning", "machine learning theory" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f0ef8fc90f6b04c9143f1e0c456623cbfdc166f6.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Memory-Efficient Self-Supervised Contrastive Learning with a Supervised Loss" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. What prompt templates are used in various experiments?\n2. How are the textual descriptions $\\theta_t$ initialized in the algorithm? How can the initial tag descriptions meaningfully be assigned to the segments?\n3. Sec 4.1 Evaluation Datasets: How do you convert the input query and the output answer of each example into segments? Do you learn tags only for segments of answers, but not queries/prompts/questions?\n4. Sect 4.1: This is titled Evaluation Datasets but indeed describes data for clustering and tagging.\n5. L195: How do you sample alternative segments?\n6. L188: What do you mean by the test documents? The datasets are query-answer examples.\n7. In the Hits@K evaluation:\n(a) Do you first generate the tag sequence based on the input of a test example or randomly? \n(b) Does a model predict an answer based on the tag sequence in one prompting call?\n(c) How do you evaluate if a sampled solution is correct or not?\n(d) Why do you say the proposed algorithm improves diversity in outputs, which is not evaluated? In fact, diversity is neither necessary nor sufficient for a model to perform a reasoning task well." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This work proposes to discover “tags” for text segments in an unsupervised fashion and a novel algorithm that is inspired by the probabilistic latent semantic analysis (PLSA).\n- The algorithm leverages the ability of an LLM to analyze textual materials and is able to find detailed and meaningful tags, as shown in the qualitative results of the paper.\n- The paper show favorable empirical results compared to multiple baselines: traditional latent Dirichlet allocation, it variant + LLM, prompting, and chain-of-thought prompting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem of discovering text segments that share common characteristics and assigning them the same tag description. It proposes an LLM-based method that iteratively assigns tags to document segments and improves each tag description based on the segment cluster. The authors aim to show that these tags are helpful for a reconstruction task and in improving “Hits@K” accuracy in evaluation sets created from WritingPrompts, MATH, and the BBH benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing in this paper is often too generic and high-level. For example, when a reader reads the motivation at L011-L014 “Humans have the ability to learn new tasks by inferring high-level concepts from existing solutions, then manipulating these concepts in lieu of the raw data. Can we automate this process by deriving latent semantic structures in a document collection using foundation models?”, they may wonder: \n - What tasks do you mean?\n - Existing solutions of the “new tasks” or other relevant tasks?\n - What does it mean to manipulate high-level concepts?\n - How do you define “semantic structures” in a document collection? It’s not precise to describe a set of “tags” as a structure.\n2. The novelty of the method is limited and its connection to PLSA and EM is loose. The proposed algorithm is simple: (1) Initialize a certain number of tag descriptions. (2) Prompt an LLM to assign a tag to each document segmentation based on the tag descriptions. (3) Let an LLM generate a new tag description that describes the shared characteristics of the segments in this cluster.\n - The main Eq. (4) is actually not used: $p(d)$, $p(x_k|d)$, $p_\\Theta(t|x_k,d)$, and $p_\\Theta(w_{1\\dots n}|t)$ are not computed.\n - L153: The parameters $\\theta_t$ are textual descriptions instead of floating-point parameters and no training is happening.\n - L157: No probability distribution is involved. An LLM is employed to greedily perform the steps in the algorithm.\n - PLSA is a generative model of the training documents that it is estimated on, and it is not a generative model of new documents. But this paper aims to find tags that apply to unseen examples.\n3. While the convergence criteria matters for an EM algorithm, this paper simply sets the number of iteration to 30. Not enough analyses is perform on the impact of the number of iterations.\n4. In the reconstruction experiments the method based on learned tags solves a multiple choice problem of picking the ground truth $x_k$ from a set of candidate segments. However, baselines such as prompting in Eq. (7) requires a language model to generate the ground truth $x_k$. These seem not comparable.\n5. Although the experiment results are positive compared to the baselines, the setups are synthetic. Would be nice to see the application of this algorithm to achieve competitive results according to standard evaluation metrics of the used datasets, which are common benchmarks.\n6. Many details are missing. See the Questions below." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "please refer to the weaknesses" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(1) Study a classic task\n\n(2) Propose a new method\n\n(3) conduct some experiments to verify the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an improved version of probabilistic Latent Semantic Analysis (pLSA), termed fPLSA (Foundation-Model-Based PLSA), which incorporates Large Language Models (LLMs) to refine the modeling of latent tags in documents for topic modeling. It conducts some experiments to verify the effectiveness of fPLSA." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) Insufficient technical contribution: The method utilizes LLMs for tag Description Generation. Specifically, the fPLSA model generates descriptive tags for the document segments by prompting the LLM with segments assigned to a particular tag to produce a cohesive summary that represents what these segments have in common. The parameters of the LLM are kept frozen during the process. This means the LLM is not fine-tuned during fPLSA training but is used in a static manner. While the integration of LLMs into pLSA offers a novel approach to document modeling, the core statistical methodologies underlying pLSA (like the EM algorithm) remain largely unchanged. This may limit the perceived novelty from a methodological standpoint.\n\n(2) Missing necessary experiments: need to involving more baselines that use LLMs for topic modeling, like Pham et al. (2024), Wang et al. (2023) mentioned in the paper. \n\n(3) Poor writing: The transition of some contents are abrupt and hard to readers to understand the points, such as the first and second paragraphs in the introduction.\n\n(4) Missing Implementation Details: all the prompts used in the experiment are not specified such as those for fPLSA and GenOutline (a baseline)\n\n(5) Unclear motivation of the experiment setting: the paper uses GPT-4 for clustering and tagging while using ChatGPT to measure the accuracy. The authors explain it’s because GPT-4 may have data contamination issues on some benchmarks. I think this explanation is lame and need more clarifications while potentially leading to unfair comparison." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How would fPLSA perform when applied to a single document at test time, especially if that document differs significantly in structure or content from the training set? Can pre-learned tags from similar training data reliably generalize to new documents in such cases, or would fine-tuning on representative samples be necessary to improve performance?\n- Given fPLSA’s structured tagging capabilities, could the authors discuss its applicability to downstream tasks like structured text summarization and content retrieval? Prior research on text summarization with text segmentation has demonstrated that segmenting texts by themes can enhance summarization quality [1]. Could fPLSA’s tags be similarly used to segment and summarize each thematic section, creating a coherent multi-level summary? Additionally, might these tags support content retrieval or indexing by allowing documents to be searchable by thematic segments? Including a brief paragraph on such applications could highlight the contribution's versatility.\n- Can the authors provide insights into fPLSA’s computational cost compared to the baselines? For instance, would a less resource-intensive model (like a smaller language model) yield competitive results without the same computational burden?\n- How sensitive is fPLSA to the choice of segment granularity (sentence, paragraph, etc.)? In testing, did certain segmentation approaches yield more cohesive or meaningful tags, and if so, could the authors elaborate?\n- Since pre-trained LLMs may encode biases, did the authors observe any potential bias issues during fPLSA’s tagging process? If so, what mitigation strategies might they recommend for fair and balanced tag generation?\n\n**References**\n1. Semantic Self-Segmentation for Abstractive Summarization of Long Documents in Low-Resource Regimes. AAAI 2022." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- **Innovative Approach:** fPLSA is a well-conceived combination of probabilistic topic modeling and LLM-based embedding, creating a tagging system that captures both low- and high-level semantics. This approach enables a nuanced understanding of document structure that extends beyond traditional methods, addressing complex relationships within text segments.\n- **Diverse Evaluation:** The method is rigorously evaluated across multiple datasets, including narrative, mathematical, and multi-step reasoning tasks, demonstrating consistent performance improvements in text reconstruction and sampling diversity. This diversity in datasets reinforces the robustness and generalizability of the approach.\n- **Potential for Cross-Domain Applications:** fPLSA’s ability to structure and tag text meaningfully is a powerful tool for hierarchical content generation, segmentation, and structured summarization, with substantial applications across various domains, such as education, content generation, information retrieval, and summarization.\n- **Foundation for Future Research in Unsupervised Document Tagging:** fPLSA provides a strong foundation for future work in unsupervised document tagging and text segmentation. Its hierarchical tagging approach encourages further exploration in transfer learning, document summarization, and adaptive segmentation, inspiring new research directions for improved document understanding and organization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces fPLSA (foundation-model-based Probabilistic Latent Semantic Analysis), a novel approach for identifying latent semantic structures in document collections by combining traditional PLSA with the contextual understanding of large language models (LLMs). fPLSA enhances probabilistic clustering and unsupervised topic modeling by assigning semantic \"tags\" to document segments through an iterative Expectation-Maximization (EM) process, where each tag captures both local meaning and broader document context. This structured tagging approach enables fPLSA to better capture complex segment relationships, making it valuable for hierarchical sampling, document analysis, and potentially other downstream tasks such as structured summarization. The paper demonstrates fPLSA’s effectiveness across diverse datasets—narrative (story writing), problem-solving (math), and multi-step reasoning—showing improvements in text reconstruction likelihood and Hits@K accuracy, underscoring its robustness and versatility." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Single-Document Applicability:** fPLSA heavily relies on cross-document patterns during training, which is not fully addressed in terms of single-document use cases. At test time, users often only have one document. It would be beneficial to clarify how fPLSA’s pre-trained tags would generalize to individual documents without access to cross-document patterns. For instance, can the model effectively apply pre-learned tags from similar training data to new documents?\n- **Lack of Efficiency Analysis:** Given fPLSA’s reliance on LLMs, a discussion on computational efficiency would be valuable. While LLMs are powerful, they are computationally expensive. Addressing the practical feasibility of deploying fPLSA at scale (or proposing more efficient variations) would make the paper’s findings more actionable.\n- **Potential LLM Biases:** Since fPLSA uses pre-trained LLMs to assign tags, there is a risk of encoding biases from the LLM's training data into the tags. The authors could explore ways to mitigate or assess the impact of these biases, especially for datasets or domains sensitive to fairness and accuracy.\n- **Segmentation Granularity:** The paper does not discuss how sensitive fPLSA is to the choice of segment granularity (e.g., sentence, paragraph) and whether different segmentation approaches yield more cohesive or meaningful tags. Further examination of this could provide clarity on best practices for applying fPLSA across different document types and tasks.\n- **Potential for Downstream Applications:** Although the paper’s results demonstrate fPLSA’s effectiveness in hierarchical sampling, the model's broader potential in downstream tasks is not explored. Given the rich, hierarchical nature of fPLSA tags, they could be valuable for applications like multi-level text summarization, where each tag could represent a theme or section for summarization. Exploring these applications would broaden fPLSA’s impact." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- I'm a bit confused by the relation between w, x, and d. If $w_{1:n}=x_k\\subset d$, how is $p(x_k|d)$ modeled in (4)? Why is it necessary to include both $x_k$ and $d$ as conditional terms?\n\n- Context window: I'm not sure I understand how the segments are selected in this article. Is a segment a fixed-length sequence of tokens? Are there any overlaps between different segments? In Line 238, the authors mentioned \"we use a context window size of 2 on WritingPrompts and use unlimited context window\". How should the \"unlimited context window\" be interpreted?\n\n- According to [1], the latent variable ($z$ in [1]) is supposed to be categorical. This article borrowed the same concept from [1] but I'm not sure whether this article follows the original setup. The authors did mention that they \"set the number of tags to 100\", but the example tags in Table 3 showed that the tags are natural language descriptions rather than categorical labels. I wonder how the tags are generated, and if calling it \"latent\" is still appropriate.\n\n- In (5), $t_k$ is sampled condition on $x_k$, which is later used to estimate the probability of reconstructing $x_k$. Is this a typo? Doesn't this lead to data leakage and make the results of (5) unfairly high?\n\n- For BBH, I'm not sure why it is necessary to \"use the step-by-step solutions produced by their automatic Chain-of-Thought\nprompt inference algorithm for clustering and tagging\". Does it mean that a part of the (ground-truth) solutions is utilized as the prompt to the model for problem-solving? I think this is a huge data leakage issue and would greatly undermine the soundness of the evaluation of the proposed method.\n\n- Since tag generation is a recursive process, what would the token consumption be for achieving the presented results? How about the baseline models?\n\n[1] Hofmann, T. \"Probabilistic latent semantic indexing.\" Proceedings of the 22nd annual international ACM SIGIR conference on Research and development in information retrieval. 1999." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The article is generally well-written except for the technical part, which I find somewhat confusing. According to the article, fPLSA's strengths lie in its enhanced semantic understanding, leveraging LLMs for capturing nuanced document structures beyond lexical co-occurrence. This approach yields more accurate text reconstruction and supports hierarchical sampling, producing diverse, high-quality outputs in applications like story generation and problem-solving. Its specific and detailed tagging outperforms generic LLM-based tags, enhancing content generation. Additionally, fPLSA’s unsupervised clustering reduces the need for labeled data, while its demonstrated adaptability across domains and improved Hits@K accuracy make it a versatile, efficient tool for semantic analysis and structured text generation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces fPLSA, a foundation-model-based extension of Probabilistic Latent Semantic Analysis (PLSA), aimed at discovering semantic structures within document collections through clustering and tagging of text segments. Unlike traditional topic modeling, which often relies on word co-occurrences, fPLSA leverages Large Language Models (LLMs) to understand segment-level semantics in a broader document context. It applies an Expectation-Maximization (EM) algorithm to iteratively refine segment tags, enhancing both text reconstruction accuracy and hierarchical sampling. Experimental results on datasets for story writing, math, and reasoning show that fPLSA significantly outperforms traditional and LLM-based tagging methods in text reconstruction and solution diversity. This makes it suitable for generating effective problem-solving guides and diverse text outputs across varied tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The usage of math symbols is sometimes confusing. Also not required, it is suggested that the authors follow the [default notation](https://github.com/ICLR/Master-Template/raw/master/iclr2025.zip) for a clearer presentation of the equations.\n\n- The proposed method is not thoroughly explained. For example, the computation of some terms in (4) is missing, as well as its optimization algorithm, e.g., how to calculate $p(x_k|d)$. From my perspective, if one chooses to express the idea using math formulas, then every term should be clearly explained except for the cases where it is extremely obvious, which I think does not apply to (4).\n\n- A figure or algorithm may better explain the proposed method.\n\n- The authors use GPT-4 for clustering and tagging but GPT-3.5 for response generation and did not provide experimental results on other combinations. The performance of the proposed method therefore may not be universally applicable.\n\n- Potential data leakage issues (detailed in Questions).\n\nOverall, I think the approach proposed by this article is rather straightforward and could be easily described with better clarity without introducing any formulae, perhaps except for the motivation part. In addition, it seems that this article may find a broader audience in the pure NLP community rather than a mixed community of different machine learning topics. Therefore I would recommend to submitting this manuscript to ACL ARR venue instead of machine learning conferences." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024fplsa,\ntitle={f{PLSA}: Learning Semantic Structures in Document Collections Using Foundation Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4NsYCAxubi},\nnote={under review}\n}" }, "abstract": { "value": "Humans have the ability to learn new tasks by inferring high-level concepts from existing solution, then manipulating these concepts in lieu of the raw data. Can we automate this process by deriving latent semantic structures in a document collection using foundation models? We introduce fPLSA, a foundation-model-based Probabilistic Latent Semantic Analysis (PLSA) method that iteratively clusters and tags document segments based on document-level contexts. These tags can be used to model the structure of given documents and for hierarchical sampling of new texts. Our experiments on story writing, math, and multi-step reasoning datasets demonstrate that fPLSA tags help reconstruct the original texts better than existing tagging methods. Moreover, when used for hierarchical sampling, fPLSA produces more diverse outputs with a higher likelihood of hitting the correct answer than direct sampling and hierarchical sampling with existing tagging methods." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Natural Language Processing", "Large Language Models", "Document Analysis", "Latent Semantic Analysis" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/45e6e2aca511f56b805e726c2ebb74e92a7084f7.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "fPLSA: Learning Semantic Structures in Document Collections Using Foundation Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- What is the size of the test set (out of 2000 images subset) used for evaluation? \n- On line 329, the authors mention \"almost correctly classified by all the evaluated models\" about the chosen subset. What are the exact accuracy numbers for every network?\nRefer to weakness section for more comments." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The attack is transferable across multiple network architectures, within CNNs as well as from CNNs to transformers.\n- Intuitive, novel method for increasing transferability of the black-box attacks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a transferable black-box attack using the concept of model augmentation through channel pruning and knowledge distillation. The authors show that the transferability of existing black-box attacks is limited due to their uneven focus on the channels. The authors also introduce a gradient regularization to enhance the transferability further. The evaluation is done using a subset of 2000 images from ImageNet." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper does not follow the format guidelines closely. Tables 1 and 2 are out of the paper margin significantly. \n- The presentation of the paper needs to be improved. The paper contains many grammatical errors, like a period in the middle of the sentences and incorrectly capitalized words. For example, on line 354, it should be \"Table\" instead of \"table\"; \",,\" on line 355; should be \"pruning\" instead of \"running\" on line 252; the sentence on lines 226-227 is incomplete; there is a blank line in Table 3 and 4 before the first result row. \n- What is the size of the test set (out of 2000 images subset) used for evaluation? \n- On line 329, the authors mention \"almost correctly classified by all the evaluated models\" about the chosen subset. What are the exact accuracy numbers for every network?\n- Limitations like an increase in computing due to model augmentation and the trade-off between the transferability and number of models should be discussed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. What are the key theoretical differences between pruning techniques like channel, kernel, and block pruning? Do these techniques exhibit different attack performances in both white-box and black-box settings?\n\n2. Could other model augmentation-based attack methods be further improved by incorporating the proposed initial parameter optimization and gradient regularization strategies?\n\n3. Could the authors provide additional details on how adversarial examples are trained on the ensemble of pruned models? In the overall loss function, does knowledge distillation by minimizing L_{ce} conflict with the perturbation training, which involves maximizing L_{ce}? Additionally, how can the proposed five loss terms be effectively trained, and how should their corresponding hyperparameters be adjusted?\n\n4. Could the authors provide additional details on why gradient regularization results in only slight changes in ASR across different layers of the surrogate model? Additionally, why is gradient regularization applied to only one layer rather than multiple layers?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors propose using channel pruning to enhance the transferability of adversarial attacks, along with knowledge distillation to recover the classification accuracy of pruned models.\n\n2. During the perturbation training, regularization of important feature maps is introduced to reduce the gradient variance to further enhance the attack performance.\n\n3. Experimental results on various target models demonstrate that the proposed transfer-based method outperforms baseline methods in attack effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a model augmentation-based method to improve the transferability of adversarial attacks. To enhance performance in black-box settings, they first introduce the technique of channel pruning to create a self-ensemble surrogate model, which mitigates the overfitting on redundant features or channels. Additionally, the authors integrate both the knowledge distillation and gradient regularization into this ensemble model to further enhance the transferability across various target models. Experiments conducted on multiple CNN and ViT networks using the ImageNet benchmark dataset demonstrate that the proposed model augmentation method achieves relatively high attack performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation for introducing channel pruning is not clearly explained. For example, why was channel pruning selected over other pruning techniques, such as kernel or block pruning? Furthermore, as noted in lines 255-257, the L2-norm distance metric lacks robustness due to its sensitivity to outlier noise. Does this criterion maintain consistent performance across different surrogate models?\n\n2. The submitted manuscript lacks certain details regarding initial parameter optimization. For example, the network structure of the self-supervision (SS) module is not specified. Additionally, the authors introduce the L_{ss} loss, as defined in lines 295-296, to maximize the similarity of all positive and negative data pairs, which is inconsistent with standard contrastive learning practices.\n\n3. In Section 4.3, the authors propose regularizing gradient variance within an intermediate feature map but do not specify how this layer is selected within the surrogate model. Additionally, as mentioned in lines 322-323, the overall loss function is designed to minimize each loss term including the L_{ce} loss, which may conflict with the goal of crafting adversarial examples by maximizing L_{ce}. Furthermore, the details regarding the optimization of this overall loss function are not clearly explained.\n\n4. As shown in Tables 3-4, the experiment is conducted on only one surrogate model, which does not provide a comprehensive evaluation. Furthermore, the proposed GRASP method is not directly comparable to other model augmentation methods, as these methods do not incorporate knowledge distillation or gradient regularization.\n\n5. The authors claim that reducing gradient variance can balance the importance of different channels within a layer. However, as shown in Figure 5, the minimal change in attack success rate after regularizing different layers does not provide strong support for this claim." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- Why are some channels redundant? Why can this explained with over-parameterization? (Line 47.)\n- Why are the pruned surrogate models containing denser information if initial parameter optimization and regularization applied? (Figure 1)\n- Why do samples overfit on the highly repetitive channels? \n- How are the hyperparameters set in 4.3?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper is easy to read through." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a method for a transfer-based blackbox attack.The main idea for the method is to augment the surrogate models by differently pruned models, and generate the adversarial examples. First to augment the models, the model is pruned with the predetermined channel pruning rate. To train the augmented models, the knowledge distillation is used to match the accuracy of the pruned model, incorporating self-supervision and input augmentation as terms in the loss function. The authors present explanations and experiments to further demonstrate their idea. The experiment is done on the ImageNet-like dataset on targeted and untargeted attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weakness of the paper is three-folds.\n- The terminology used in the paper lacks objective justification and includes many subjective and inaccurate expressions.\n-- See questions.\n- Lack of Novelty.\n-- The main methodology comes from the idea of channel-pruning based model augmentation. There are plenty of existing studies that analyze channel-wise pruning in terms of robustness, and numerous studies demonstrate that model augmentation can achieve higher transferability. [1-3] As it stands, there doesn't seem to be much take-away for the reader from the combined ideas in this study.\n- The overall explanation and experiments are insufficient.\n-- Lack of Validation to the proposed method.\nThe experiments are limited to a single dataset, which does not demonstrate generalized results.. \nhe paper primarily explains the performance of the proposed method through intuition and explanation, but lacks supporting evidence. Adding intermediate empirical or theoretical validations for the proposed method would improve its persuasiveness.\nTo enhance the persuasiveness of this study, more ablation studies are needed.\n\n-[1] Bai et al., \"Improving Adversarial Robustness via Channel-wise Activation Suppressing\", ICLR 2021\n-[2] Borkar et al., Defending Against Universal Attacks Through Selective Feature Regeneration, CVPR 2020\n-[3] Tramer et al., \"Ensemble adversarial training: Attacks and defenses\", ICLR 2018" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "-\tWhy did you do not discuss in the related work the loss-based transfer methods such as for example [C] or [D]? \n-\tFor the contrastive learning, how did you choose these three transformations? Why not other ? \n\n[C] Naseer, M., Khan, S., Hayat, M., Khan, F. S., & Porikli, F. (2021). On generating transferable targeted perturbations. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 7708-7717).\n\n[D] Zhao, A., Chu, T., Liu, Y., Li, W., Li, J., & Duan, L. (2023). Minimizing maximum model discrepancy for transferable black-box targeted attacks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (pp. 8153-8162)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "-\tStudying the transferability of adversarial examples is an important topic.\n-\tThe paper is easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Adversarial examples have recently received much attention in the black-box transfer-based attack scenario due to its more realistic attack setting. To enhance the transferability of the generated adversarial example, the paper introduces GRASP, a model augmentation method that uses channel pruning to generate different models. These different pruned models are used to generate an adversarial example that may be less specific to the potential channel redundancy of the source model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-\t**Experiments are insufficient and do not sufficiently support claims.** The comparison of the proposed method against other model augmentation methods is only done for one source model (ResNet50). Despite the source network being the same for Table 3 and Table 4, the targeted networks are different, which is surprising. The superiority of the proposed method cannot be established. Moreover, since the proposed method uses three pruned networks to generate the adversarial example, I expect the authors to compare their method against an ensemble method to ensure that the gain obtained is not due to ensembling networks.  It would be interesting to compare against ensemble-based methods such as [A] and [B]. \n\n\n- **Motivation of the method.** In the introduction, the authors say, “On the contrary, some\nof them are vague and seem to only contain weak features about the object in the input image. If\nwe only conduct adversarial attacks based on original CNN models, the adversarial samples tend to “overfit” on those highly repetitive features.” I do not understand why, if a feature is not useful for predicting an object (whereas this feature may be useful for another object), this feature would be exploited by an attack that tries to fool the network. I would expect the attack to disrupt features responsible for predicting the target class and not those that are not useful. To validate their intuitions, can the authors provide an experiment or proof showing that adversarial examples tend to “overfit” those highly repetitive, unuseful features? Moreover, I find it strange that increasing the pruning rate, which is the core of the method, degrades the transferability of the generated adversarial examples. Can the authors clarify this point, please? \n\nTypos:\n- In Table 3 and 4, there is a blank row. \n- Table 1 and 2 are overextended. \n\n[A] Tang, B., Wang, Z., Bin, Y., Dou, Q., Yang, Y., & Shen, H. T. (2024). Ensemble Diversity Facilitates Adversarial Transferability. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 24377-24386).\n\n[B] Chen, B., Yin, J., Chen, S., Chen, B., & Liu, X. (2023). An adaptive model ensemble adversarial attack for boosting adversarial transferability. In Proceedings of the IEEE/CVF International Conference on Computer Vision (pp. 4489-4498)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024enhance,\ntitle={Enhance the Transferability of Adversarial Attacks through Channel Pruning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4NtrMSkvOy},\nnote={under review}\n}" }, "abstract": { "value": "Recent studies have shown that neural networks are vulnerable to adversarial attacks, where attackers generate adversarial samples by imposing tiny noise. The tiny noise can not misguide human perception, though leading the neural networks to generate wrong predictions. Transfer-based black-box attacks play a more significant role in recent studies due to their more realistic setting and considerable progress in performance. Previous studies have shown that some different channels of the same layer in convolution neural networks (CNN) contain lots of repetitive information, and we find that existing transferable attacks tend to exploit those redundant features more, which limits their transferability. Hence, we advocate using channel pruning and knowledge distillation to conduct model augmentation. In addition, we introduce a method of regularization on the gradients of intermediate feature maps of augmented models, which further enhances the transferability of our method. Comprehensive experiments demonstrate that imposing our method of model augmentation on existing methods can significantly improve the transferability of adversarial attacks in untargeted or targeted scenarios. Furthermore, our method outperforms state-of-the-art model augmentation techniques without the usage of additional training datasets." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "adversarial attacks transferability", "channel pruning", "model augmentation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/657596a1ed284099bffc1ea50db75407c6cc146f.pdf" }, "presentation": null, "primary_area": { "value": "alignment, fairness, safety, privacy, and societal considerations" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/cd3cd0f87f68487a10b286a556c8e7ee2803be02.zip" }, "title": { "value": "Enhance the Transferability of Adversarial Attacks through Channel Pruning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How do the authors anticipate their findings will generalize to other complex reasoning tasks not covered in the study? Can the observed ineffectiveness of self-critique mechanisms be extrapolated to different types of LLMs or reasoning models? \n2. Could the authors elaborate on the choice of domains for the study? Why were these specific domains chosen, and how do they represent the broader spectrum of reasoning tasks applicable to LLMs? \n3. What additional mechanisms or modifications do the authors suggest could potentially improve the self-verification capabilities of LLMs? Is there ongoing work to develop more effective internal critique mechanisms within LLMs? \n4. How do the authors envision the impact of their findings on the future development and deployment of LLMs in safety-critical applications? What precautions or additional measures would they recommend based on their study’s outcomes?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper addresses a novel aspect of LLMs—self-critique and iterative verification—that is underexplored in the existing literature. It challenges the assumption that LLMs can effectively self-critique by demonstrating that external verification offers more reliable improvements. \n2. The experimental setup is clearly described, allowing for reproducibility and understanding of how iterative prompting affects LLM performance in reasoning tasks. The paper methodically outlines its methodology and the rationale behind using specific domains for testing (Sections 4 and 3). \n3. The findings significantly contribute to understanding LLM limitations in self-verification tasks, which is critical for deploying these models in real-world applications where accuracy and reliability are paramount. \n4. The study is well-structured and robustly empirically analyzed, providing a comparative assessment of LLMs with and without external sound verifiers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper evaluates the self-verification abilities of LLMs in reasoning and planning tasks using iterative prompting and critiquing. It contrasts the performance of LLMs self-verifying their solutions against external sound verifiers across three domains: Game of 24, Graph Coloring, and STRIPS planning. Findings indicate that LLMs underperform in self-verification and that external sound verification significantly improves the accuracy of solutions. The study suggests the ineffectiveness of self-critique mechanisms in LLMs and recommends integrating external verifiers for better performance in reasoning tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper’s focus on only three specific domains might limit the generalizability of the findings. While these domains are relevant, more varied tasks could provide a broader understanding of LLM capabilities across different reasoning types (Section 3). \n2. The analysis of the self-critique mechanism lacks depth regarding why LLMs fail at self-critique. Specific instances of LLM outputs and their failures would enrich the discussion by pinpointing the flaws in LLM reasoning strategies (Section 5.1). \n3. There is no detailed discussion on the computational cost and efficiency of using external verifiers versus self-verification. This information would be crucial for practical implementations where resource constraints are a significant consideration. \n4. The paper does not thoroughly explore the theoretical implications of its findings on the computational complexity theories surrounding LLMs and self-verification. A deeper theoretical analysis could provide insights into the fundamental limitations of LLM architectures (Section 2)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See point 2) in weaknesses and more specifically: are there any arguments that the findings do not only hold for the examined LLMs but in general for transformer-based auto-regressive models, or even other models like discrete diffusion models.\n\n\nMinor comment:\nThe use of \\citep and \\citet is not correct. Also there are some small spelling/typesetting issues here there that should be easily fixable." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) The paper is really well written and easy to to follow. The non-technical nature might have helped here. \n2) I enjoyed reading the (extensive) related work section where problems in prior work are made explicit.\n3) All claims made in the paper are substantiated by appropriate experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This is an experimental paper studying the the popular technique of self-verification of LLMs for enhancing apparent reasoning capabilities. The stringent experimental protocol shows that the self-verification does not really work and shows that other techniques like using formal (symbolic) verifiers are better suited to achieve automated reasoning with LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) For someone familiar with the field the findings might be a little bit obvious. However, given the amount of papers being published using self-verification, I nevertheless believe this to be an important study when it comes to combating the delusion of self-verification.\n\n2) Given that this is an experimental study it might be good to also point out weaknesses in the experimental protocol. Specifically, making explicit all the assumptions that were made and what might change if they were not to hold. Concretely, what gives the authors the confidence that their findings have a high probability of standing the test of time." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please discuss the points mentioned in the weakness part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The presentation and writing are clear.\n2. The authors made adjustments to the test datasets to ensure their validity. For example, they generated new instances to prevent test set memorization and modified the task solution format to better evaluate the self-verification pipeline.\n3. The systematic analysis of self-verification provides insights into the cases where the verification module can help improve LLM performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper discusses the effect of self-verification in LLM reasoning and planning, challenging the assumption that verifying correctness is easier than generating solutions. Through experiments in three domains (Game of 24, Graph Coloring, and STRIPS planning), it finds that self-verification does not improve accuracy and can even reduce performance. In contrast, a \"sound external verifier\" significantly enhances accuracy, with simpler re-prompting methods maintaining most of these benefits." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiment design is not entirely persuasive.\nThe authors attempt to challenge the assumption that verifying correctness is easier than generating solutions, suggesting that self-verification does not improve performance. However, task difficulty should be considered. Apart from Blocksworld, the accuracy in other tasks is quite low. According to Table 2, \"LLM Verification Results,\" verification performance varies across tasks, especially in terms of the False Negative Rate (FNR). In Blocksworld, which has the lowest FNR, self-verification actually improves task accuracy from 40% to 55%. This suggests there are cases where verification is easier than generation and where self-verification contributes positively to task accuracy. More tasks should be added for more comprehensive results.\n\n2. The authors use a “sound” verifier as a comparison, but it’s unsurprising that a ground-truth verifier significantly improves performance. With ground-truth verification, the model can eliminate incorrect answers and focus on sampling correct ones, or at least provide ground-truth information. Taking the first point into account, a more nuanced conclusion could be that in tasks where verification is easier than generation, self-verification helps; otherwise, it has no benefit or even harms performance. The improvement limit for self-verification is thus bounded by the effectiveness of the “sound” verifier. \n\n3. The exact GPT version should be specified, and more models, particularly advanced ones, should be tested for comprehensive results." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "I don't have more questions but authors are encouraged to address my concerns in the weakness section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper has the following strengths.\n\nOriginality:\nThis paper systematically evaluates LLMs' self-verification capability by careful ablation and comparison to symbolic processors. It finds that LLM verification is not very helpful in some planning cases. This is novel and interesting.\n\nQuality:\nThis paper considers problems of memorization and lack of ground truth during evaluation, which are important concerns.\n\nClarity:\nThe experiment setup is quite clear.\n\nSignificance:\nKnowing what LLMs are actually doing when people claim they can do a lot of things is very important." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper mainly evaluates the self-critique and verification capabilities of LLMs in comparison to sound verifiers. It concludes that the self-verification loop conducted by LLMs is not as helpful as previous works claimed. On the contrary, it might hurt LLM performance in some planning domains." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Is it fair to compare LLMs with oracle verifiers? The finding that LLM self-critique can sometimes downgrade their generation performance is interesting, but oracle processors are not always accessible in all tasks (as the authors mentioned in their paper, some tasks such as creative writing do not have a ground truth answer). I'm not surprised that oracle verifier is improving LLM performance, but I wonder if it is possible that LLM can serve as a decent alternative when sound verifier is absent in general areas?\n\n2. The task domains are constrained while the general conclusion is very strong. The authors evaluated three main planning tasks and one LLM (GPT-4) with four datasets, with no clear objective as to why the conclusion drawn from these three specific tasks can be generalized. It is unclear why these tasks and datasets can represent general LLMs' lack of self-verification capability in a wide ranges of tasks.\n\n3. Some other details in this paper are also overclaimed, e.g., on page 2, the authors claim \"...the state-of-the-art GPT-4\" while GPT-4 is not SotA in many benchmarks anymore (in [1], inter alia).\n\n4. There are quite some typos. Writing needs to be done with more care.\n\n[1] Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., ... & Ganapathy, R. (2024). The llama 3 herd of models. arXiv preprint arXiv:2407.21783." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024on,\ntitle={On the self-verification limitations of large language models on reasoning and planning tasks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4O0v4s3IzY},\nnote={under review}\n}" }, "abstract": { "value": "There has been considerable divergence of opinion on the reasoning abilities of Large Language Models (LLMs).\nWhile the initial optimism that reasoning might emerge automatically with scale has been tempered thanks to a slew of counterexamples--ranging from multiplication to simple planning--there persists a wide spread belief that LLMs can self-critique and improve their own solutions in an iterative fashion.\nThis belief seemingly rests on the assumption that verification of correctness should be easier than generation--a rather classical argument from computational complexity--which should be irrelevant to LLMs to the extent that what they are doing is approximate retrieval.\nIn this paper, we set out to systematically investigate the effectiveness of iterative prompting in the context of reasoning and planning.\nWe present a principled empirical study of the performance of GPT-4 in three domains: Game of 24, Graph Coloring, and STRIPS planning.\nWe experiment both with the model critiquing its own answers and with an external correct reasoner verifying proposed solutions.\nIn each case, we analyze whether the content of criticisms actually affects bottom line performance, and whether we can ablate elements of the augmented system without losing performance. We observe significant performance collapse\nwith self-critique and significant performance gains with sound external verification.\nWe also note that merely re-prompting with a sound verifier maintains most of the benefits of more involved setups." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large Language Models", "Reasoning", "Planning", "Self-Critique", "Verification" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/59b96e7d81b724fb9f2c26a190c4ca157110d100.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "On the self-verification limitations of large language models on reasoning and planning tasks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Do you have an intuition on why SAM doesn't perform better on Walker2d-v3 for high friction factor?\n- Have you tested SAM+PPO on non-MuJoCo environments to assess robustness in discrete action spaces or varying reward structures?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper provides a formal link between flat reward surfaces and robustness in policy space. Proposition 1 establishes a clear theoretical foundation for the paper's main claim.\n- The authors comprehensively test SAM+PPO across multiple challenging environments and scenarios, including noisy actions and varying transition probabilities, to demonstrate robustness.\n- The authors compare SAM+PPO with RNAC, PPO, and RARL, which shows both performance and computational efficiency, which strengthens their findings.\n- The use of reward surface visualizations and flatness metrics strengthens the paper's argument by providing visual and quantitative evidence for the flatness achieved by SAM+PPO." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates the relationship between flat reward maxima in policy parameter space and the robustness of reinforcement learning (RL) agents. It claims that flatter reward maxima lead to more robust policies, particularly against action perturbations. The paper presents a theoretical proposition linking flat reward to action robustness and supports this claim through empirical experiments in MuJoCo environments (e.g., Hopper-v3, Walker2d-v3, HalfCheetah-v3). The authors demonstrate that an RL algorithm enhanced with Sharpness-Aware Minimization (SAM), called SAM+PPO, consistently outperforms standard PPO and a recent robust RL baseline (RNAC) in various robustness tests, including action noise, transition probability changes, and reward function variations. The paper also provides visualizations and quantitative measurements of reward surfaces, further confirming the link between flatness and robustness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While SAM is shown to be effective, the paper lacks a discussion of its potential limitations, such as computational overhead or sensitivity to hyperparameter tuning.\n- The justification for reward noise being added during training for reward function robustness evaluation could be clearer: The paper mentions this difference in methodology but could expand on why this is necessary for a valid evaluation.\n- I don't know if the preliminary experiment is best placed in the introduction, it feels a bit out of place for me.\n- typos 234 \"objeective\", 249 \" funciton\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Is there a specific advantage to using PPO with SAM, or could any PG or even AC algorithm be used? It might be that the clipping approximation to the trust region synergizes well with the SAM objective? I think this is an optional extension to the paper." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The authors propose a simple yet intuitive approach for robust RL. I was somewhat surprised that this combination has apparently not been tried in the literature, but a brief literature survey has not brought up any similar algorithms. I actually think the authors are somewhat underselling their contributions here! While SAM has been used to train PPO before, the authors appropriately cite prior work here, previous papers have not drawn any connections to robust RL at all and the authors should feel entitled to proudly claim this connection as their connection! They do not merely provide theoretical backing, as far as I can tell, they make a connection that was wholly absent in cited work.\n\nThe theoretical statements are mostly correct as far as I can tell. See questions below however." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a study on using sharpness-aware regularization to obtain robust reinforcement learning policies. Drawing a theoretical connection between flatness in the reward, action and parameter space to action-robust RL, the authors present both a theoretical justification and experiments to show that the proposed method achieves good robustness properties." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main problem with the paper as it stands are writing problems and baseline comparisons.\n\nEspecially the beginning of the paper, abstract and introduction, suffer from very frequent grammar mistakes which make the paper much harder to read. I strongly encourage the author to revise the paper wrt to the writing.\n\nIn definition 1, I'm unsure if $\\epsilon$ is added to the policy, parameters or action? From the proof it seems this is a parameter perturbation, this should be stated directly. I think adding parentheses in the equation would already make this much clearer, as we have two nested subscripts here.\nIn addition, the state is sampled from the policy, which seems strange?\n\nAs the theoretical statement depends on the Jacobian of the policy network, which is not bounded anywhere, I'm slightly skeptical that the theoretical results are sufficient to practically guarantee robust RL. Does the SAM objective guarantee or incentivize a flat Jacobian?\n\nGiven the surprisingly (?) bad results of RNAC - it barely seems to outperform PPO - I think it would be appropriate to apply SAM+PPO in the same environments as used in the RNAC paper. As far as I can tell, the code is available, so this should be feasible within the rebuttal timeline? If not, I will not hold this against the authors. I think it is important to verify that used examples are not cherry-picked to make the presented algorithm look stronger. This is the higher priority comment in terms of baseline comparisons.\n\nI would encourage the authors to present some additional baselines. I acknowledge that more baselines is a somewhat lazy comment. However, given that there are several different formulations of robust RL, I believe it would be helpful to pick a variety of environments and algorithms presented with different robust formulations for comparison to understand how well the algorithm does in comparison to others. This doesn't have to be many or complex environments, just a larger variety of formalisms. This is a soft concern and not a large barrier to acceptance for me.\nBoth safe-control-gym [1] and Safety Gymnasium [2] provide a variety of tasks and implemented baselines to speed up experimentation.\n\n[1] https://github.com/utiasDSL/safe-control-gym\n[2] https://github.com/PKU-Alignment/safety-gymnasium" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Is the perturbation domain $\\rho$ in Eq.8 known to the agent? Probably the optimization of the objective in Eq.8 needs elaboration, and with pseudo-code.\n\n- Why in \"Nominal\" SMA+PPO still has a higher reward, e.g. Table 1+2, Fig. 3,4. Similarly, experiment in 5.2, why when action noise is small, i.e. even equal to 0, SAM+PPO still performs better than the others, because the objectives of PPO and SAM+PPO would converge to the same one? And in 5.3, SAM+PPO has a higher return, while with variation in Friction Coefficient shows mixed results. \n\n- Joint variation of friction and mass shows quite clear that SAM+PPO is performing better than baselines, except on Walker2d-v3 with a mixed result. Can the authors elaborate on why or provide ablation to explain the mixed performance of SAM+PPO?\n\n\n- The proof of proposition 1 is a bit not standard. The policy is sometimes referred as a distribution, but sometime used as a deterministic mapping. It needs revised." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Provide a link of flat reward to action robustness. The authors show this through both theoretical results in section 4, and various experiment results. The motivation of having a robust objective is good. The theoretical result seems correct.\n\n- Positive experiment results showing the benefit of optimizing for a flat reward maxima. The authors show this through different experiment settings: variation to physics properties of the underlying MDP, and visualization of the reward surface." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the impact of flat minima in reinforcement learning (RL), linking flatter reward surfaces to improved model robustness. The authors show that flatter rewards lead to more consistent actions despite parameter changes, enhancing robustness against variations in state transitions and reward functions. The authors show through extensive experiments to confirm that flatter rewards significantly bolster RL model performance across diverse scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The performance of SAM + PPO is mixed in comparisons to the baselines, e.g. some visible ones at Fig 5.c, 4.b.\n\n- Ablations are not provided to understand how such an objective can bring benefits in comparisons to similar approaches, e.g. RNAC or robust RL." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "I'm really curious about \"flat rewards\" in general. Definitions 1 and 2 seem too strict at first glance (the equalities therein), so it is actually a bit surprising to me that they are even possible at all; however IIUC, Fig 6 does give evidence of this. I think that these definitions can be further elaborated on (do you have a toy example where it is easy to see in parameter or action space?) Realistically, what values of $\\epsilon$ do you think are reasonable? Something like $10^{-11}$ or $10^{-2}$? (I might've missed it somewhere, sorry.) If these are novel definitions not previously given in the literature, that can be stated as a contribution of the paper. I think it can spark future work in both theory and experimental directions.\n\nHere are some follow up questions/comments:\n\n- In Sec 5, how long are those agents trained for? Equal number of env steps for each? How were hparams tuned for each algo?\n- What is the agent's action scale for these environments (cf L337)? What do you do if the noise added is outside the action range?\n- Do you have any ideas about the sharp dropoff in Fig 3b for SAM PPO? it looks interesting, but I'm not sure what to make of it... is there some \"critical\" mass ratio? I.e., if we zoom in, how sharp is that transition, and have you averaged over enough random seeds?\n- you mention \"flatter reward maxima\" in L70. I think a formal definition or good visualization of this phenomenon early on would really improve the paper.\n- How does this work relate at all to other trust region methods like TRPO? How about e.g. [1]\n\n[1]: https://arxiv.org/abs/2103.06257\n\nTypos/minor\n\n- Fig 3 caption \"nomial\"\n- some missing +/- signs in Table 1 (in parens)\n- citations in sec 2 often have a missing leading space.\n- can you improve the visual in Fig 1? I think it's important but not quite capturing the essence. Maybe just to remove axes and grid and zoom in a bit: is there indeed a channel for the agent? It's hard to see\n- The introduction paragraphs have some grammatical issues. A cleanup/re-write here can help to crystallize the main message early on\n\nWith a rewrite to clean up the presentation, deeper explanation for SAM (i)-(iv), and perhaps a few more visualizations, this could be a really strong paper; but unfortunately I don't think it's quite there yet." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This use of robustness in policy parameter space seems to be fairly new\n- The experiments demonstrate a strong performance boost across a range of perturbations \n- The visualizations in Figure 6 offer an interesting insight into the optimizations produced by PPO vs SAM+PPO. The Hopper example is quite striking. Could you elaborate on the distinction and sharp dropoffs seen there?\n- Theory provides a potential link between flatness in parameter space and action robustness\n- Provided a solid comparison wrt computational overhead / sample complexity and wall time versus other algorithms\n- Figure 5 is quite nice, I think it should be emphasized\n\nOverall, the paper seems like a nice first step in the direction of understanding the relationship between robustness in reward, policy parameter, and dynamics spaces. The notion of \"flat rewards\" is an interesting one." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes a new method to ensure robustness in RL based on variations in the loss landscape: \"SAM\": Sharpness-Aware Minimization. By posing the policy optimization as a min/max objective with respect to perturbations in the parameter space, the authors show robustness to changes in reward and dynamics. A theoretical result is given, linking parameter and reward robustness, and a diverse set of experiments on 3 MuJoCo environments is provided." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Writing:**\n\n- Overall, I think the clarity of the paper can be enhanced with a re-write fixing grammar and overall structure:\n- It would be helpful for example, to also include some visualizations of the definitions 1 & 2.\n- Also, Proposition 1 and the following remarks are not very clear. As an example, for Prop1, if I understand correctly, the result would be better phrased as \"if $\\mathcal{E}$-flat, then $\\Delta$-robust, with $\\Delta \\leq ...$ otherwise the current phrasing is a bit confusing.\n\n\n**Discussion:**\n\n- The discussion of the main idea, \"SAM\" is lacking:\n- After it is introduced in Sec 3.3, the authors give a way to solve the optimization problem in Eq (3) by their steps (i)-(iv). However, (to me at least), it is not clear why this method is used. Is there prior work demonstrating the efficacy of this method? Are there experiments or maybe some minimal example illustrating the utility of this setup? E.g. why is $\\epsilon$ chosen to be in the direction of the previously computed gradient, if theoretically it should represent an arbitrary direction in the ball. \n- At the very least, can the authors provide some visual demonstration as to what is happening here in the loss landscape? Getting a better intuition would help to understand the core method of the paper. \n- Remark 1.1 seems to be a restatement of Prop 1 unless I am missing something. Could you please explain?\n- Remark 1.2 can be improved by using more technically accurate statements (i.e. what is meant by \"when a reward function slightly changes\")? What is meant by the \"direct [correspondence] to the changes of loss function in the supervised learning case\"? I think the latter is very unclear, and maybe even misleading.\n\n\n**Experiments:**\n- My only issue with the experiments (minor) is that you are missing RNAC in Table 2 (why?). Also why not compare against RARL? Missing explanation of the shaded regions in each figure caption." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024flat,\ntitle={Flat Reward in Policy Parameter Space Implies Robust Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4OaO3GjP7k},\nnote={under review}\n}" }, "abstract": { "value": "Investigating flat minima on loss surfaces in parameter space is well-documented in the supervised learning context, highlighting its advantages in model generalization. However, limited attention has been paid to the reinforcement learning (RL) context, where the impact of flatter reward in policy parameter space remains mostly unexplored. Beyond the naive guessing from the lesson of supervised learning, which makes us anticipate a link from flat rewards to the enhanced generalization, we here aim to formally bridge the flatness in reward surface to the robustness of RL models. For a policy model case, where the deep model determines actions, the flatter behavior of rewards against the parameter perturbations primarily leads to consistent rewards against perturbed actions. Moreover, the action robustness further affects to the robustness against other variations from the changes of state transition probabilities and reward functions. We extensively simulate various RL environments, confirming the consistent gains of flatter reward in bolstering the robustness of RL in varying circumstances, including action selection, transition dynamics, and reward functions." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Reinforcement learning", "Flat Minima", "Robust Reinforcement learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/11a15eb57bf648fedf0b0d59787bd7e338cdf1da.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Flat Reward in Policy Parameter Space Implies Robust Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Besides the concerns raised in the weakness part, I have the following additional questions:\n\n1. PPO has also been applied to solve other combinatorial optimization problems like routing problems, where the horizons are also very large. Could you give some intuitions why PPO is particularly unstable for this problem?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well-written and easy to follow\n2. Special designs on the actor and critic network to have more efficient embeddings and long-range interaction modeling\n3. customized gradient for stabilizing the PPO training for the online settings\n4. Experiments are conducted on many online and offline settings" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel Graph-assisted Offline-Online Deep Reinforcement Learning (GOODRL) approach for Dynamic Workflow Scheduling (DWS) in cloud environments. The authors introduces three main innovations: a task-specific graph representation with a Graph Attention Actor Network for focused task assignments, a system-oriented graph representation with a Graph Attention Critic Network for real-time state evaluation, and a hybrid offline-online RL method to improve adaptability. The offline stage uses imitation learning for stable initial policy development, while the online stage applies advanced PPO techniques for continuous adaptation. Experiments demonstrate GOODRL’s superiority over state-of-the-art methods in minimizing mean flowtime and enhancing performance in several offline and online settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper presents a comprehensive learning pipeline for addressing the dynamic workflow scheduling (DWS) problem. I appreciate the authors for their efforts in adapting various components to suit the unique challenges of DWS.\n\nThe primary concern with this paper is the applicability of the proposed pipeline. Many of the modifications and design choices appear closely tailored to DWS, leaving it unclear how generalizable this approach might be to other scheduling problems, such as flexible job shop scheduling. Can these designs be readily adapted for other problem domains? The paper would be significantly strengthened by demonstrating the pipeline’s transferability to other scheduling scenarios.\n\nSeveral techniques are introduced throughout the pipeline, though not all are rigorously validated in the ablation study. A more thorough investigation into the contributions of each component would enhance our understanding of their individual benefits.\n\nOverall, I appreciate the contributions of this work and currently lean toward a borderline acceptance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In a real-world data cluster, how does the arrival pattern change over time? Consider plotting the 95%-99% quantile of the number of simultaneous arrivals. Please also consider adding other more background information.\n2. How long does the model take to make a decision at inference time, compared to prior approaches?\n3. It seems that in Table 1 and 2, the arrival rates during training and testing are always the same for each scenario. However, in a real-world data center, the arrival pattern might fluctuate over time, especially in the case of extreme events (e.g., holidays or deadlines). How robust is your approach to such distribution shifts?\n4. In Table 1 and 2, \"Ours-Offline\" already achieves a very good performance. If due to distribution shifts, the offline version gets a much lower performance, can the online algorithm quickly adapt to such changes?\n5. Many real-world scenarios involve some type of resource contention or performance interference. For example, two tasks are both memory intensive, so maybe we should allocate them on different machines. How does GOODRL address this issue?\n\nMinor:\n- Line 45, \"In fact existing\" --> \"In fact, existing\"" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper has many diagrams, which help the readers to understand.\n- GOODRL shows strong performance against other baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Prior works often consider homogeneous setups and static conditions, and fails to consider the dynamic nature of the workflow scheduling problem. To this end, the paper proposes a GNN-based DRL approach with an offline stage as well as an online stage." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- My main concern is that the paper seems to be applying some existing algorithms to a scheduling problem. There are some simple modifications at different parts of the overall method, and ended up giving us good performance. However, what are the broader insights of this work?\n- Consider adding more background on the DWS problem and studies on real traces." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How does this model differ from the model presented in Wang et al 2022 [7]?\n\n- How does the heterogeneity of the agent-task is accounted for in the graph representation?\n\n- It is unclear what the novelty of this work is compared to similar works published in Multi-Agent Coordination, and Task Allocation and Scheduling Domains." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Clear and detailed explanation of the approach being used.\n- Easy to understand figures.\n- Compared against Multiple Benchmarks and show that the model outperforms the baselines in most instances" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Graph assisted Offline-Online Deep Reinforcement Learning (GOODRL) for Dynamic Workflow Scheduling for Cloud Computing. The presence of heterogenous configurations, dynamic arrival of workflows, and constant evolving environment makes this a challenging problem for State of the Art Models.\n\nThe Contributions presented in the paper are:\n\n1) A Task-Specific Graph Representation and Graph Attention Actor Model that dynamically assign focused tasks to heterogenous machines.\n2) Explicit Consideration of the future impact of the crucial state.\n3) A combination of Offline Imitation Learning followed by Online PPO." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The topic of Resource Opimization using Graph Neural Networks is an open problem that has applications not limited to Cloud Computing. The problem itself is also explored under Multiple Travelling Salesman Problems [1, 2], Vehicle Routing Problem [3], Job Shop Scheduling [4] and Task Allocation and Scheduling in Multi-Agent Systems [5, 6]. While the application of the problem into Cloud Computing is novel, the use of Reinforcement Learning and Graph Attention Networks to similar optimization problems exists.\n\n- It is unclear how the proposed method differs from Online Predictive Scheduling using Heterogenous Graph Attention presented in Wang et al 2022 [7]:\n\n- The enhancement provided by the Online RL part of the model is unclear. The experimental results show that the Offline Learning allows for the model to be within 2% of the Online Training results. The significance of this improvement is unclear and needs to be discussed clearly.\n \n[1] Yujiao Hu, Yuan Yao, and Wee Sun Lee. 2020. A reinforcement learning approach for optimizing multiple traveling salesman problems over graphs. Knowledge-Based Systems 204 (Sept. 2020), 106244. https://doi.org/10.1016/j.knosys.2020. 106244\n\n[2] Yujiao Hu, Zhen Zhang, Yuan Yao, Xingpeng Huyan, Xingshe Zhou, and Wee Sun Lee. 2021. A bidirectional graph neural network for traveling salesman problems on arbitrary symmetric graphs. Engineering Applications of Artificial Intelligence 97 (Jan. 2021), 104061. https://doi.org/10.1016/j.engappai.2020.104061\n\n[3] Steve Paul and Souma Chowdhury. 2022. A scalable graph learning approach to capacitated vehicle routing problem using capsule networks and attention mechanism, Vol. 86236. American Society of Mechanical Engineers, V03BT03A045\n\n[4] Song, Wen, et al. \"Flexible job-shop scheduling via graph neural network and deep reinforcement learning.\" _IEEE Transactions on Industrial Informatics_ 19.2 (2022): 1600-1610.\n\n[5] Z. Wang and M. Gombolay, \"Learning Scheduling Policies for Multi-Robot Coordination With Graph Attention Networks,\" in IEEE Robotics and Automation Letters, vol. 5, no. 3, pp. 4509-4516, July 2020, doi: 10.1109/LRA.2020.3002198.\n\n[6] B. Altundas, Z. Wang, J. Bishop and M. Gombolay, \"Learning Coordination Policies over Heterogeneous Graphs for Human-Robot Teams via Recurrent Neural Schedule Propagation,\" _2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)_, Kyoto, Japan, 2022, pp. 11679-11686, doi: 10.1109/IROS47612.2022.9981748.\n\n[7] Wang, Z., & Gombolay, M. (2022). Stochastic Resource Optimization over Heterogeneous Graph Neural Networks for Failure-Predictive Maintenance Scheduling. _Proceedings of the International Conference on Automated Planning and Scheduling_, _32_(1), 527-536. https://doi.org/10.1609/icaps.v32i1.19839" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Could the authors elaborate on how their model adapts to significant changes in workflow patterns or cloud configurations without extensive retraining? How does the method maintain robust performance under dynamic conditions that differ from the training data?\n- What are the practical deployment requirements for GOODRL in terms of computational resources, and how do they compare with simpler heuristic-based solutions in large-scale, real-time applications?\n- Can the authors provide more insights into how the method handles noisy or incomplete data, which is a common challenge in real-world cloud scheduling environments?\n- How might the proposed approach be extended or adapted to incorporate multi-objective optimization, such as balancing energy efficiency with flowtime reduction, and what specific challenges would need to be addressed to achieve this?\n- Could the authors comment on potential scalability issues when deploying GOODRL in larger cloud infrastructures or in environments with highly heterogeneous machine configurations, and how these challenges could be mitigated?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper proposes a unique combination of graph representations tailored for actor and critic networks, enhancing action differentiation and value estimation.\n- GOODRL demonstrates improvements in mean flowtime over established baseline algorithms, showcasing robust performance in both offline and online settings.\n- The offline-online learning approach with imitation learning and gradient control addresses challenges in adapting to dynamic environments, adding practical value.\n- The ablation studies and performance comparisons are thorough, providing strong evidence for the contributions and architectural decisions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces an innovative approach, GOODRL, for handling dynamic workflow scheduling in cloud computing environments. This method integrates a task-specific graph representation with Graph Attention Networks (GATs) for actor-critic networks and incorporates offline imitation learning alongside online reinforcement learning to adapt to changing conditions. The proposed system is evaluated in various online and offline scenarios and is shown to outperform existing state-of-the-art methods in terms of mean flowtime and adaptability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper's reliance on simulations limits its generalizability to real-world cloud environments. Practical tests in real cloud data centers would bolster the validity of the results.\n- The experiments primarily focus on a specific set of workflow types and machine configurations, potentially limiting the applicability of findings to other types of DWS problems.\n- The computational overhead associated with the proposed GAT-based architectures is not discussed in detail, raising questions about deployment feasibility in large-scale, real-time applications.\n- While the method performs well in flowtime reduction, other practical objectives, such as energy efficiency and cost, are not explored, which would be valuable for broader applicability.\n- The paper lacks discussion on how the model generalizes to varied workloads, impacting its robustness in dynamic cloud environments." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper proposes an offline-online DRL framework that uses novel graph representations to effectively and efficiently schedule dynamic workflows across heterogeneous machines, significantly improving flowtime compared to state-of-the-art methods." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024graph,\ntitle={Graph Assisted Offline-Online Deep Reinforcement Learning for Dynamic Workflow Scheduling},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4PlbIfmX9o},\nnote={under review}\n}" }, "abstract": { "value": "Dynamic workflow scheduling (DWS) in cloud computing presents substantial challenges due to heterogeneous machine configurations, unpredictable workflow arrivals/patterns, and constantly evolving environments. However, existing research often assumes homogeneous setups and static conditions, limiting flexibility and adaptability in real-world scenarios. In this paper, we propose a novel *Graph assisted Offline-Online Deep Reinforcement Learning* (GOODRL) approach to building an effective and efficient scheduling agent for DWS. Our approach features three key innovations: (1) a *task-specific* graph representation and a *Graph Attention Actor Network* that enable the agent to dynamically assign focused tasks to heterogeneous machines while explicitly considering the future impact of each machine on these tasks; (2) a *system-oriented* graph representation and a *Graph Attention Critic Network* that facilitate efficient processing of new information and understanding its impact on the current state, crucial for managing unpredictable workflow arrivals/patterns in real-time; and (3) an *offline-online* method that utilizes imitation learning for effective offline training and applies gradient control and decoupled high-frequency critic training techniques during online learning to sustain the agent’s robust performance in rapidly changing environments. Experimental results demonstrate that GOODRL significantly outperforms several state-of-the-art algorithms, achieving substantially lower mean flowtime and high adaptability in various online and offline scenarios." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "workflow scheduling", "graph attention neural network", "reinforcement learning", "online learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/8da076d1ee04d72fbd76639c6c98dceda72b2962.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Graph Assisted Offline-Online Deep Reinforcement Learning for Dynamic Workflow Scheduling" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Line 290: It is not clear about how the first gradient term in Eq. 4 would lead to optimising the LLM policy to generate higher quality rationales. Further elaboration is needed as to why generating/optimising the likelihood of input question/instruction conditioned on the rationale would lead to better reasoning.\n2. Lines 414-415: Please support with examples about why improvements on ARC-Challenge are relatively lesser. The magnitude of improvements is not at all an issue, however, it should be better demonstrated that why better reasoning chains are not leading to higher improvements. Does this make ARC-Challenge ill-suited for this study since it involves limited reasoning scope?\n3. Line 430: Why is it needed to generate rationales that are as long as 500 tokens? It would be good to show the usefulness of the information contained in the longer rationales and why does it lead to better performance before saturating." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Enhancing the reasoning ability of LLMs without relying on feedback from any external LLM during the training phase is an important research question.\n2. Suitability of applying the variational objective is well motivated, explained and justified in Sections 3 and 4.1.\n3. Accuracy gains over the base and the SFT version of the LLMs are shown with some ablation studies on greedy decoding vs. self consistency based sampling. Further, the authors show that inference time scaling of number of reasoning samples obtained using LaTRO can additionally enhance the accuracy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work focuses on enhancing the reasoning abilities of large language models (LLMs) during the training phase without relying on external feedback. The authors motivate the work by raising an important question on improving the reasoning ability of LLMs during training phase since most prior works have focussed at achieving this at inference time. To do so, the authors propose to sample diverse reasoning rationales from latent distribution and optimize the LLM through a variational framework using the sampled rationales. The intuition behind application of the variational framework is well motivated through the objective of self-consistency based chain-of-thought prompting. Further, the work proposes to usethe likelihood of generating correct answer conditioned on a rationale as a proxy to explicit reward models to optimise the LLM towards generating better reasoning chains. Results demonstrate that the proposed LaTRO method helps in improving the accuracy achieved for multiple LLMs such as Mistral-7B, Llama-3.1-8B etc. on GSM8K and ARC-Challenge datasets compared to the corresponding base and the SFT versions of the LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Presentation of the introduction section of the paper needs improvement - It should better motivate the need behind reducing the reliance on external LLMs/feedback for improving a given LLM. Further, it should provide information about why the variational framework is suitable and elaborate some details about the proposed method in the introduction itself.\n2. Discussion on related work is very limited. Even though the proposed method is an effort to reduce reliance on external LLMs, it should still discuss and contrast against existing methods that leverage external LLMs (for example, [1,2]) as well as employ self-play based techniques to improve an LLM by itself with requiring help from any other LLM (for example, [3, 4]). The example references are only indicative (but not exhaustive) of the type of citations that should be included.\n3. Lack of appropriate baselines - Even though the work claims and focusses at improving the base/SFT version of the LLM through the proposed LaTRO framework, it should still compare against some of the existing self-play trainable methods (by adapting them for 7B LLMs) as baselines (eg. [3] - on datasets where the ground-truth rationale is available and [4]). Such methods improve the rationales generated by an LLM by exploring the generation space and discriminating better rationales from the inferior ones.\n4. More reasoning datasets such as CSQA, Hellaswag etc. should be considered for evaluation studies.\n\n[1] Distilling Step-by-Step! Outperforming Larger Language Models with Less Training Data and Smaller Model Sizes. Cheng-Yu Hsieh, Chun-Liang Li, Chih-kuan Yeh, Hootan Nakhost, Yasuhisa Fujii, Alex Ratner, Ranjay Krishna, Chen-Yu Lee, and Tomas Pfister. In Findings of the Association for Computational Linguistics: ACL 2023, pages 8003–8017, Toronto, Canada. Association for Computational Linguistics.\n\n[2] Zephyr: Direct Distillation of LM Alignment. Lewis Tunstall and Edward Emanuel Beeching and Nathan Lambert and Nazneen Rajani and Kashif Rasul and Younes Belkada and Shengyi Huang and Leandro Von Werra and Clementine Fourrier and Nathan Habib and Nathan Sarrazin and Omar Sanseviero and Alexander M Rush and Thomas Wolf. First Conference on Language Modeling, 2024.\n\n[3] Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models. Zixiang Chen and Yihe Deng and Huizhuo Yuan and Kaixuan Ji and Quanquan Gu. ICML 2024.\n\n[4] Self-Rewarding Language Models. Weizhe Yuan, Richard Yuanzhe Pang, Kyunghyun Cho, Xian Li, Sainbayar Sukhbaatar, Jing Xu, Jason Weston" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "see weakness" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The writing is clear and easy to follow.\n- The discussed topic and motivation are both innovative and significant." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposesLaTRO, a novel framework aimed at enhancing the reasoning capabilities of LLMs. LaTRO addresses the challenge of improving reasoning during the training phase by formulating reasoning as sampling from a latent distribution and optimizing it through variational approaches. This method allows LLMs to self-improve their reasoning process and ability to evaluate the quality of reasoning without external feedback or reward models. The paper validates LaTRO's effectiveness through experiments on GSM8K and ARC-Challenge datasets, demonstrating significant improvements over base models and supervised fine-tuning approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Although I'm not familiar with the topic discussed in this article, I believe the experiments presented are too few and not comprehensive enough. Moreover, the datasets considered are limited to only GSM8K and ARC-Challenge, which lacks persuasiveness.\n- The number of case study examples is too limited, with only a few instances in Figure 4 and the appendix, which is not convincing.\n- The proposed method, LaTRO, especially the stability of the Self-reward component, has not been adequately considered." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In Proposition 1, $p(y|x) = \\int p(y|z,x) p(z|x) dz$ holds for any CoT-based method. Why, then, is CoT-SC introduced here?\n\n2. What is the definition of \"golden rationales\", and why can’t ARC-Challenge have golden rationales?\n\n3. What can the experimental results on ARC-Challenge demonstrate? The two baselines are too weak, as the SFT baseline did not utilize rationales." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper offers a novel perspective by framing reasoning as a process of sampling from a latent distribution and addressing it through variational methods.\n\n2. The paper leverages the model's own probability estimates as an implicit reward, unifying the training of the policy and reward models.\n\n3. The paper is well-organized and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces LaTRO, a novel approach that formulates Chain-of-Thought (CoT) reasoning as sampling from a latent distribution, optimized through variational techniques. By leveraging the probability of generating the correct answer as an implicit reward, LaTRO unifies the learning of both the policy and reward models, allowing large language models to refine reasoning paths in a self-rewarding manner. The authors demonstrate LaTRO’s effectiveness through experiments on the GSM8K and ARC-Challenge datasets across various model architectures. Their results indicate that latent reasoning capabilities within pre-trained language models can be unlocked and enhanced using this self-improvement framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experimental setup lacks sufficient strong baselines, which are essential for a robust evaluation. Two key baselines to consider are:\n - A stronger SFT Baseline: Fine-tuning the policy model with correct reasoning paths. Given the availability of ground truth answers, multiple reasoning paths could be sampled, retaining only those that align with the ground truth. This baseline would provide a more rigorous comparison for evaluating LaTRO’s effectiveness in reasoning.\n - DPO Baseline: The authors could further fine-tune the policy model using the DPO algorithm, incorporating both correct and incorrect reasoning paths. Actually, the DPO algorithm aligns closely with LaTRO in its approach, as both methods aim to avoid training an explicit reward model. Including DPO as a baseline would highlight LaTRO’s strengths relative to an approach that similarly leverages implicit reward mechanisms.\n\n2. The experimental scope is limited, as only two datasets and small models were tested. \n - Expanding the experiments to include a wider range of reasoning datasets would better assess the model's reasoning capabilities. Standard practice for evaluating reasoning in large language models includes diverse datasets that cover arithmetic reasoning (only GSM8K is not enough for arithmetic reasoning evaluation), commonsense reasoning, symbolic reasoning, and other reasoning types. Incorporating these would provide a more comprehensive evaluation.\n - Testing across varying model scales, especially with larger models, could provide insights into how the approach scales with model size and whether larger models yield better reasoning performance.\n\n3. Although the authors claim that training did not rely on external feedback, the ground truth answers effectively serve as a form of implicit external feedback or reward." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. About self-evaluation task in LaTRO: how does LaTRO evaluate the probability of each reasoning path producing the correct answer? What does the conditional probability mention in Section 4.1 mean? Is there a task restriction for this evaluation method? What is the accuracy of self-evaluation? The authors did not discuss these questions in the paper, nor did they explore them in depth in the experiments.\n\n2. The authors only used a formula to explain the use of self-reward signals to achieve parameter updates, so what exactly does this update parameter refer to during the training phase? How is it implemented? Suggest provide more detailed information." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. LaTRO regards the reasoning process as sampling from a latent distribution and optimizes it using a variational method. This approach is different from prompt-based methods such as CoT and is closer to unsupervised learning. Besides, the feasibility of LaTRO is verified by mathematical proof.\n\n2. This paper focuses on a very interesting topic, which enables an LLM to improve itself through a self-reward mechanism without external supervision and feedback signals." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces LaTRO, a framework that enhances LLM’s reasoning abilities by treating reasoning as sampling from a latent distribution and optimizing it with variational methods. LaTRO allows LLMs to improve their reasoning process and evaluation of reasoning quality simultaneously, without external feedback. Experiments on GSM8K and ARC-Challenge datasets demonstrate the effectiveness of LaTRO compared with the SFT training method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Request a step-by-step description or flowchart of the LaTRO pipeline. Although a large number of formulas are used in this paper to explain each part，but it lacks a detailed description of the proposed method, which makes it difficult for readers to understand the complete pipeline of the proposed method and reproduce it. \n\n2. As far as I know, there are some works to gradually enhance the capabilities of LLM (not limited to reasoning task), including prompt-based [1][2][3] and training-based methods [4][5][6], some of which do not use any feedback information to enhance the capabilities of LLM [7][8]. The author should discuss the differences between this work and these works and compare the performance of these works. It is necessary to add a dedicated subsection in the Related Work section discussing these specific works and their methods. If possible, include performance comparisons with these methods in the experimental section.\n\n[1] When can llms actually correct their own mistakes? A critical survey of self-correction of llms\n\n[2] Learning From Mistakes Makes LLM Better Reasoner\n\n[3] Mirror: A Multiple-perspective Self-Reflection Method for Knowledge-rich Reasoning\n\n[4] REFINER: Reasoning Feedback on Intermediate Representations\n\n[5] CRYSTAL: Introspective Reasoners Reinforced with Self-Feedback\n\n[6] SELF: Language-driven Self-evolution for Large Language model\n\n[7] Small language modes can self-correct\n\n[8] Think Thrice Before You Act: Progressive Thought Refinement in Large Language Models\n \n3. In the experiments, the author only employs SFT training method and the base model as baselines, without comparing the performance with COT-SC mentioned in Section 3 and the self-improvement classic work mentioned above, making it difficult to demonstrate the advantages of the proposed LaTRO. So, please provide a more comprehensive analysis of how LaTRO performs relative to these additional baselines across different tasks and metrics." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See \"Weaknesses\"." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The structure of this paper is clear, and it's easy to follow the main idea and the contribution.\n2. The research problem this paper targets is very important to the community.\n3. The motivation is sound and the benefits of the approach are clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a principled framework, LaTRO, to treat the reasoning process as sampling from latent distribution and enable LLMs themselves as reward models to evaluate the quality of reasoning rationales. The proposed LaTRO outperforms supervised fine-tuning baselines on 2 datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although modeling the rationales as latent variables to sample is well defined and proposed, the paper lacks a discussion with previously proposed reasoning methods that formulate the reasoning process as latent variables [1, 2] as well. Even though they are cited in the related work. Specifically, [1] proposes to sample diverse reasoning rationales to improve the prediction performance and also proposes an EM-like algorithm to improve the reward model and LLM alternatively. It would be great to have a discussion and comparison with these methods.\n2. The paper does not compare with various prompting-based reasoning baselines mentioned in the related work section, such as tree-of-thought[3], and RAP[4], as well as fine-tuning baselines such as STaR[5], which is a missed opportunity to demonstrate its effectiveness. It would be better to compare them with metrics like training / inference computational cost and accuracy.\n3. The paper does not provide a confidence interval, leading to unawareness of how the proposed LaTRO is robust to the initialization and randomness. It would be great to report the results from at least 3 times repetitions.\n4. As the proposed LaTRO is a general approach, it would be great to evaluate it on more benchmark datasets to verify its effectiveness, such as HumanEval[6] and MBPP [7], which are popular in the current LLM reasoning community.\n\n\n[1] Hu, Edward J., et al. \"Amortizing intractable inference in large language models.\" arXiv preprint arXiv:2310.04363 (2023).\n\n[2] Hoffman, Matthew Douglas, et al. \"Training chain-of-thought via latent-variable inference.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[3] Yao, Shunyu, et al. \"Tree of thoughts: Deliberate problem solving with large language models.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[4] Hao, Shibo, et al. \"Reasoning with language model is planning with world model.\" arXiv preprint arXiv:2305.14992 (2023).\n\n[5] Zelikman, Eric, et al. \"Star: Bootstrapping reasoning with reasoning.\" Advances in Neural Information Processing Systems 35 (2022): 15476-15488.\n\n[6] Chen, Mark, et al. \"Evaluating large language models trained on code.\" arXiv preprint arXiv:2107.03374 (2021).\n\n[7] Austin, Jacob, et al. \"Program synthesis with large language models.\" arXiv preprint arXiv:2108.07732 (2021)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024language,\ntitle={Language Models are Hidden Reasoners: Unlocking Latent Reasoning Capabilities via Self-Rewarding},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4Po8d9GAfQ},\nnote={under review}\n}" }, "abstract": { "value": "Large language models (LLMs) have shown impressive capabilities, but still struggle with complex reasoning tasks requiring multiple steps. While prompt-based methods like Chain-of-Thought (CoT) can improve LLM reasoning at inference time, optimizing reasoning capabilities during training remains challenging. We introduce LaTent Reasoning Optimization (LaTRO), a principled framework that formulates reasoning as sampling from a latent distribution and optimizes it via variational approaches. LaTRO enables LLMs to concurrently improve both their reasoning process and ability to evaluate reasoning quality, without requiring external feedback or reward models. We validate LaTRO through experiments on GSM8K and ARC-Challenge datasets using multiple model architectures. On GSM8K, LaTRO improves zero-shot accuracy by an average of 12.5\\% over base models and 9.6\\% over supervised fine-tuning across Phi-3.5-mini, Mistral-7B, and Llama-3.1-8B. Our findings suggest that pre-trained LLMs possess latent reasoning capabilities that can be unlocked and enhanced through our proposed optimization approach in a self-improvement manner." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large language model", "Optimizing LLM reasoning capabilities", "Self-improvement", "Reward model-free optimization", "Reinforcement learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/a01dd007cfd17c70238397405c1e340140150e36.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Language Models are Hidden Reasoners: Unlocking Latent Reasoning Capabilities via Self-Rewarding" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "My main concern is that there are lots of assumptions made in this paper that are unsupported by evidence or experiments, I would change my opinion if the authors present more ablation experiments that carefully study each of their design decisions. I would also love to see more qualitative examples (instead of just descriptions). Lastly, the authors should give more details when comparing the baselines. Are the performance gain simply caused by a better network architecture or a bigger network capacity?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The authors do a good job summarizing previous findings about causal confusion problems in self-driving tasks\n* Proposed lots of interesting strategies to potentially solve or alleviate the causal confusion problems" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present a new imitation learning algorithm that aims to solve some causal confusion problems in previous imitation learning methods on self-driving tasks. Specifically, the authors propose to 1) learn a more representative state representation; 2) reduce the chance of learning spurious-correlation by inferring delta actions from latent states, and 3) weight training samples by the discrepancy between prediction and ground truth." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Authors made lots of assumptions:\n * It's not sufficient to directly map ot to at\n * This remains an untested hypothesis \n * learning a decoder for \\hat{s}t helps it match with expert st\n * on the contrary, learning a decoder for \\hat{s}t could force the encoder to focus on every detail in the image, even the ones that do not directly contribute to ground st.\n * Since delta(at) is inferred from (st), it doesn't learn the spurious correlation\n * This assumption can be wrong since st would contain information from a(t-1)\n * The proposed method is better than baselines in most scenarios, but is that because of the design choices or just better models or bigger capacities?\n* Figure 2 could have more annotations\n * It would be better if the authors could annotate the different colors and shapes of each node" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The improvements in those scenarios look promising. But, how to demonstrate it's coming from reducing the confounding factors as shown in previous sections?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Formulate the imitation learning problem from a causal perspective and tries to prevent confounding factors using representation learning.\n- Proposes to use supervised contrastive learning to learn an image representation that aligns with expert actions.\n- The improvements in the tested scenarios is promising." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to solve causal confusion problem in imitation learning using supervised contrastive learning, residual prediction and sample weighting. It draws insights from causality that motivates to learn a representation of history observations without spurious correlations.\nExperiments on CARLA shows solid improvements over baselines, such as CIL and Premier-TACO." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of comparisons with some related work, such as Wen et.al. Key-frame focused visual imitation learning, which proposes a weighting strategy based on action predictability.\n- It would be nice to have more quantitative and qualitative analysis of the improvement. Can we attribute those to improvement in reducing spurious correlations?\n- Lack of evaluation on CARLA benchmark instead of self-constructed scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the Weaknesses part." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The proposed framework is well-motivated.\n\nThe experimental result looks good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To deal with the causal confusion problem in imitation learning, authors take the inspiration from causal learning and propose the 3CIL framework which integrates contrastive learning, action residual prediction, and importance weighting techniques together. Testing on autonomous driving benchmark CARLA, 3CIL achieves good performance with higher success rate and lower collision times than the baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tFor importance weighting, would you please intuitively or theoretically explain the motivation for using the errors of action residual prediction as weights, instead of other choices, e.g., AdaBoost or Keyframe? Or would you please compare with them?\n\n2.\tWhy is it necessary to divide the imitating process into two separate stages? I suggest comparing with training the representation modules and the policy end-to-end.\n\n3.\tIn Table 1, why is the performance of 3CIL in scenario 3 worst across all scenarios? Even worse than the unseen ones, i.e., 5 and 6.\n\n4.\tMore ablation studies, analysis experiments, and visualizations are necessary. The current experimental results only contain the comparison with baselines and two simple ablation studies. I suggest having more experiments and visualizations to verify your arguments that 3CIL successfully removes the spurious correlation and importance weighting correctly finds the rare scenarios.\n\n5.\tMissing some important references [2,3,4].\n\n\n[1] Keyframe-focused visual imitation learning\n\n[2] Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst\n\n[3] Fighting Fire with Fire: Avoiding DNN Shortcuts through Priming\n\n[4] Shaking the foundations: delusions in sequence models for interaction and control" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "### **Some high-level questions**\n1. In \"3CIL: Causality-Inspired Contrastive Conditional Imitation Learning,\" the abbreviation \"CIL\" is first defined in line 162. However, it’s unclear what the \"C\" stands for. Is it \"Causality\" or \"Conditional\"? Could you clarify this term for consistency?\n\n2. Does the input for the VAE include all the past history observations $o_{1:t}$?\n\n3. Some important details are in the appendix. The authors should consider move them to the major context, e.g., Fig. 4 and Fig. 5. \n\n### **Some questions about theoretical guarantee**\n1. As discussed in (Ruan et al., 2022; Ruan & Di, 2022; Kumor et al., 2021), unobserved confounders (UCs) often complicate causal identifiability, and certain variables must be considered to mitigate the influence of spurious correlations or shortcuts during the learning process. In your approach, which specific variables are crucial to achieving this objective? E.g., which variables are required to block all the backdoor paths.\n\n2. Following up on the previous question, if all the past ground-truth states $s_{1:t}$ are unobserved (common in POMDPs), and $a_{t-1} \\leftarrow s_{t-1} \\rightarrow s_{t} \\rightarrow a_{t}$ is active, in other words, there is an active backdoor path between $a_{t-1}$ and $a_{t}$ which can never be blocked. Given that this path cannot be blocked, is the policy learning process identifiable or robust? How do you ensure convergence in policy learning under these circumstances? \n\n3. The idea of training a representation model $G$ is not that novel. Especially, when $\\hat{s}_{t}$ is unobserved, it is very hard to directly determine whether the representation model is good or not. While simulations may provide insights, evaluating the model’s practical effectiveness in real-world conditions can be significantly harder. What methods do you suggest to compare model performance outside of a simulation environment?\n\n4. The proposed method appears to be a pipeline structure (i.e., representation model + policy model).\n \n 4.1 IIf overall performance is not expected, what strategy would you recommend to isolate and improve the specific component responsible? How can one effectively determine whether limitations stem from the representation model or from the policy?\n \n 4.2 Will there be any cascaded errors from upstream to downstream tasks? Could you elaborate on any mechanisms in place to mitigate such cascading errors?\n\n### **Some questions about experiments**\n\n1. In your experiments, does the imitator have access to the reward $R$? Additionally, are the expert demonstrations generated from RL algorithms that use the same reward function $R$? \n\n2. For the observations, the expert is able to observe $s_{t}$, but the imitator is only able to observe $o_{t}$. Is that correct?\n\n3. To what extent does the reward $R$ reflect real-world driving behaviors? How accurately does it capture the dynamics observed in actual driving scenarios?\n\n4. Could the authors add more details to the reward $R$? Specifically, how are the four components $r$ defined? If the primary objective is to evaluate route adherence, is $r_{position}$ alone sufficient, or are the other rewards essential? Please clarify the role of each reward component.\n\n5. Additional metrics could enhance the evaluation process, such as the RMSE between predicted positions and target routes. Would the authors consider including these metrics for a more comprehensive analysis?\n\n**Should the authors address these questions thoroughly, I would consider raising my evaluation score.**\n\n---------\n\nReferences:\n- Pearl, Judea. Causality. Cambridge university press, 2009.\n- Peters, Jonas, Dominik Janzing, and Bernhard Schölkopf. Elements of causal inference: foundations and learning algorithms. The MIT Press, 2017.\n- Ruan, Kangrui, and Xuan Di. \"Learning human driving behaviors with sequential causal imitation learning.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 36. No. 4. 2022.\n- Ruan, Kangrui, et al. \"Causal imitation learning via inverse reinforcement learning.\" The Eleventh International Conference on Learning Representations. 2023.\n- Kumor, Daniel, Junzhe Zhang, and Elias Bareinboim. \"Sequential causal imitation learning with unobserved confounders.\" Advances in Neural Information Processing Systems 34 (2021): 14669-14680." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The idea of combining causality, contrastive learning and conditional imitation learning is quite interesting. The performance seems to be good for all scenarios. Fig. 1 is intuitive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose to combine causal reasoning techniques to assist imitation for autonomous driving. Specifically, the paper presents a novel approach, causality-inspired contrastive conditional imitation learning (3CIL), which integrates contrastive learning and action residual prediction. The framework is based on POMDP, trying to mimic the scenarios when the expert and imitator share different views." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My major concerns lie in theoretical clarity, experimental design, and practical applicability. In the absence of strong theoretical guarantees, the paper would benefit greatly from robust experimental results. Additionally, it’s crucial to provide a clear rationale for the integration of causality, contrastive learning, and conditional imitation learning, explaining why this combination is necessary. I have outlined specific questions below." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024cil,\ntitle={3{CIL}: Causality-Inspired Contrastive Conditional Imitation Learning for Autonomous Driving},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4QVgnxXVDB},\nnote={under review}\n}" }, "abstract": { "value": "Imitation learning (IL) aims to recover an expert's strategy by performing supervised learning on the demonstration datasets. Incorporating IL in safety-crucial tasks like autonomous driving is promising as it requires less interaction with the actual environment than reinforcement learning approaches. However, the robustness of IL methods is often questioned, as phenomena like causal confusion occur frequently and hinder it from practical use. In this paper, we conduct causal reasoning to investigate the crucial requirements for the ideal imitation generalization performance. With insights derived from modeled causalities, we propose causality-inspired contrastive conditional imitation learning (3CIL), a conditional imitation learning method equipped with contrastive learning and action residual prediction tasks, regularizing the imitator in causal and anti-causal directions. To mitigate the divergence with experts in unfamiliar scenarios, 3CIL introduces a sample-weighting term that transforms the prediction error into an emphasis on critical samples. Extensive experiments in the CARLA simulator show the proposed method significantly improves the driving capabilities of models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Imitation Learning", "Autonomous Driving", "Causal Reasoning", "Causal Confusion" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/56180408d355331d3fa6a181b1ef6d6746b86aea.pdf" }, "presentation": null, "primary_area": { "value": "applications to robotics, autonomy, planning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "3CIL: Causality-Inspired Contrastive Conditional Imitation Learning for Autonomous Driving" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper addresses an important research problem, KV cache optimization for LLM inference, and proposes two interesting techniques: center of gravity eviction (CGE) and remote gap localization (RGL).\n\n- Empirical results are competitive with prior work and baseline models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper develops a KV-cache compression technique called IntelLLM to optimize inference in LLM on long text tasks. IntelLLM consists of two cache eviction strategies: center of gravity eviction (CGE) and remote gap localization (RGL). CGE mitigates the domain semantic imbalance by redirecting attention away from the center of gravity attention (cluster of important KVs). RGL solves the issue of time span vanishing caused by cache compression by assigning cache position values to distant KVs. Empirical evaluation shows IntelLLM saves KV cache memory by 50% with similar performance on long text tasks compared to baseline." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Using sparsity in attention to compression KV cache is not new. Two ICLR 2024 papers: StreamingLLM (https://openreview.net/forum?id=NG7sS51zVF) and FastGen (https://openreview.net/forum?id=uNrFpDPMyo) both observe the attention patterns and use it to compress KV cache.\n\n- Missing important work in both related work and baseline comparison. The paper did compare with StreamingLLM, but does not discuss it in related work. In fact, the paper misses many important KV cache prior works:\n(1) Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs, ICLR 2024, https://openreview.net/forum?id=uNrFpDPMyo\n(2) SnapKV: LLM Knows What You are Looking for Before Generation, https://arxiv.org/abs/2404.14469\n(3) XC-Cache: Cross-Attending to Cached Context for Efficient LLM Inference, https://arxiv.org/abs/2404.15420\n(4) Layer-Condensed KV Cache for Efficient Inference of Large Language Models, https://arxiv.org/abs/2405.10637 \n(5) PyramidInfer: Pyramid KV Cache Compression for High-throughput LLM Inference, https://arxiv.org/abs/2405.12532\n(6) PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling, https://arxiv.org/abs/2406.02069 \n\n\n- Many writing sections are unclear. \n(1) The introduction (line 55) says prior work has two limitations, does IntelLLM have these two limitations too? What is the conflict between increased computational overhead and memory optimization? Line 83 says significant, how much memory is saved? Any speed gains? \n(2) The analysis in Section 3.1 is weak without any evidence or citations, the two conclusions are not convincing either. For example, line 172 says they compromise the robustness of the attention distribution. What is the robustness of attention in the first place? And why would the covariates compromise this? Line 180 says sliding window fails to reason effectively about long texts, any evidence or citation? Line 183 says they contribute to the collapse of the LLM, again, no evidence or justification. \n(3) The two theorems in sections 3.2 and 3.3 are not theorems, and the paper provides no proof.\n\n- Evaluation is weak and flawed. Table 1 (line 399) presents the results of IntelLLM on longbench. But unclear why the window size is 4K for IntelLLM and 8K for others. There is no side-by-side efficiency comparison either. Line 443 says the latency increased by 2.63% but the memory saves 50%, is it the case that IntelLLM always saves 50% memory? There is no ablation study on that." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "As weakness" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The combination of CGE and RGL provides a novel solution to the KV cache memory challenge, enhancing memory efficiency without fine-tuning or substantial performance loss. The paper presents a strong theoretical basis for its methods, including insights into the sparsity of attention weights and the impact of key tokens, strengthening the validity of the proposed strategies. IntelLLM is easy to integrate into existing LLM frameworks, as it requires minimal modifications, making it highly practical for real-world deployment, especially in resource-constrained environments. The extensive benchmarking on LongBench with models like Llama-3-8B-Instruct and Mistral-7B demonstrates IntelLLM's efficiency and adaptability across diverse tasks, validating the approach. Achieving 50% KV cache reduction with a negligible increase in latency is a noteworthy achievement, making IntelLLM suitable for long-text inference tasks in various settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces IntelLLM, a framework that aims to optimize the key-value (KV) cache compression for large language models (LLMs) without compromising performance. It addresses the challenge of high memory consumption during long-sequence inference by using two innovative techniques: Center of Gravity Eviction (CGE) and Remote Gap Localization (RGL). CGE prioritizes important tokens in attention mechanisms to ensure efficient memory use, while RGL preserves essential long-range dependencies using positional features. These strategies enable significant memory savings, reducing KV cache usage by 50%, with only a minimal impact on inference latency. The authors demonstrate IntelLLM's effectiveness through comprehensive experiments, achieving performance comparable to or better than full KV models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Sparse attention is already well explored in several previous works as [1] [2]. This will weaken the novelty of this work. H2O [3] has already well-explored the feedback of using sliding window.\n2. Lack of baselines (i.e., H2O [3], SnapKV [4], PyramidKV [5])\n3. Evaluation of Needle in a Haystack is required to help illustrate your motivation of maintaining long-range dependencies\n4. GCE is pretty close to previous methods like H2O [3] and SnapKV [4]. I can only see limited novelty over this method.\n\n[1] https://arxiv.org/abs/2402.17762\n[2] https://arxiv.org/pdf/2309.17453\n[3] https://arxiv.org/abs/2306.14048\n[4] https://arxiv.org/abs/2404.14469\n[5] https://arxiv.org/abs/2406.02069" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. What is the rationale behind choosing StreamingLLM and InfiniteLLM as KV cache compression baselines?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "**Significance:** The paper claims 50% KV cache compression without a significant drop in performance, outperforming other KV cache compression methods in a majority of datasets in LongBench and outperforming full KV cache in some datasets. The method does not require fine-tuning, making it easy to apply to existing LLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to further the Pareto frontier of KV-cache compression rate and performance. By employing strategic eviction strategies, the method leverages the observation that only a small subset of tokens in long texts capture the majority of attention weights." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There does not seem to be a rigorous proof or set of empirical observations to substantiate the theorems proposed in sections 3.2 and 3.3.\n2. There is no discussion on the method used to choose \"k\" – the number of top keys to be treated as \"centers of gravity\".\n3. Implementation details are not provided, making it hard to reproduce the results.\n4. There does not seem to be an ablation study to evaluate how the method performs when only either CGE or RGL is used.\n5. There is no discussion on specific deployment environments where a 50% memory saving will enable new use cases, or how this method can be combined with other methods to further increase memory savings.\n6. The experiments only provide two KV cache compression methods as baselines, leaving out other KV cache compression methods that do not require fine-tuning, such as static prefix caching, paged attention, or radix attention. Additionally, the experiments do not compare with other approaches that do not involve KV cache compression.\n7. Best performing methods are not clearly marked in the experiment result tables.\n8. The explanations and visualizations of CGE and RGL are unclear." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "There is no reference to table 2, why the experiments are different on Mistral and Llama?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The problem investigated in the paper, KV cache compression, is crucial for long-context generation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents IntelLLM, a method for compressing the key-value (KV) cache in Large Language Models (LLMs) to address memory constraints in long-sequence processing. Drawing on the sparsity of attention mechanisms, IntelLLM focuses on retaining only essential tokens, significantly reducing the KV cache size without compromising model performance. The proposed approach combines two strategies: center of gravity eviction (CGE), which prioritizes important tokens to preserve key semantic information, and remote gap localization (RGL), which maintains long-range dependencies using positional features. IntelLLM can integrate smoothly with existing LLMs, requiring minimal modifications and no fine-tuning. Experimental evaluations show IntelLLM achieves performance close to StreamLLM while halving KV cache requirements." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "a.\tThe paper lacks substantial references from the past year, missing important studies like SnapKV[1], FastGen[2], H2O[3], PyramidKV[4] and etc.\nb.\tThe observation that only a limited subset of tokens is critical for long-context generation has been extensively discussed in these and other recent works, which should be cited to provide a more comprehensive background.\nc.\tThe experimental baseline used in the study is relatively weak; including stronger baselines from the above-mentioned works would enhance the robustness of the comparative analysis and strengthen the validity of the results.\nd.\tIn Section 3, the two presented \"theorems\" are more accurately findings, as no formal proofs are provided to substantiate these claims.\n\n[1] SnapKV: LLM Knows What You are Looking for Before Generation\n[2] Model Tells You What to Discard: Adaptive KV Cache Compression for LLMs\n[3] H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models\n[4] PyramidKV: Dynamic KV Cache Compression based on Pyramidal Information Funneling" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Suggestion on improving the overall presentation. First, there are some notation used without definition, e.g. $n, m, l_{head}, l_{tail}$ in Algorithm 1, and the operation $[:]$ (which dimension is applied to?) used in Algorithm 1. Second, terms such as \"Center of Gravity\" should be defined formally in the KV cache eviction topic. Third, there remains some minor grammar and expression mistakes in this paper, such as the misuse of citation format. Forth, the presentation of the tables and figures should be improved. The average score should be reported in Table 1 & 2. The meaning of the x-axis and y-axis of Figure 1 & 2 should be clarified.\n\n2. The description of Algorithm 1 is confusing. Could you provide a very detailed explanation of the calculation process in natural language? \n\n3. There isn't any model named \"Llama3-7B\". Do you mean Llama-3-8B?\n\n4. What are the detailed settings of the baselines, e.g. how much KV cache is retained? Could you provide more information on the hyperparameters used on baselines?\n\n5. The experimental results mentioned in line 316-319 should be included in this paper if such conclusion is drawn. If there is no space, it should be placed in the Appendix. The experiments in line 359-361 should also be included.\n\nOverall, I really like the idea proposed by this paper, which I personally find very inspiring. I will raise my rating on soundness and overall rating if good discussion is made with necessary experimental results during the rebuttal phase." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper provides a clear explanation of why the attention mechanism tends to allocate excessive weights to a few tokens with high attention scores while neglecting others. The analysis of the softmax function is insightful and well-reasoned.\n\n2. The proposed method, IntelLLM, demonstrates strong performance in benchmarks and shows compatibility across multiple models, including widely-used ones like Llama and Mixtral.\n\n3. The method is simple and easy to implement. IntelLLM can be integrated into various scenarios and inference frameworks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents IntelLLM, a novel KV cache eviction algorithm designed to alleviate the storage and computational burden of Transformer-based large language model inference. IntelLLM leverages the sparsity of attention mechanisms and strategically evicts certain tokens to improve inference efficiency. The authors propose two key strategies: Center of Gravity Eviction (CGE) and Remote Gap Localization (LGL). CGE addresses the semantic loss caused by dominant attention scores in softmax. LGL reorganizes token positions by creating a large gap between global and local tokens, further enhancing processing efficiency. IntelLLM is evaluated on LongBench with two models, Llama-3-8B and Mixtral-7B-instruct-v0.2, and outperformed all the baselines, including Full KV cache, StreamingLLM, and LM-Infinite. The algorithm achieved 2x KV cache compression ratio for Llama and an 8x compression ratio for Mixtral, maintaining strong inference efficiency despite the reduced cache size. Additionally, the paper includes ablation studies focusing on Head Gravity and the RGL Gap, demonstrating the soundness and effectiveness of IntelLLM's design." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The presentation is to some extent unclear and confusing. Some notations and terminologies are used without a clear definition. The description of Algorithm 1 is confusing and hard to understand. Tables and figures and be organized in a better format. Details of presentation issues are described in question Q1.\n\n2. Insufficient baseline selection. The proposed algorithm of this paper is an eviction algorithm that compresses KV cache, thus there should be comparison between other popular eviction methods, including H2O [1], SnapKV [2], SirLLM [3] and IntervalLLM (a baseline created in SirLLM). It would be better to compare IntelLLM with other end-to-end methods, including InfLLM [4] and MInference [5], as they utilize attention sparsity. (This is only a suggestion and the authors are not required to test all the methods above during the review process, but futher discussion is expected.)\n\n3. Insufficient benchmark datasets. Although LongBench is a classic benchmark for long context inference, the average length is relatively short and the hardness is limited. It is expected to benchmark IntelLLM on harder benchmark, e.g. L-Eval [6], and longer benchmark, e.g. RULER [7] or InfiniteBench [8]. It would also be better to test IntelLLM on accurate context retrieval tasks, e.g. Ret.PassKey in InfiniteBench and RULER. (This is also a suggestion, and the authors are not required to test all of the benchmarks mentioned above. However, some simple supplementary benchmarks are welcomed.)\n\n4. Lack of ablation studies. Ablations on $L_{comp}$ and $L_{near}$ are expected, as they can be used to prove which part of the retained KV cache is more important and to what extent can the hyperparameters affect the overall performance. Also, there should be comparison between RGL and other methods of assigning position information, such as assigning continuous position ids and not assigning any position information for distant tokens.\n\n5. Unclear presentation of research intention. KV cache compression is a technique developed to enhance model generation speed or reduce memory consumption. The paper should clearly state out which purposed is focused mainly, and conduct the corresponding experiments. For example, the model generation speed should be tested and it is expected to be faster than full KV cache inference. The peak memory is also expected to be lower if the system implementation is delicate enough. The speed reported in line 445-446 cannot state the research intention as the pre-fill speed is slightly lower. Reporting a faster generation speed (the speed only tested on the decoding stage and excluding the pre-fill stage) might be helpful.\n\n6. No limitation and future work discussion. This paper should include some vital limitations and potential future work of the proposed methods.\n\n\n[1] H2O: Heavy-Hitter Oracle for Efficient Generative Inference of Large Language Models\n\n[2] SnapKV: LLM knows what you are looking for before generation\n\n[3] SirLLM: Streaming Infinite Retentive LLM\n\n[4] InfLLM: Unveiling the intrinsic capacity of LLMs for understanding extremely long sequences with training-free memory\n\n[5] Minference 1.0: Accelerating pre-filling for long-context llms via dynamic sparse attention\n\n[6] L-eval: Instituting standardized evaluation for long context language models\n\n[7] RULER: What’s the Real Context Size of Your Long-Context Language Models?\n\n[8] ∞ Bench: Extending Long Context Evaluation Beyond 100K Tokens" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024intelllm,\ntitle={Intel{LLM}: Little Hints Make a Big Difference for {LLM} {KV} Cache Compression},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4QWPCTLq20},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) have demonstrated exceptional capabilities in integrating contextual knowledge, but their deployment is often constrained by the substantial computational resources required for long text sequences. To mitigate the inference time cost associated with attention mechanisms, LLMs utilize key-value embedding caching techniques (KV cache), which introduce significant storage pressure. In this paper, we propose IntelLLM, a novel and efficient approach to KV cache compression that strikes a balance between compression rate and performance. Drawing inspiration from sparse attention mechanism, we observe that only a small subset of tokens in lengthy texts capture the majority of attention weights. This sparsity, intrinsic to the attention mechanism, serves as the foundation for improving the KV compression ratio through a strategic eviction method. IntelLLM is composed of center of gravity eviction (CGE) strategy and remote gap localization (RGL) strategy. CGE is designed to address the potential loss of important semantic dependencies when evicting high-sparsity tokens, which prioritizes the retention of key tokens by shielding the center of gravity of attention during inference, thereby preserving critical information and optimizing the efficiency of attention computation. Additionally, RGL is proposed to leverage implicit positional features to maintain long-range dependencies, inspired by advancements in location encoding research. Our KV compression approach integrates seamlessly with existing LLMs, requiring minimal code modifications without the need for fine-tuning or model parameter changes. IntelLLM not only significantly reduces the storage requirements for KV cache but also consistently outperforms full KV models in long text processing tasks, while utilizing only 50% of the typical KV cache expenses." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "LLM", "KV cache compression", "CGE", "RGL" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/e70bdc24592a789f5a144d6ea6d61377169446db.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "IntelLLM: Little Hints Make a Big Difference for LLM KV Cache Compression" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. Are there any issues with the citation format in the paper?\n2. Does the paper lack an appendix?\n3. In Table 2, there is no comparison of results with environment tools but without evolving. This should be added." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper proposes a multi-agent collaboration approach to address software development problems.\n2. This paper introduces a software development benchmark called RSD-Bench, which provides more detailed and structured software requirements for documenting user needs compared to previous benchmarks.\n3. Extensive experiments demonstrate that EvoMAC outperforms other single-agent and multi-agent methods on both RSD-Bench and HumanEval." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a multi-agent collaboration approach to address software development problems. EvoMAC obtains text-based environmental feedback by verifying the match between the MAC network's output and the target proxy, and it updates the network using a novel textual backpropagation technique, thereby achieving the final development outcome. Additionally, this paper introduces a software development benchmark called RSD-Bench, which provides more detailed and structured software requirements for documenting user needs compared to previous benchmarks. The final experimental results show that the proposed EvoMAC outperforms other single-agent and multi-agent methods on both RSD-Bench and HumanEval." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The benchmark proposed in this paper lacks data analysis and some basic statistical information, such as prompt length, the number of final generated files/functions, etc.\n2. The benchmark proposed in this paper is relatively easy, with the EvoMAC method already achieving around 90% accuracy." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How sensitive is EvoMAC to the quality and specificity of feedback from unit tests? If the unit tests are incomplete or overly general, would EvoMAC still produce reliable code, or would it require stricter validation criteria?\n2. Can EvoMAC work effectively with models of different sizes, or does it rely on the power of high-capacity LLMs? Would it perform satisfactorily with smaller models that might be more efficient in constrained environments?\n3. Can the self-evolving mechanism be applied to other domains outside software development? If yes, how?\n4. Given EvoMAC’s iterative approach, how would it handle larger software projects with thousands of lines of code and extensive requirements? Are there specific design considerations for scaling it to more extensive projects?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This self-evolving paradigm allows MAC networks to adapt iteratively based on environmental feedback. The framework employs a mechanism similar to neural network backpropagation, where the output of the MAC network is verified against a target proxy, facilitating continuous learning and improvement.\n- The RSD-Bench provides a structured benchmark for software-level coding tasks, focusing on comprehensive requirements rather than isolated functions.\n- By incorporating unit tests and compilers as feedback mechanisms, EvoMAC reduces subjectivity and provides reliable feedback, which is critical for verifying the correctness of generated code. The objective environment-based feedback is an effective alternative to critique agents, which can introduce bias and hallucinations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel framework called EvoMAC, aimed at enhancing the capabilities of LLM-driven multi-agent collaboration (MAC) systems in software development. The authors argue that traditional MAC systems are heavily reliant on human-designed workflows, which limits their adaptability and performance in real-world scenarios. EvoMAC seeks to overcome these limitations by enabling self-evolution of agents and their connections during task execution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The EvoMAC framework introduces significant complexity with its multi-agent setup, dynamic adjustments, and textual backpropagation mechanism. This complexity may limit the framework's accessibility and implementation ease for real-world adoption outside of specialized research contexts.\n- Although EvoMAC performs well with large models like GPT-4o-Mini, its performance with smaller or less capable models is unclear. This reliance may restrict its applicability, particularly in environments with limited computational resources.\n- RSD-Bench focuses on website and game software types, which may not comprehensively represent the diversity of real-world software development tasks. Expanding the evaluation to include other domains, such as enterprise applications or data processing software, would enhance the generalizability of the results." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Please refer to the questions in Weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- Effectively demonstrates EvoMAC's evolution process as analogous to neural network training, establishing a reliable target proxy for evaluation and constructing a clear objective.\n\n- The paper is well-structured and easy to follow, with thorough explanations of EvoMAC’s self-evolving process,the design of the RSD-Bench, and detailed descriptions of experimental procedures. Figures and benchmarks illustrate the methodology effectively, aiding comprehension.\n\n- By addressing limitations in traditional MAC systems and demonstrating EvoMAC’s efficacy on challenging software-level tasks, this work sets a promising precedent for adaptive agent frameworks in automated software development, making it valuable for both research and practical applications.\n\n- The RSD-Bench is more practical in software generation evalaution, as it aligns closely with real-world software development process. By incorporating unit tests at both the task and function levels, it establishes a rigorous and precise mechanism for evaluating software generation quality. Additionally, RSD-Bench demonstrates strong human alignment (0.9922), providing a reliable evaluation metric. This paper also conducts analysis the reasonality of RSD-Bench in comparison to existing benchmarks, demonstrates its necessity. \n\n- This paper provides thorough experiments and analyses that robustly demonstrate EvoMAC’s effectiveness and performance. EvoMAC’s strong performance on both the RSD-Bench and HumanEval benchmarks highlights its high quality and efficacy in handling complex coding tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents EvoMAC, a self-evolving multi-agent collaboration (MAC) network designed to advance LLM-based multi-agent systems beyond function-level tasks to software-level coding tasks. EvoMAC employs a unique textual backpropagation mechanism to iteratively update agents and their connections in response to text-based environmental feedback, effectively enhancing task completion without human intervention. By formulating the evolving process to analogize neural network training, EvoMAC provides a clear structure for defining and extracting improvements. This approach underscores the significance of iterative refinement in the software generation process, enabling continuous improvement and adaptability in complex coding tasks. \n\nTo evaluate EvoMAC, the authors introduce RSD-Bench, a novel benchmark with complex and diverse software requirements that includes automated requirement verification. EvoMAC demonstrates superior performance on RSD-Bench and the HumanEval function-level benchmark, outperforming state-of-the-art methods and showcasing its robustness and adaptability across various evolving iterations and LLM configurations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Existing studies, such as [1-4] have also explored automation in LLM-based multi-agent collaboration. Please compare the differences between EvoMAC and these works. \n\n- EvoMAC's updating process includes removing agents that have completed their tasks. Can the entire agentic workflow be replayed once a task is finished, or is the removed agent permanently excluded from further iterations? \n\n- Given that EvoMAC includes multiple evolutionary iterations, direct comparisons with standard multi-agent frameworks may not be entirely fair. Could you also provide the number of LLM calls for tasks in RSD-Bench? This metric would offer a more clearer understanding of EvoMAC’s performance.\n\n- EvoMAC primarily focuses on models like GPT-4, Claude 3.5, and Gemini, but it is unclear if the framework can adapt to less powerful models, such as GPT-3.5 or open-source options like DeepSeek. Presenting results across a broader range of LLMs would support EvoMAC’s claims of robustness and adaptability.\n\n- Can the authors provide additional examples of the unit tests designed within RSD-Bench?\n\n- The Testing Team’s performance significantly impacts the Coding Team's potential, particularly in the HumanEval benchmark. How is the Testing Team’s performance evaluated to ensure alignment with target performance objectives and to prevent divergence? Additionally, how is the Testing Team’s performance quantified within RSD-Bench? \n\n- The paper does not specify the stopping criteria for EvoMAC’s iterative evolution process. Could the authors provide details on the stopping mechanism or criteria? \n\n- The paper lacks specific settings for the Coding Team in both the HumanEval and RSD-Bench benchmarks. Please provide these details to improve clarity on the experimental configuration and consistency across benchmarks.\n\n- Could the authors showcase additional examples of the textual gradient analysis and the updating process during the evolution for HumanEval and RSD-Bench?\n\n\n\n[1] Liu Z, Zhang Y, Li P, et al. Dynamic llm-agent network: An llm-agent collaboration framework with agent team optimization[J]. arXiv preprint arXiv:2310.02170, 2023.\n\n[2] Zhuge M, Wang W, Kirsch L, et al. GPTSwarm: Language Agents as Optimizable Graphs[C]//Forty-first International Conference on Machine Learning.\n\n[3] Hu S, Lu C, Clune J. Automated design of agentic systems[J]. arXiv preprint arXiv:2408.08435, 2024.\n\n[4] Qian C, Xie Z, Wang Y, et al. Scaling Large-Language-Model-based Multi-Agent Collaboration[J]. arXiv preprint arXiv:2406.07155, 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- When explaining their approach (section 3.1), the authors did not mention problems regarding the context window of most LLMs. Since requirements can contain quite a large amount of textual data, is the approach capable of dealing with it without extra techniques (e.g., RAG)? If not, even though the authors did not highlight it as a problem, the context window limitations should be mentioned." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is very well written and clearly explained. The two proposed contributions (the EVoMAC approach and RSD-Bench dataset) are consistent with the objective expressed in the introduction.\nAltough the problem of how to (self-)organize multi-agents is not exactly a new problem, the rise of agentic approaches using LLMs brought new traction to this challenge, and the authors addressed a pain point on these approaches when dealing with complex software development. Indeed, most solutions rely only on function-level and ignore the requirements engineering perspective, which ultimately leads the developers to half-baked solutions. This is not the case with the proposed EvoMAC. It takes account of the particularities of user requirements when organizing the agents initially and also considers the evaluation of the generated code against initial requirements.\nTheir inspiration for neural network algorithms is broad. However, it is also a clever idea that sounds original and demonstrates a creative adaptation of backpropagation principles to multi-agent collaboration.\nThe authors also provide sound experimentation on how they implement their approach and when comparing with similar solutions.\nTogether with the approach, the authors provided a well-defined benchmark to overcome the limitations of the existing ones. A more detailed description of the RSD-bench could indeed compose a contribution per se.\nThe RSD bench is tailored to requirements engineering, which could address common limitations in agent-based collaboration research by bridging functionality-based assessments with requirement-based evaluations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors provide a paper with twofold contribution: i) a new approach to developing multi-agent collaboration systems for software development and ii) a benchmark to compare their approach with existing approaches to solving the same problem.\nRegarding (i), their approach (called EvoMAC) intends to overcome the limitations of similar multi-agent collaboration systems development workflows, mainly on adaptability and generalization. EvoMAC was designed to mimic standard neural network development, meaning that the errors are \"backpropagated\" throughout the agents, creating a self-adaptive multi-agent network.\nRegarding (ii), the benchmark dataset RSD-bench was created based on the requirements of the software being developed, in contrast to the existing ones, which are usually based on the functionality of the generated code/software (i.e., unit tests).\nThe paper's results show that the EvoMAC approach outperforms other approaches when applying the RSD-Bench to adapted versions of standard benchmark datasets like HumanEval." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Most of the paper's weaknesses are minor problems that can improve its quality, even though I don't consider them mandatory.\n- I truly believe that the mathematical explanation of the problem (mainly between lines 169-184) is unnecessary, even though I see that this somehow facilitates some later explanations. Altough it provides some kind of generalization, the approach still relies on LLMs that are probabilistic by nature and then are not mathematically generalizable (i.e., the function $\\phi(\\cdot,\\cdot)$ is essentially a message sent to an LLM, so it does not behave exactly as a mathematical generic function as the authors may want to demonstrate). I advise the removal of the mathematical explanation. It can help sometimes but does not add value to the paper.\n- The paper lacks an actual example/use case in the opposite (or complementary, if the authors decide to keep it) of the mathematical explanation. The authors use this type of explanation in lines 270-271. This kind of example can be used as a running example throughout the paper.\n- Minor problems:\n\t- Typo on the Figure 3 caption. \"indrection\"\n\t- Figures 6 and 8 can be improved to ensure legibility and accessibility, maybe by adjusting font size and contrast.\n\t- Figure 1 is a bit confusing. I think it can be split into 3 different figures or be better explained in the paper. If kept as is, I suggest adding brief explanations of the arrows, such as \"add,\" \"revise,\" and \"remove.\"\n\t- Figure 5 can be rethought with a better color choice, especially considering the accessibility of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024selfevolving,\ntitle={Self-Evolving Multi-Agent Networks for Software Development},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4R71pdPBZp},\nnote={under review}\n}" }, "abstract": { "value": "LLM-driven multi-agent collaboration (MAC) systems have demonstrated impressive capabilities in automatic software development at the function level. However, their heavy reliance on human design limits their adaptability to the diverse demands of real-world software development.\nTo address this limitation, we introduce EvoMAC, a novel self-evolving paradigm for MAC networks. Inspired by traditional neural network training, EvoMAC obtains text-based environmental feedback by verifying the MAC network's output against a target proxy and leverages a novel textual backpropagation to update the network.\nTo extend coding capabilities beyond function-level tasks to more challenging software-level development, we further propose RSD-Bench, a requirement-oriented software development benchmark, which features complex and diverse software requirements along with automatic evaluation of requirement correctness.\nOur experiments show that:\ni) The automatic requirement-aware evaluation in RSD-Bench closely aligns with human evaluations, validating its reliability as a software-level coding benchmark.\nii) EvoMAC outperforms previous SOTA methods on both the software-level RSD-Bench and the function-level HumanEval benchmarks, reflecting its superior coding capabilities." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Software development", "LLM", "Multi-agent collaboration" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/ad7adb5f92daccd3d18429f235d6e265635b7dd8.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/bccf047de25755d94d059c3f5f1dcc5000452fe4.pdf" }, "title": { "value": "Self-Evolving Multi-Agent Networks for Software Development" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- For KV-caches, the total KV caches are a little increased by the amount of one layer for cross-attention in upper layers, rights?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- StagFormer architecture is interesting, and has very good potential for both performance and throughput.\n- The idea of parameter sharing and recurrent decoding looks good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a new architecture StagFormer, which stagger the time dependency between the lower and upper layers. The overall design seems a little non-intuitive, but has a lot of potential for throughput and performance. For example, parameter sharing or local cross-attention could yield better throughput." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I like the concept and potential of this paper, but I believe that this paper is not well-organized, and looks like unfinished work yet. For example, there is missing reference in L.267 (I guess this refers to Table 3), there are a few results for proof of concept.\n- Table 3 is showing few-shot results for gray, blue, red lines in Figure 4 (correct me if I’m wrong.) I wonder why shared-weights StagFormer (blue) outperforms Baseline 2.8B params (red) in some benchmarks, even though it shows higher loss values.\n- What makes StagFormer 2.9B to outperform Baseline 2.8B params in Table 1? Is it due to cross-attention in upper layers? This looks somewhat interesting and also confusing because I thought the changed structure (using previous timestep’s intermediate activations) could degrade performance a lot.\n- How did the authors measure the decoding time in Table 2? Running separate parameters in parallel is not trivial, I believe. Is it actual time or hypothetical time by assuming parallel execution of them?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Can you describe the architecture shape (vocab size, qkv heads, embedding dimensions) and its justification? The vocab size of 256K is quite high for models of this size.\n1. In Lines ~499-501, the authors mention that cross-attention is linear to input length instead of quadratic with window size 1. Isn't it linear with any fixed window size? Considering that the cost of attention mainly stems from KV cache IO during decoding, I think the constant factor with a window size as small as 128 makes the cost of cross-attention negligible compared to self-attention (especially when expanding to modern context lengths of 8K or more).\n 1. However, the *increase* in performance when going from full cross-attention (1024) to windowed attention with window size 512 and 128 is strange. Can the authors justify this increase in performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The idea and architecture design are very novel\n1. The authors propose numerous variants which showcase the potential extension of the idea across various axes–parallel execution, weight sharing, recurrent computation.\n1. The architecture shows clear advantages over vanilla transformers across its variants\n1. The writing is easy to follow and visual depiction of the architecture and its variants are superb." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present the Staggered Transformer (StagFormer) and its variants which relieve sequential dependancies in the decoding pipeline to enable higher levels of parallel execution.\n\nConsider a transformer with two stacks of layers, A (bottom half) and B (upper half). In vanilla transformers, the input token embedding is passed to stack A. Then, the output of stack A is passed to stack B. All layers apply self-attention on outputs of the previous layer.\n\nIn the baseline StagFormer (`Separate-Weights`), stack A is the same. However, stack B takes in the input token embedding rather than the output of stack A.\nTo supplement this, stack B applies cross-attention on the final outputs of stack A, up until the previous token. In other words, stack B cross-attends to the outputs of *all previous input tokens* from stack A, instead of directly inputting that of the *current* input token. This relieves the dependency of stack B on stack A, within a single decoding step, thus both A and B can be computed simultaneously.\n\nThe authors investigate many variants of this design:\n1. `Shared-Weights`: this is where stack A and stack B share the same model parameters (excluding the cross-attention layers which are unique to stack B).\n2. `Recurrent, Shared-Weights`: this is a unique decoding method for the `Shared-Weights` trained model. In `Shared-Weights` stack A and B are identical, except that stack B applies cross-attention to outputs from stack A. Essentially, the shared stack S (= A = B) is first forwarded without cross-attention, and then forwarded a second time *with* cross-attention, attending to outputs from the first forward pass. The `Recurrent` setting refers to that where the first forward pass is skipped, and cross-attention in the second pass attends to outputs of the \"second\" pass from the previous decoding step.\n3. `p > 2`: this is where more than two stacks are considered.\n\nWhen compared to vanilla transformers pretrained from scratch, StagFormers show various advantages, mainly:\n- `Shared-Weights 2x18L`: StagFormer outperforms the vanilla 18L baseline (with roughly same parameters) in both perplexity and average task performance. Using recurrent decoding (roughly matching 18L baseline computation), average task performance lies between the two. StagFormer underperforms the vanilla 36L baseline with roughly same computation in perplexity, but performs comparably on tasks.\n- `Separate-Weights 2x18L`: StagFormer outperforms the vanilla 36L baseline (with roughly same parameters and compute) in both perplexity and task performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Memory bottlenecks during decoding may hinder benefits of parallel execution, which is not discussed**: LM decoding is typically bottlenecked by memory rather than compute (see references below). When batch size x context length is small, memory access is dominated by model parameter access. Otherwise, memory access is dominated by KV cache access. While StagFormer can conceptually *parallelize* execution of layers, the associated memory access load cannot be parallelized. In fact, the cross-attention layer will add additional KV cache access overhead. These are critical to assessing the actual wallclock time benefits of decoding with StagFormers compared to vanilla transformers, but is not discussed.\n 1. Different variants of StagFormers will have different memory bottlenecks. Examples:\n 1. All variants: cross-attention is added in half of layers. Therefore, the overall KV cache access overhead will increase by 50% (relative to that of self-attention, used in all layers). This will have a larger effect on decoding time as batch size x sequence length becomes large.\n 1. `Separate-Weights`: both stacks can be executed in parallel, but the memory load is identical as the parameters of both stacks must be retrieved from memory. This means that wall-clock time should typically be identical to vanilla transformers, as decoding is bottlenecked by memory access. `Shared-Weights` can solve this issue.\n 1. **It is unclear which StagFormer variant is used in Table 2, raising questions on the performance vs latency comparison**: While Table 2 states that a \"comparable quality StagFormer\" is 33% faster than baseline transformer during decoding, the exact variant is unclear. Given the reasons above, it seems likely that this is the `Shared-Weights 2x18L` variant. While its average task performance is comparable to baseline 36L, its PPL is in the middle of that between vanilla 18L and 36L. It would be misleading to describe this variant as \"comparable quality\" to vanilla 36L.\n 1. **Missing comparison of performance vs latency across model variants**: Expanding on the point above, a comparison of prefill/decode time across model variants will provide a clear picture on the performance vs latency benefits of each model variant. This could take the form of a single table that lists the PPL, task performance, and prefill/decode time for each model. In the case of `p > 2, Shared-Weight` variants, I believe this may actually reveal some advantages in terms of latency.\n 1. **The additional KV cache overhead of cross attention may slow down decoding for longer contexts**: Since KV cache overhead is quadratic to context length, the decode time advantages as shown in Table 2 will likely diminish with longer contexts, especially in batch decoding. Given the relatively short context length of 1024 tokens considered in this study, compared to modern LLMs with 8K+ context, measurement on longer contexts and larger batch sizes can help gauge the potential of the architecture.\n1. **Misleading task performance of `Recurrent` variant**: In Table 3 (for example), the performance of various tasks are identical between the `Shared-Weights 18L` model and its `Recurrent` counterpart. This is likely because the tasks are measured in a teacher-forcing setting, where the outputs of the prefill stage are used for evaluation. This does not represent the task performance of the `Recurrent` setting, as recurrence is only applied to decoding, as explained in Section 3.2.\n1. **Experimental results on model variants are hard to follow**: The organization of the results section could be improved to make the comparison between different model variants more clear.\n 1. Within tables, variations could be better indicated with separate columns, task names could be shortened for space, latency metrics could be included, etc.\n 1. Results on different variants are presented in multiple tables without a clear organization.\n1. **Incomplete writing**: \"(TODO)\" in Line 385, the reference error \"??\" in Line 267, and numerous typos suggest that this is an incomplete manuscript that is not ready for review.\n\nReferences on memory bottlenecks during inference\n- [Efficiently Scaling Transformer Inference](https://arxiv.org/abs/2211.05102)\n- [LLM Inference Unveiled: Survey and Roofline Model Insights](https://arxiv.org/abs/2402.16363v4)\n- [Taming Throughput-Latency Tradeoff in LLM Inference with Sarathi-Serve](https://arxiv.org/abs/2403.02310)\n- [Block Transformer: Global-to-Local Language Modeling for Fast Inference](https://arxiv.org/abs/2406.02657)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How does varying the depth of individual stacks in StagFormer affect the trade-off between decoding speed and model quality?\n2. What factors determine the optimal number of stacks for a given application, balancing computational efficiency and performance?\n3. Could the staggering concept be extended to encoder-decoder Transformers, like those used in machine translation?\n4. How well could StagFormer be combined with other techniques, like quantization or knowledge distillation, to further enhance decoding efficiency?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. StagFormer introduces a unique method to break the sequential dependency of layers in Transformers, enabling parallel execution.\n2. Experiments demonstrate significant latency reduction while maintaining or even exceeding the quality of a standard Transformer.\n3. The paper investigates different StagFormer variants, offering flexibility and adaptability to various scenarios and resource constraints.\n4. The paper effectively explains the StagFormer concept and its variants, supported by clear diagrams and algorithms." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel Transformer architecture called StagFormer designed to improve the efficiency of decoding in Transformer-based language models by enabling the parallel execution of layers along the depth axis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited exploration of p > 2. While the paper explores StagFormer with more than two stacks, it acknowledges performance degradation and the need for further research in this area.\n2. The paper mentions the communication cost associated with parallel execution but doesn't offer concrete solutions to mitigate it.\n3. While the Pile dataset is comprehensive, evaluating on additional datasets would strengthen the generalizability of the findings.\n4. Comparing StagFormer with other methods for efficient Transformer inference, such as speculative decoding, would provide a more comprehensive perspective." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Addressing the weaknesses outlined above would improve the paper." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- staggered computation leads to significant improvements in per-time-step decoding speeds while slightly improving performance\n- provides results and analysis of different variants of staggered transformers that further explores the architecture's efficacy" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel transformer architecture that effectively reduces the number of sequential steps (layers) during the decoding process by staggering the computation across different time-steps. This allows for improved parallelism during decoding individual sequences, providing speedups during inference." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Biggest critique is that it lacks comparative analysis of staggering computation vs. simply increasing the width of the model and lowering the number of layers, as this increases per layer parallelism while decreasing the number of layers leading to a similar improvement in decoding speed.\n- This technique is possibly only useful for speeding up decoding when only a single sequence is being decoded. A non-staggered model could in theory process twice the batch size as it has half the parallelism (and hence half the per layer memory requirement) of a model staggered with p=2. \n- StagFormer is possibly slower to train (as inferred by its slower pre-filling speed)\n- Paper could be further refined (minor critique): \n - Some references are configured incorrectly (Table ?? in page 5, \"TODO\" in page 8)\n - Plots have unnecessary information (Figure 4 doesn't need texts like /1/summarize/train)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024stagformer,\ntitle={StagFormer: A Staggered Transformer for Decoding Layers in Parallel},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4RHdGVimNA},\nnote={under review}\n}" }, "abstract": { "value": "Standard decoding in a Transformer based language model is inherently sequential as we wait for a token’s embedding to pass through all the layers in the network before starting the generation of the next token. In this work, we propose anew architecture StagFormer (Staggered Transformer), which staggered execution along the time axis and thereby enables parallelizing the decoding process along the depth of the model. We achieve this by breaking the dependency of the token representation at time step $i$ in layer $l$ upon the representations of tokens until time step $i$ from layer $l−1$. Instead, we stagger the execution and only allow a dependency on token representations until time step $i−1$. The later sections of the Transformer still get access to the ”rich” representations from the prior section but only from those token positions which are one time step behind. StagFormer allows for different sections of the model to be executed in parallel yielding up to 33% speedup in decoding while being quality neutral. We also explore many natural variants of this idea. We present how weight-sharing across the different sections being staggered can be more practical in settings with limited memory. We show how one can approximate a recurrent model during inference using such weight-sharing. We explore the efficacy of using a bounded window attention to pass information from one section to another which helps drive further latency gains for some applications. We also explore demonstrate the scalability of the staggering idea over more than 2 sections of the Transformer." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "decoder only language models", "transformers", "staggered execution", "pipelining", "parallel decoding", "efficiency" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/60ecfc1e9a442a4b3208ef478624fffea1214859.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "StagFormer: A Staggered Transformer for Decoding Layers in Parallel" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weaknesses" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The work handles an important ML problem; Active Learning (AL) with differentiable strategy search.\n- The proposed method is technically sound\n- Writing is clear and easy-to-follow" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces AutoAL, a differentiable active learning (AL) strategy search method that builds on existing AL sampling strategies. AutoAL contains two neural networks, SearchNet and FitNet, which are co-optimized through a differentiable bi-level optimization framework to identify optimal AL strategies for different tasks. Experimental results show that AutoAL outperforms individual AL algorithms and other selective approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Hybrid AL methods that combine uncertainty and diversity have been demonstrated to perform effectively in a variety of situations. It would be beneficial to include examples where the proposed AutoAL approach is particularly necessary or advantageous for specific applications.\n- The bi-level optimization within AutoAL relies on labeled data. How does the algorithm perform if the labeled data is skewed or imbalanced? For instance, if the initial labeled set suffers from class imbalance, might this severely impair the algorithm? The assumption of a randomly selected initial set, as used in the current experiments, appears to be less practical.\n- Similarly, is there a guarantee that the AutoAL approach, trained with labeled data from the current AL round, will identify the most informative samples from the unlabeled pool in the subsequent AL round? A more detailed analysis of the algorithm's guarantees is necessary.\n- The approach of training an additional network for sample selection shows similarities to [1] employing meta-learning with an additional network for querying.\n- Can the proposed method be applied to the open-set AL problem [1]?\n- The datasets used in the experiments are of small scale. It is imperative to validate the performance on large-scale datasets, such as ImageNet.\n\n---\n[1] Meta-Query-Net: Resolving Purity-Informativeness Dilemma in Open-set Active Learning, NeurIPS, 2022" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- How does AutoAL perform in terms of computational efficiency compared to traditional methods on large-scale datasets?\n- What mechanisms ensure robustness against convergence issues and local minima in the bi-level optimization?\n- Can AutoAL be extended to generate new active learning strategies dynamically rather than relying solely on a predefined candidate pool?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- A new approach that automates active learning strategy selection through differentiable optimization, surpassing manual and non-differentiable methods.\n- Effective integration of strategy selection and data modeling via the bi-level optimization of SearchNet and FitNet.\n- Flexibility and adaptability, allowing incorporation of multiple existing strategies and tailoring to specific tasks and data distributions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents AutoAL, a framework for automated active learning that optimizes query strategy selection using differentiable methods. Traditional active learning approaches often rely on predefined strategies like uncertainty sampling or diversity sampling, which may not perform optimally across different datasets or tasks. AutoAL addresses this limitation by integrating existing active learning strategies into a unified framework. It employs two neural networks, SearchNet and FitNet, within a bi-level optimization structure to automate the selection process. By relaxing the discrete search space of active learning strategies into a continuous domain, AutoAL enables gradient-based optimization, enhancing computational efficiency and adaptability. Experimental results demonstrate that AutoAL consistently outperforms individual strategies and other selective methods across various natural and medical image datasets, highlighting its effectiveness and versatility." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Increased complexity and computational overhead due to the additional neural networks and bi-level optimization, potentially challenging scalability on large datasets.\n- Dependence on a predefined pool of candidate strategies, which may limit performance if optimal strategies are not included.\n- Lack of in-depth theoretical analysis explaining the method's effectiveness and the conditions under which it performs best, possibly affecting generalizability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 1 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "see above" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The studied strategy selection problem for active learning is important.\n2. The bi-level optimization strategy is rational." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an active strategy query algorithm where the optimal query strategy is selected by a bi-level optimization network. In particular, the authors aggregates the query strategies by a scoring function implemented as a Gaussian Mixture Model. Then, the authors split out a validation set from the labeled samples to guide scoring function calculation. Experimental results show that the proposed method supasses the baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is still room for improvement in the paper writing.\n\n 1.1. It's unnecessary to name the two networks \"fitnet\" and \"searchnet,\" as it seems intended to make people think this is a significant innovation. However, in meta-learning, this kind of separated network design and bi-level optimization paradigm is very common.\n\n 2.1. The notations are somewhat confusing. For example, the authors didn't clearly define the output of the search net in Section 3.2, making it hard to understand Sec 3.2. It wasn’t until I finished reading the method section that I realized the output is actually a sample-wise score, forming an aggregation of scores for different queries.\n\n2. The novelty of this paper is relatively limited. The proposed meta-learning/bi-level optimization has been applied to AL [1,2]. Also, I think the algorithm design is too complicated.\n\n3. The motivation for modeling the scores by GMM distributions is unclear. Why is the score function of each strategy distributed as a Gaussian Distribution? Why is the final score function a linear weighted aggregation of different strategies? The authors should provide a concrete application or example.\n\n4. The comparison methods are too outdated, with the latest ones being LPL and BADGE from 2019. Additionally, the datasets are quite limited; despite the complexity of the method design, only CIFAR and MNIST datasets were used. Validation should be conducted on the ImageNet dataset (at least Image100). Otherwise, given that the algorithm design is much more complex than the baselines, its effectiveness cannot be convincingly demonstrated.\n\n[1] Kunkun Pang, Mingzhi Dong, Yang Wu, and Timothy Hospedales. 2018. Meta-learning transferable active learning policies by deep reinforcement learning. International Workshop on Automatic Machine Learning (ICML AutoML 2018).\n\n[2] https://grlplus.github.io/papers/96.pdf" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "1. My only concern is the efficiency of the AutoAL algorithm. Although more efficient solutions have been proposed to solve second-order optimization problems, I cannot find any relevant experiments to verify them." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the Weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The problem studied in this paper is valuable. This paper presents the first differentiable AL strategy search method. \n2. The proposed AutoAL approach is interesting and easily followed.\n3. The paper is well organized. \n4. The experiments validate the effectiveness of the proposed approach, and the ablation study in Figure 5 is insightful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper attempts to tackle the \"generalization problem\" of active learning (AL) algorithms across data scenarios. I believe this is a core issue in the current active learning field. This paper proposes AutoAL, a differentiable AL strategy search method to select the most effective AL sampling strategies in each iteration. It consists of two neural nets, named SearchNet and FitNet, which are optimized concurrently under a differentiable bi-level optimization framework. The experiments on multiple datasets validate the effectiveness of the proposed approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My only concern is the efficiency of the AutoAL algorithm. Although more efficient solutions have been proposed to solve second-order optimization problems, I cannot find any relevant experiments to verify them." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "To our knowledge, we present the first automatic AL query strategy search method that can be trained in a differientiable way." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024autoal,\ntitle={Auto{AL}: Automated Active Learning with Differentiable Query Strategy Search},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4RRmy9iw3c},\nnote={under review}\n}" }, "abstract": { "value": "As deep learning continues to evolve, the need for data efficiency becomes increasingly important. Considering labeling large datasets is both time-consuming and expensive, active learning (AL) provides a promising solution to this challenge by iteratively selecting the most informative subsets of examples to train deep neural networks, thereby reducing the labeling cost. However, the effectiveness of different AL algorithms can vary significantly across data scenarios, and determining which AL algorithm best fits a given task remains a challenging problem. This work presents the first differentiable AL strategy search method, named AutoAL, which is designed on top of existing AL sampling strategies. AutoAL consists of two neural nets, named SearchNet and FitNet, which are optimized concurrently under a differentiable bi-level optimization framework. For any given task, SearchNet and FitNet are iteratively co-optimized using the labeled data, learning how well a set of candidate AL algorithms perform on that task. With the optimal AL strategies identified, SearchNet selects a small subset from the unlabeled pool for querying their annotations, enabling efficient training of the task model. Experimental results demonstrate that AutoAL consistently achieves superior accuracy compared to all candidate AL algorithms and other selective AL approaches, showcasing its potential for adapting and integrating multiple existing AL methods across diverse tasks and domains." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Active Learning", "Differentiable Bi-level Optimization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/8cdd3d608b6f90a698fca18f279a8639dc3683d3.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "AutoAL: Automated Active Learning with Differentiable Query Strategy Search" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Many of the evaluations are on datasets for specific molecules, which are very useful for understanding specific model behavior, but should be complemented by more general evaluations, especially in the context of examining scaling behavior. Can the authors comment on or provide additional evidence that these specific, bespoke evaluations are connected to more general relationships between pre-training setups and downstream evaluations?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper continues an important line of work in exploring the utility of pre-training and scaling GNNs for molecular learning tasks. Unlike other dominant areas of deep learning, all-atom molecular representation learning relies on GNNs and does not directly benefit from advances in scaling sequence-based models.\nThe authors explore model size, aspect ratio, nearest neighbor cutoff radius, and architecture to provide a comprehensive look into the scaling behavior of molecular GNNs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper the authors extend previous work on scaling laws for all-atom graph neural networks applied to self-supervised and supervised training. They investigate aspects of pre-training task choice, different downstream evaluations, GNN model size and aspect ratio, as well as the radial cutoff for constructing nearest neighbor graphs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Like most molecular GNN works, the authors are limited by the available evaluations (e.g., QM9). xxMD is an interesting evaluation, but these datasets are limited to a small set of specific molecules. QM9 with B3LYP in particular is not informative, and the authors might consider newer benchmarks like POLARIS or a subset of the Therapeutic Data Commons to strengthen their evaluations.\nThe paper does not offer a clear and concise summary of recommendations based on the empirical findings, which is essential to achieving the stated aim of inspiring practitioners to rethink GNN training." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The paper claims to demonstrate zero-shot transfer learning using pre-trained Geom-GNNs. However, the described method seems closer to downstream task fine-tuning, given that a separate network is trained with the VAMP score objective. Could the authors clarify how this approach qualifies as zero-shot transfer rather than fine-tuning?\n\n2. Limited Pre-Training Approaches:\nThe pre-training in this work is restricted to coordinate denoising. Given that prior work has successfully used a combination of node mask prediction and coordinate denoising for improved performance. Specifically, how might adding a node mask objective influence for the molecular kinetics tasks? Would it enhance the model’s ability to generalize across different molecular conformations?\nAdditionally, could the authors hypothesize the potential impact of such an extended pre-training approach on scaling behavior? \n\n3. While the effectiveness of ET and ViSNet is demonstrated on several tasks, the study lacks comparisons with invariant feature-based networks (e.g., SchNet, DimeNet, GemNet) and tensor product-based networks (e.g., Equiformer, MAC). Could the authors provide insights into how their method might perform relative to these alternative architectures?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Demonstrates substantial performance improvements in kinetic modeling on VAMP score metrics by utilizing denoising pretraining techniques.\n2. Applies the self-supervised pretraining approach to a variety of downstream tasks, successfully proving its effectiveness across applications.\n3. Examines scaling laws in both standard equivariant models and pre-trained equivariant models, finding that even pre-trained models diverge from typical neural scaling laws due to issues like under-reaching and over-smoothing. The author suggests that, while scaling models has its benefits for supervised and unsupervised tasks, it may be more effective to focus on addressing data label uncertainty and using active token mixing to mitigate information bottlenecks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates denoising pretraining and potential scaling laws for geometric graph neural networks (GNNs) on supervised tasks. These GNNs are pre-trained on noisy coordinate data to learn how to denoise and reconstruct the original coordinates. The effectiveness of this approach is tested on various downstream applications, including molecular kinetics, fold classification, and energy and force prediction. Additionally, the paper examines the scaling behavior of these models and highlights specific limitations in supervised prediction tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The author claims to be the first to demonstrate the pre-trained Geom-GNNs’ capability as zero-shot transfer learners. However, I would consider this more appropriately described as downstream task fine-tuning. In zero-shot learning, a model test on data source is directly applied to unseen data class without additional training (e.g., training on English and French, then testing on Chinese without further adjustments). Unlike this approach for molecular kinetics, the paper involves training a separate network with the VAMP score objective.\n2. The pretraining methods in this work are limited to coordinate denoising. Other approaches [1,2] that leverage both node mask prediction and coordinate denoising have already proven effective.\n3. Although the paper demonstrates the effectiveness of ET and VisNet on several tasks, it does not include evaluations on invariant feature-based networks (such as SchNet, DimeNet, or GemNet) or tensor product-based networks like Equiformer and MAC.\n\n[1] Cui, Taoyong, et al. \"Geometry-enhanced pretraining on interatomic potentials.\" Nature Machine Intelligence 6.4 (2024): 428-436.\n[2] Zhou, Gengmo, et al. \"Uni-mol: A universal 3d molecular representation learning framework.\" (2023)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "as Weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Innovation**: The study presents a novel perspective, exploring an area that remains under-researched (to my knowledge), offering significant insights for the development of Geom-GNNs.\n2. **Clarity and Structure**: The article is well-organized, with clear presentation and summary of experiments and viewpoints, facilitating reader comprehension.\n3. **Robust Experimentation**: The experimental design is thorough, effectively supporting the authors’ conclusions.\n4. **Exploration of Zero-Shot Transfer Capability**: The investigation into the zero-shot transfer ability of Geom-GNNs is intriguing, with experiments indicating their potential as excellent zero-shot transfer learners.\n5. **Pre-training Insights**: Through extensive denoising pre-training tasks, valuable experiences have been gained regarding the pre-training of Geom-GNNs, including aspects such as model width, depth, aspect ratio, and the cutoff radius in geometric atomic graph construction, providing rich guidance for pre-training.\n6. **Advancement of Unified Architecture**: Given the widespread attention and efforts in the research of all-atom Geom-GNNs, this study effectively inspires researchers to reconsider the design of Geom-GNN architectures and the adjustment of training strategies, thereby promoting the development of a unified Geom-GNN architecture." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents a novel investigation into whether pre-trained Geom-GNNs (Graph Neural Networks for Conformational Molecules) possess efficient and transferable geometric representation capabilities, particularly in addressing the low generalization of models typically trained on specific tasks. The authors also aim to introduce “Neural Scaling Laws” to summarize the performance behavior of these pre-trained Geom-GNNs. However, it is unfortunate that the experimental results indicate that Geom-GNNs do not adhere to power-law scaling laws and fail to demonstrate predictable scaling behavior across various supervised tasks. Furthermore, the findings reveal that the all-atom embedding graph representations derived from Geom-GNNs exhibit excellent expressive capabilities, suggesting that Geom-GNNs can function effectively as zero-shot transfer learners." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Figure 6, which explores different model widths, even though the x-axis represents the total number of parameters (with model depth held constant), it would be more beneficial to indicate the model width for each point in the legend to enhance result presentation. Similarly, in Figures 4 and 7, which demonstrate the impact of model depth, using the legend to specify the exact number of layers might be more effective. In general, clear legends are always advantageous.\n2. The comparison between models trained from scratch and those fine-tuned in Section 6.1 could be more comprehensive if extended to include model depth. Previous discussions (albeit in the context of pre-training) have presented certain viewpoints, and it is anticipated that these would also have significant effects during fine-tuning." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- Why is the VAMP task considered zero-shot?\n- Appendix D: Why is there a difference between datasets for pretraining and downstream? According to Table 2, QM9 there seem also experiments with pretraining. Are models in Table 3 not fine-tuned?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "relevant research topics in the paper:\n- pretraining of (geometric) GNNs\n- scaling behavior of GNNs\n- oversmoothing, underreaching\n\nexperiments:\n- experiments considered not only random splitting, but also scaffold and temporal splits" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In their work the authors empirically studied properties of pre-trained geometric GNNs (i.e., GNNs with coordinates attached).\nThey especially seemed to consider pretraining via denoising as introduced by Zaidi et al. (2022).\nThe authors consider several downstream tasks: molecular dynamics (kinetics modeling), fold classification, prediction of quantum-chemical properties, and, molecule conformations\nThey study properties such as power-law scaling and find that the geomtric GNNs do not follow that on the pre-training task.\nThe authors conclude that geometric GNNs \"saturate early and are bounded by issues such as under-reaching and over-smoothing\" and further that \"other aspects become more crucial, such as addressing data label uncertainty and employing active token mixing to mitigate the information bottleneck.\"" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "In general the largest problem to my opinion is that the authors lack to specify a clear research goal. Methodologically there seems not too much novelty.\nBut if it is primarily an empirical research paper it seems very important to me that the research goal is clearly defined and dependent on that authors have\nto reason why they are selecting the problems they look at and why they are selecting the methods they compare to. If a new benchmark dataset is used (as e.g., scaffold QM9), then empirical evaluations/comparisons should take into account at least some previously suggested methods (which are ideally complementary to each other to better see the potentials of each of the methods on the new benchmark and being able to compare to what authors say).\n\\\n\\\nDetails:\n\\\n\\\n**hard to see the novelty in the paper**:\\\nFact that pretraining is useful was already found e.g. by Zaidi et al. (2022).\\\nNo new methodology seems to be suggested (or is \"token mixing\" the new method)?\\\nStudy of scaling behavior might be interesting and to a certain degree novel for geometric GNNs, but no follow-up investigations seemed to be employed to draw formal connections to GNN properties like oversmoothing, underreaching, etc.\n\\\n\\\n**paper is structured in a strange way, which makes it hard to follow the paper**:\\\nIt is actually hard to understand what the research aim of the authors was.\\\nchapter 4, 5, and, 6, actually seem to be about empiricial experiment results.\\\nThe setups are partly however already explained in chapter 3.\\\nchapter 4 and 6 show performances on problems\\\nchapter 5 however studies power-law behavior and other ablations.\\\nIn sum it gets hard to follow the paper. Better reasoning why some experiments are applied at all (with respect to the general research goal) and why they are done as they are done is necessary (e.g., why which methods are selected for comparison or why which dataset is used).\n\\\n\\\n**experiments**:\\\nalthough there are some good points as mentioned in strengths (such as the splits), it will get hard to understand how large the impact of pretraining really is, as the authors only test very few method once with pretraining and once without pretraining.\\\n\\\nalso the there is no good argument in the paper, why the authors exactly compare to those methods they selected to compare to \ne.g.:\\\nfor molecular dynamics, why don't they compare to Timewarp\\\nfor QM9 and xxMD they could e.g. compare to \"E(n) Equivariant Graph Neural Networks\", \"SE(3)-Transformers\", DimeNet++, MACE, etc.\\\nAn option to get more impression on the significance of the author's results would be to additionally compare to standard QM9, etc. (where there are also a lot of method comparisons out).\n\\\n\\\n**minor points**:\\\n\"token mixing\" not defined, but heavily used\\\ngrammatical errors/typos: \"In silico molecular computation and simulation are indispensable tool in...\"" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper explores using pre-trained geometric graph neural networks as effective zero-shot transfer learners for molecular conformation representation in out-of-distribution scenarios, and their scaling behavior under various setups." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024pushing,\ntitle={Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4S2L519nIX},\nnote={under review}\n}" }, "abstract": { "value": "Constructing transferable descriptors for conformation representation of molecular and biological systems has been a long-standing challenge in drug discovery, learning-based molecular dynamics, and protein mechanism analysis. Geometric graph neural networks (Geom-GNNs) with all-atom information have transformed atomistic simulations by serving as a general learnable geometric descriptors for downstream tasks such as interatomic potential and molecular property prediction. However, common practices involve supervising Geom-GNNs on specific downstream tasks, which suffer from the lack of high-quality data and inaccurate labels, potentially leading to poor generalization and performance degradation on out-of-distribution (OOD) scenarios, especially with quantum chemical data. In this work, we explored the possibility of using pre-trained Geom-GNNs as transferable and highly effective geometric descriptors for improved generalization. To explore their representation power, we studied the scaling behaviors of Geom-GNNs under self-supervised pre-training, supervised and unsupervised learning setups. We found the expressive power of different architectures can differ on the pre-training task. Interestingly, Geom-GNNs do not follow the power-law scaling on the pre-training task, and universally lack predictable scaling behavior on the supervised tasks with quantum chemical labels. More importantly, we demonstrate how all-atom graph embedding can be organically combined with other neural architectures to enhance the expressive power. Meanwhile, the low-dimensional projection of the latent space shows excellent agreement with conventional geometrical descriptors." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Geometric Graph Neural Networks", "Self-supervised Pre-training", "Scaling", "Zero-shot Transfer", "Molecular Representation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/80c45a708014eea392ed8db0e423c825ba02b4d0.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling, and Zero-Shot Transfer" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Can authors provide visualizations of generating consistent scenes based on slightly different conditions (e.g., the ego car moves differently), which is a key aspect for closed-loop evaluation?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This submission targets an important problem in the field of autonomous driving: how to properly evaluate the performance of end-to-end systems. The submission introduces a closed-loop evaluation method, which is more reflective of real-world driving conditions compared to open-loop evaluations. It will be useful for practical applications.\n2. The platform utilizes road networks from cities worldwide and allows for the generation of diverse traffic scenarios with varying styles, which is essential for training and evaluating driving agents across different driving environments. \n3. The submission provides a clear and detailed explanation of the technical aspects of DriveArena. The figures, tables, and appendices enhance the understanding of the system's components and their interactions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The submission introduces DriveArena, a high-fidelity closed-loop simulation platform designed for testing and developing autonomous driving agents in real-world scenarios. The platform consists of two main components: the Traffic Manager and the World Dreamer. The Traffic Manager is responsible for generating realistic traffic flow on any global street map, while the World Dreamer is a high-fidelity conditional generative model that creates infinite autoregressive simulations. DRIVEARENA enables the generation of diverse traffic scenarios with varying styles and allows driving agents that can process real-world images to navigate within its simulated environment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The submission lacks novelty. The proposed DriveArena is just a combination of several existing methods: LimSim for traffic simulation and condition generation; DriveDreamer/Vista for generating images from the conditions; NAVSIM and Carla for closed-loop evaluation. \n2. The World Dreamer model is trained primarily on the nuScenes dataset, which may not capture diverse driving scenarios. To improve the model's generalizability, it would be beneficial to incorporate additional datasets that represent different geographical locations, driving cultures, and road conditions.\n3. This submission fails to address an important issue for closed-loop evaluation: the model should be able to generate the same scene captured from different positions (similar to actual scenarios of driving differently in the same scene). No visualization was found addressing such an issue. \n4. Experiments are not enough. The submission primarily focuses on the simulation platform itself rather than an in-depth evaluation of various driving agents within the platform. Expanding the experimental section to include a broader range of driving agents and more extensive testing can help provide a clearer picture of DRIVEARENA's capabilities and limitations.\n5. Minor issues in writing and presentation. For example, The figures are not vectorized for zooming in and they are suggested to be replaced." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In Lines 261-263 it is written \"We also verify that extending to a multi-frame auto-regressive version (using multiple past frames as reference and outputting multi-frame images) and adding additional temporal modules can enhance temporal consistency.\" - How does this work in practice? In closed-loop, the ego-vehicle can drive however it wants and so there can still be inconsistencies between generated videos at time t and t+1, right? Further, how are the generated videos used? Are T frames predicted but only the 1st one is shown to the ego-policy? And then a new prediction is made (similar to model-predictive control)?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The topic of generative simulation is an especially timely one for the AV field, with better and better generative video models coming out regularly and frameworks such as this one being able to capitalize on these parallel investments.\n\nWhile all of the individual pieces of this framework have existed before, connecting them all together into a usable simulation framework for the broader community to use is appreciated and presents a potential path towards practical generative simulation.\n\nThe paper is written well and it is easy to follow the core ideas as presented.\n\nLeveraging layout-conditional generation is a sound and sensible idea for maintaining consistency across time." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents a generative simulation framework for autonomous driving. In particular, a layout-conditional diffusion model is proposed as a sensor simulator, with bounding boxes and road graphs serve as underlying world state." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The core weakness is that any geometric or semantic consistency is not guaranteed. Layout conditioning certainly helps, but in the linked webpage's videos there are clear inconsistencies across timesteps (e.g., car color and types changing over time). This is something that is not brought up in Lines 190 - 198, but perhaps should be as it is the core reason why works leverage NeRF or 3D Gaussian Splatting (for their geometric/semantic/temporal consistency over purely-2D generative models).\n\nWhile static images of generations appear to be of good quality, there are significant temporal consistency issues when viewed as part of a video on the linked project page (most videos appear to be static even with the ego-vehicle theoretically moving forward in the world). Do the authors have any idea for why that is? It almost appears that the video diffusion model suffers from mode collapse when tasked with generating building walls (taking example from the linked webpage).\n\nThe AV results in Tables 1 and 2 still show a significant gap to real data, indicating that, while the core points of DriveArena are sensible, there is still much work to be done to leverage it for practical AV development." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* Evaluation of sim-to-real gap: How are the videos generated? Is this in open- or closed-loop? For correctness, it should be closed-loop. If so, do you notice any suffering from DAgger issues?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Highly relevant research direction: The work correctly argues for the need of closed-loop evaluation of autonomous driving behavior models.\n* Novelty: High-fidelity closed-loop image-based simulation with clear controllability (e.g. via text prompts).\n* Performance: Evaluation of sim-to-real gap shows superiority over MagicDrive and reasonable results for open-loop and closed-loop behavior eval.\n* Well presented: The paper is easy to follow and all information is well presented." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents DriveArena, an image-based high-fidelity simulator for closed-loop simulation of agents for autonomous driving. DriveArena consists of a Traffic Manager and a World Dreamer, and its modular design allows for easy replacement of each of the components. Traffic Manager enables the dynamic control of all traffic agents and supports various HD maps, both of which are inputs to World Dreamer, which uses conditional diffusion to generate realistic images. The diffusion model is conditioned on map and object layout and generates images autoregressively for temporal consistency and variable length simulations.\n\nDriveArena is evaluated in two ways. First, its fidelity is evaluated, among others, with UniAD's performance. Then, open- and closed-loop evaluation of VAD and UniAD models is performed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Evaluation of sim-to-real gap: The presented evaluation is pretty short and for example lower L2 distances do not necessarily imply higher quality images. Additional evaluation of the fidelity would be helpful. Are there perception metrics such as FID that could be used or other metrics that compare statistics between ground truth and generated images? Otherwise, user studies are another possibility to judge the quality.\n* Unclear takeaway from VAD vs. UniAD open- and closed-loop comparison: In open-loop, UniAD performs better on nuScenes but worse on DriveArena than VAD. This difference is explained with better open-loop generalization of VAD. However, it's unclear what role the fidelity of DriveArena plays. Is it possible to e.g. run an experiment with different DriveArena datasets, some that are closer and some that are further from nuScenes? In closed-loop eval, UniAD outperforms VAD in DriveArena. It's unclear whether these differences are due to open- / closed-loop model gaps or issues in DriveArena. I acknowledge that this difficulty of correct attribution is inherent to research in this area but you might be able to shed more light on this. For example, would it be possible to evaluate the models on various levels of fidelity in DriveArena to disentangle open- / closed-loop eval from it?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "+ About constructing traffic scenarios from OSM.\n + How is the HD map built from the OSM data, and how is traffic demand generated in this kind of scenario? \n + OSM maps are not high quality in many areas, is there a way to solve this?\n + Is this part of the work mainly based on the tools provided by SUMO or LimSim?\n+ Are there any indicators to evaluate the generated results directly, like FID or FVD?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Compared to previous work using video generation methods as world models to achieve realistic traffic simulation, this work uses a two-step pipeline, including rule-based traffic simulation in geometric space and diffusion-based video generation conditioned on trajectory layouts of vehicles. I believe this approach can achieve better physical realism and temporal consistency.\n\nText prompts are introduced to achieve diverse driving scenarios and plenty of demos are presented to clearly show the generation results.\n\nThe codes are well organized with a modularized design and the whole platform is open-source to better support downstream tasks in research of autonomous driving." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a traffic simulation platform for testing autonomous driving algorithms. The simulation architecture is built based on LimSim, using Monte Carlo tree search for vehicle motion planning. A diffusion-based renderer is applied to achieve realistic and controllable surround images for vision-based driving agents. The platform supports driving scenario construction from nuScenes and OSM; codes and tools are open-source for the community to use." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The author says that one of the main contributions is scalability, which means simulation on any region can be achieved by using map info from OSM. As far as I know, OSM only contains road-level map information, and extra efforts like completing lane topology and planning vehicle OD are needed to construct simulations based on it, this part of the work seems unclear in this paper.\n\nAs the dreamer is the most important part of this paper, it would be better if the author could provide some indicators that can directly evaluate the generated results, like FID.\n\nMinor note: it seems that Figure 3 is not in vector format." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "DriveArena is a pioneering closed-loop autonomous driving simulator based on conditional generative models for training and testing driving agents." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024drivearena,\ntitle={DriveArena: A Closed-loop Generative Simulation Platform for Autonomous Driving},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4S9bBbX1be},\nnote={under review}\n}" }, "abstract": { "value": "This paper introduces DriveArena, the first high-fidelity closed-loop simulation system designed for driving agents navigating real-world scenarios. DriveArena comprises two core components: Traffic Manager, a traffic simulator capable of generating realistic traffic flow on any global street map, and World Dreamer, a high-fidelity conditional generative model with infinite auto-regression. DriveArena supports closed-loop simulation using road networks from cities worldwide, enabling the generation of diverse traffic scenarios with varying styles. This powerful synergy empowers any driving agent capable of processing real-world images to navigate in DriveArena's simulated environment. Furthermore, DriveArena features a flexible, modular architecture, allowing for multiple implementations of its core components and driving agents. Serving as a highly realistic arena for these players, our work provides a valuable platform for developing and evaluating driving agents across diverse and challenging scenarios. DriveArena takes a significant leap forward in leveraging generative models for driving simulation platforms, opening new avenues for closed-loop evaluation of autonomous driving systems." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Autonomous Driving", "Diffusion Model", "Closed-loop Simulation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f71f3b768df9f40d995082ce12f052f327c6cb3b.pdf" }, "presentation": null, "primary_area": { "value": "applications to robotics, autonomy, planning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/ca33a46441fac4362d7871fbf6ef590e90fffc06.zip" }, "title": { "value": "DriveArena: A Closed-loop Generative Simulation Platform for Autonomous Driving" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "None" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "To generate reasoning paths for VideoQA, would it be more effective to design a Chain-of-Thought dataset and perform supervised fine-tuning (SFT)? The O1 model currently adopts this approach, achieving clear reasoning paths through an end-to-end structure." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The MoST-Grounding module introduces two standard modules—temporal localizer and spatial localizer—which can be flexibly assembled in sequence based on the question parser. This structure is robust and allows for reliable generation of reasoning paths.\n\n2. The authors present a clear motivation: to create a framework for generating reasoning paths for the black-box nature of VideoQA tasks. The comprehensive visualization of reasoning paths further demonstrates the effectiveness of the model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the use of multimodal Large Language Models in understanding tasks across various multimodal scenarios, specifically focusing on applications in Video Question Answering. However, current Multimodal Large Language Models are largely black-box systems for VideoQA tasks, lacking the ability to provide an understandable reasoning path and, thus, suffer from limited interpretability. To address this, the authors propose MSR-ViR, which constructs a modular reasoning structure designed to generate reasoning paths, and incorporates a reinforcement learning framework to prevent the model from generating unreasonable reasoning paths.\n\nWhile the proposed approach is interesting, it relies on the integration of four existing models. This ensemble-based structure shows only marginal performance improvements (1-2%), and the manuscript does not discuss crucial aspects such as reasoning time costs or memory overhead." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The question parser and LLM reasoning components both rely on LLM structures, leading to high computational costs.\n\n2. Both the temporal localizer and spatial localizer use existing models, specifically UniVTG and YOLO-World, which contribute to significant parameter overhead. As the complexity of VideoQA tasks increases, relying on these two models alone may not only limit predictive accuracy but also compromise the completeness of reasoning paths. Future work may need to explore additional modules to support diverse combinations (see [1] for reference).\n\n3. The ablation study lacks comprehensiveness. While the authors assess model performance on QA and grounding tasks and provide an effectiveness analysis of each module, they do not evaluate inference speed, parameter count, or other metrics compared to end-to-end models. Given that the proposed framework integrates multiple existing large models, an analysis of inference speed is both important and currently missing.\n\n[1]. Neural Module Networks" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "see weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tSuccessfully develop a ground-then-answer framework for interpretable video question parsing. The question parser policy is able to be optimized via answer feedback. \n\n2.\tThe approach is well presented and easy to understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to improve the interpretability of Multimodal LLMs in performing VideoQA. To achieve the goal, the authors design a modular self-reflection framework MSR-ViR. The framework primarily comprises a spatial-temporal grounding module and a self-refection learning mechanism based on DPO. MSR-ViR basically decouples video grounding from videoqa, enabling the interpretaion of intermediate results to understand the answers. The experiments on related datasets have demonstrated the strength of the approach in both accuracy and interpretability (grounded accuracy)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe paper does not improve video grounding but just uses existing method UniVTG. According to table 2, the grounding performance in terms of IoP@0.5 is worse than previous VLMs (VGT and TempCLIP). This severely limits the improvements of QA performance. \n2.\tAccording to the model ablation results in Table 3, the global representation g_v (which opens back door for grounded QA) seems more crucial than other components. Such results slightly depart from the major claim of interpretable VQA where correct answers are anchored on correct visual content.\n3.\tShould compare with SeViLA which also finetunes a localizer on QVHighlight (like UniVTG) for grounded VideoQA." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See Weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) The paper is well written and the motivation is clear, with a strong focus on the interpretability challenge in VideoQA, making it highly relevant to the field.\n2) The MoST-Grounding module integrates multiple submodules to effectively localize temporal and spatial regions, improving transparency in the input to the MLLM.\n3) The Alternate Self-Reflection strategy introduces a novel reinforcement-based method to align the question parser and the MLLM, enhancing performance and consistency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the black-box problem of multimodal large language models (MLLMs) in VideoQA by proposing MSR-ViR, a novel framework. MSR-ViR introduces two core components: (1) the MoST-Grounding module, which localizes relevant temporal segments and spatial regions in videos, and (2) an Alternate Self-Reflection Training Strategy, which iteratively enhances the question parser and the MLLM. Evaluations on datasets such as NExT-QA, STAR, and NExT-GQA demonstrate that MSR-ViR achieves competitive performance on VideoQA tasks and improves interpretability by providing visual evidence for answers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The framework is relatively heavy, relying on multiple external tools for tasks and additional operations such as resizing and truncation, which increases computational overhead. Moreover, what if these external tools are unreliable, which can lead to further exposure bias? It's necessary to further investigate the choice of the sub-modules in the MoST-Grounding module.\n2) While the approach improves the selection of input information, it does not make the internal reasoning process of the MLLM more interpretable. It still focuses on the process of 'input' to decide which information should be fed into the MLLM as soft prompts. \n3) The paper misses references with related works such as SeViLa [1] and GCG [2], which also focus on VideoQA with grounding elements. Including these baselines would strengthen the empirical validation.\n\n[1] Yu et al. \"Self-Chained Image-Language Model for Video Localization and Question Answering\", 2023 NIPS\n[2] Wang et al. \"Weakly Supervised Gaussian Contrastive Grounding with Large Multimodal Models for Video Question Answering\", 2024 ACM MM" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024msrvir,\ntitle={{MSR}-ViR: Modularized Self-reflected Video Reasoner for Video Question Answering},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4Sv5MQ931E},\nnote={under review}\n}" }, "abstract": { "value": "Recently, multimodal large language models (multimodal LLMs) have been applied to a wide range of video understanding tasks, particularly for Video Question Answering (VideoQA). However, existing multimodal LLMs suffer from the following challenge: the classic end-to-end training strategies of multimodal LLMs for VideoQA tasks are black-box, thus lacking interpretability as they can neither present a reasoning path nor indicate where the answer is derived from the video. To tackle this challenge, we propose MSR-ViR (Modularized Self-Reflected Video Reasoner), a self-reflected framework that introduces a Modularized Spatial-Temporal Grounding (MoST-Grounding) module to multimodal LLMs for VideoQA tasks. MoST-Grounding utilizes a question parser LLM to generate execution policies, which serve as a reasoning path from questions to answers providing interpretability for our VideoQA framework. Based on the execution policies, MoST-Grounding invokes various small modules to localize temporal segments and spatial regions in videos which provide multimodal LLMs with most relevant visual information, while presenting visual evidence of our final answers. To avoid the question parser LLM generating unreasonable policies, we further propose a reinforcement learning-based Alternate Self-reflection training strategy to optimize the Multimodal LLM and the question parser LLM. Experiments on VideoQA datasets (NExT-QA and STAR) and grounded VideoQA dataset (NExT-GQA) demonstrate that our method significantly improves video understanding capabilities of multimodal LLMs, while providing interpretable reasoning paths together with temporal and spatial localization evidence within the video." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Video Question Answering", "Multimodal LLM", "Modular Network", "Self-reflected Training" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/bc2578be11dcf758ee1e9d5318f4727faf706324.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "MSR-ViR: Modularized Self-reflected Video Reasoner for Video Question Answering" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "For quite a few models on the leaderboard, the MMLU score will be random chance (~25%, which you can see in Figure 1). Would it be a useful preprocessing step to subtract out random chance from the score and renormalize? E.g. take (score - 0.25) / (1 - 0.25)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper is well-written and clearly communicates its ideas and methods. In the choice of the IRT, multiple models and methods for estimating the ability are explored. The method proposed produces a much smaller benchmark which the authors demonstrate has better predictive power than randomly subsampling items (Figure 1B). Careful consideration is given to potential limitations of the method, including assumptions about the conditional independence of the LLMs used for the study. The work also considers the interesting idea of a benchmark that performs adaptive testing in which items are selected sequentially based on a current estimate of the model's ability.\n\nOverall I think the paper makes meaningful contributions to studying LLM benchmarks and making model evaluation more efficient, and I thus lean towards acceptance. However, I do think the benchmarks considered are missing some of the abilities that people seek to measure in LLMs (e.g. coding), somewhat limiting the work's impact. I seek to provide concrete suggestions regarding this in the next section." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers the six LLM benchmarks included in the Open LLM Leaderboard (ARC, GSM8K, HellaSwag, MMLU, TruthfulQA, and WinoGrande) and seeks to create a much smaller benchmark that is predictive of the original suite by subselecting items. This is done using data from more than 5000 LLMs included in the leaderboard and a pyschometric method called item response theory (IRT) which in essence fits a model that estimates the item's difficulty and how well the item discrimates between models whose \"abilities\" are close to the item's difficulty. (Note this model ability is also fit by the method in an alternating fashion.) The presented method results in a benchmark that is only 3% the size of the original benchmark but is able to effectively reconstruct both the original individual benchmark scores and the joint score. Finally, using factor analysis, the authors demonstate that a single latent is predictive of all 6 benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My comments in this section are not intended to be required changes to the paper but rather a discussion of what I think the authors could add to have more significant impact.\n\nCurrently the main output of the paper is a much smaller benchmark that can be used to efficiently rank models on the six benchmarks as well as evidence from factor analysis that all six benchmarks are measuring a single latent ability. However, across the broader field of LLM benchmarks, it is generally assumed that there are multiple latent dimensions to the abilities of LLMs. For example, if a code benchmark was added into the set, I would assume this would require another latent dimension to fit model performance, and it would be intriguing if this was not true! Also I would be curious if a larger fraction of the test items is required to reconstruct the scores when the set of included benchmarks require multiple latent ability dimensions to represent.\n\nIn essence, the most interesting direction I see for this work is to apply the methods to a more comprehensive set of benchmarks to try to discover latent ability dimensions that might be interpretable as what we think of as LLM capabilities. This should then also provide a characterization of which of these abilities each benchmark measures." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Will a small benchmark lead to a large variance in evaluation?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper distills six prominent LLM benchmarks into a much smaller one with less than 3% of the size, which enables more streamlined and cost-effective evaluation methods;\n2. The new sparse benchmark yields estimators able to reconstruct the original benchmark score." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces metabench, a sparse benchmark distilled from six prominent benchmarks (ARC, GSM8K, HellaSwag, MMLU, TruthfulQA, and WinoGrande). Simple criteria, cross-validated subsampling, and information-based filtering are used to reduce the size of the benchmark. Original scores are reconstructed in a cross-validated manner." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "As mentioned in the limitations section, a smaller benchmark has the risk of being memorized." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could authors elaborate on potential limitations when applying Metabench to other domains?\n2. How might Metabench handle scenarios where specific benchmarks assess unique skills not captured by a general latent factor?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The paper’s technical approach is methodologically sound, with robust use of IRT and statistical modeling to identify informative items.\n2. It is well-organized, with a clear explanation of Metabench’s goals and psychometric techniques.\n3. It makes a substantial contribution to LLM evaluation, providing a novel, efficient, and scalable benchmarking solution." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Metabench, a sparse benchmarking method designed to evaluate large language models (LLMs) with minimal redundancy and resource demands. By analyzing data from over 5000 LLMs across six benchmarks (ARC, GSM8K, HellaSwag, MMLU, TruthfulQA, and WinoGrande), Metabench distills these into a much smaller subset, reducing the combined item count by over 97%. Using psychometric techniques such as Item Response Theory (IRT), Metabench selects the most informative items, facilitating efficient and accurate evaluation while maintaining the integrity of the original benchmarks. The sparse benchmark achieves impressive fidelity, reconstructing original scores with less than 1.24% RMSE on average, and identifying a single common latent factor strongly correlating with general model ability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The framework currently focuses on six benchmarks; additional work could explore its applicability across a broader range of LLM tasks or domains.\n2. Metabench’s dependence on psychometric models, especially IRT, could be limiting if these models do not fully capture the complexities of LLM behavior, as they were traditionally designed for human subjects." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Could alternative item selection methods (beyond Fisher information) yield better results?\n2. How stable are the results across different random seeds and model subsets?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Novel application of psychometric methods to LLM evaluation\n2. Impressive compression ratio (<3% of original size) while maintaining accuracy. Low reconstruction error (1.24% RMSE for individual benchmarks, 0.58% for total score).\n3. Comprehensive ablation studies and baseline comparisons. Thorough investigation of factor structure across benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes metabench, a compressed version of six popular LLM benchmarks (ARC, GSM8K, HellaSwag, MMLU, TruthfulQA, and WinoGrande) that achieves comparable evaluation capability while using less than 3% of the original items. The authors leverage psychometric techniques, particularly Item Response Theory (IRT), to identify the most informative test items and estimate latent abilities that can reconstruct original benchmark scores with high accuracy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Memorization risks: \n (1) Smaller benchmark size increases memorization vulnerability\n (2) Proposed mitigation strategies need further validation\n2. Theoretical Assumptions:\n (1) IRT assumptions about LLMs need more justification\n (2) Independence assumptions between models may be violated due to shared architectures/training data" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "metabench is the data-driven distilled version of six popular LLM benchmarks, drastically reducing their size with minimal information loss" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024metabench,\ntitle={metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4T33izzFpK},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) vary in their abilities on a range of tasks. Initiatives such as the $\\texttt{Open LLM Leaderboard}$ aim to quantify these differences with several large benchmarks (sets of test items to which an LLM can respond either correctly or incorrectly).\n However, high correlations within and between benchmark scores suggest that (1) there exists a small set of common underlying abilities that these benchmarks measure, and (2) items tap into redundant information and the benchmarks may thus be considerably compressed.\n We use data from $n > 5000$ LLMs to identify the most informative items of six benchmarks, $\\texttt{ARC}, \\texttt{GSM8K}, \\texttt{HellaSwag}, \\texttt{MMLU}, \\texttt{TruthfulQA}$ and $\\texttt{WinoGrande}$ (with $d=28,632$ items in total). From them we distill a sparse benchmark, \\texttt{metabench}, that has less than $3\\%$ of the original size of all six benchmarks combined. This new sparse benchmark goes beyond point scores by yielding estimators of the underlying benchmark-specific abilities.\n We show that these estimators (1) can be used to reconstruct each original \\textit{individual} benchmark score with, on average, $1.24\\%$ root mean square error (RMSE), (2) reconstruct the original \\textit{total} score with $0.58\\%$ RMSE, and (3) have a single underlying common factor whose Spearman correlation with the total score is $r = 0.94$." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "llm", "benchmarking", "item response theory", "factor analysis", "information" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/dc39dc3c648417a95ca13fd0f37e3b3c859d724e.pdf" }, "presentation": null, "primary_area": { "value": "datasets and benchmarks" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/b2ee3999e2552d5ae5a9fd8cd89f5e720281bae2.zip" }, "title": { "value": "metabench - A Sparse Benchmark of Reasoning and Knowledge in Large Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Could \"scaling the field of view of the network so the size remains constant while increasing to full resolution\" make the full-resolution model handicapped in terms of field of view? There is a chance this is already answered in the paper but the reader missed it.\nOther than that, fixing the presentation issues in the previous section." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper is original in the sense of creative combinations of existing ideas (the components of the model that enable scaling), application to a new domain (forecasting on minimally preprocessed / raw data from weather to neural data), as well as removing limitations from previous works (removing dependency on segmentation mask accuracy and avoiding loss of information caused by conversion to trace-matrix). Writing style is very high quality, and all the methods and results get across in a relatively clear manner to the reader. To the specific line of research (forecasting neural data), the method proposed seems to be a significant step forward in the field’s development, moving from hand-crafted to learnable features." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel method for neural activity forecasting of zebrafish that works on the raw volumetric video instead of using standard preprocessing that reduces the original 4D space to 1D (trace matrix) and disregards spatial relationships between neurons. To do this, a u-net architecture is employed, taking advantage of a large scale neural dataset and performing extensive ablations for model selection. Multiple measures were taken to enable scaling the architecture for this computationally expensive problem, such as using the temporal context dimension as input channels, lead-time conditioning, and distributed training. The ablation results show that (1) pretraining on other specimens does not help, (2) there is a trade-off between spatial and temporal context, and (3) that downsampling input resolution up to 4x is beneficial to performance. Compared to the best trace-matrix models, the proposed multivariate model achieves 10% reduction in MAE for the short context forecasting setting in both the test and the holdout sets, while it is comparable to the trace-matrix models for long-context forecasting in the test set and 12% better in the holdout set." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are some presentation issues that impact the understanding of a reader that is not very familiar to forecasting neural data. \n* These are mainly in the abstract and introduction, while once entering section 2 all misunderstandings of this reader were resolved. Nevertheless it would be beneficial for the paper to have them fixed. (1) Until section 2 it was not very clear that the goal is to predict future steps from previous steps, and not one neural modality (e.g. electrical signal) from another (e.g. blood oxygen). (2) It was not clear that the neuron segmentation mask is applied in both the trace models and the proposed model, but at different points in the pipeline, i.e. in the latter the forecasting itself is done on the volumetric video and afterwards the mask is applied before computing the error - without knowing this it is not clear how the two methods can be compared fairly. It would also help if in Figure 1 the same notation was used between orange (trace) and blue (proposed) in the segmentation mask block, i.e. instead of “Extract Neurons” and “Mask Neurons” say “Apply segmentation mask” in both cases. (3) In the abstract the phrase “we design a model to handle the high resolution and large receptive fields…” is structured in a confusing way where the reader does not understand if the large receptive fields are an aspect of the model or the recordings. (4) Minor - the footnote on page 2 is not so much footnote information, but rather more suitable for the main text.\n* Additionally, Figure 2 should have a more informative caption that explains better what is shown, e.g. it is not clear what the colored blobs are, segmentation masks?\n* Is H=32 the only setting that is tested and why? Not sufficiently described in the paper.\n* In the conclusion, calling the findings counterintuitive seems excessive; there is no reason to assume that high input resolution, pretraining, or increased model capacity works well for all domains and applications. Results sufficiently showcase enough reasons why these sometimes useful settings might lead here to overfitting, distribution shifts, etc." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Major\n- It would be nice to test whether or not the trace-base model performed (relatively) poorly due to the lack of inter-cell activities or imperfect masking as suggested in Section 2 and Figure 2. I suggest applying the segmentation masks to the video, i.e. set regions outside of the identified cells as background, and train the video-based model on the masked videos. This should give us a sense of the influence of inter-cell activities and imperfect masking, and isolate the influence of spatial organization of the cells. \n- I suggest the authors include metrics that are commonly used in neural response prediction so that readers can have a sense of how well these models are performing, such as normalized correlation ($CC_\\text{norm}$) [1], or fraction of explainable variance explained (FEV) [2], basically metrics that takes trial-to-trial variability into account.\n- Can the authors comment on the computational cost of the models? The authors stated in the hyperparameter search section and appendix A.3 that 16 A100 40GB GPUs are used to train the video-based model, and ~5k GPU hours (so ~300 hours in wall-time) was used in the loss ablation experiment in Figure 4, which is a considerable amount of computational time and cost. Can the authors share the time it took to train the final (best) video-based and trace-based models? I believe the authors should discuss the trade-off between the two approaches in computation cost if they are indeed substantially different. To clarify, I think it is fine for the method to be more computationally expensive than other methods, but it is important to point it out.\n\nMinor\n- What is the frame rate of the video? \n- Why and how are the two temporal context lengths (4 and 256) selected? Does it make sense to predict the future 32 frames from only 4 frames?\n- In the hyperparameters section and Figure 1, it is stated that the models optimize the trace-based MAE. Does this include the video-based model? Since the video-based model inputs and outputs a video, does it make a difference to optimize the recorded and predicted video MAE?\n- How are the hyperparameters selected? Hand-picked or via some form of hyperparameter search (random search, bayesian search, etc.)\n\n[1] Schoppe, Oliver, et al. \"Measuring the performance of neural models.\" Frontiers in computational neuroscience 10 (2016): 10.\n\n[2] Cadena, Santiago A., et al. \"Deep convolutional models improve predictions of macaque V1 responses to natural images.\" PLoS computational biology 15.4 (2019): e1006897." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- To my knowledge, neural response prediction/forecasting directly from volumetric video is novel. This allows minimal preprocessing of the data which can be beneficial to deep learning-based methods.\n- A wide range of training and evaluation conditions are compared, including trade-offs of spatial and temporal resolution, pre-training vs direct training, and training set size and combinations. These empirical results can guide future work in modeling neural responses." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new approach to neuronal response modeling by predicting/forecasting the volumetric video instead of the per-neuron calcium trace (dF/F) or spike train, which is the norm in neural response prediction. This approach allows the model to take advantage of the inter-cell activity and spatial organization of the population that is typically discarded when deconvolving the volumetric video to individual response traces. The authors evaluated a range of video and trace-based models on ZAPBench and showed that the video-based model outperforms trace-based models in short temporal context length conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Please find my suggestions for the following points in the Question section.\n- A key motivation of this work, as stated by the authors in Section 2, is that the typical deconvolution step to convert volumetric video to dF/F response traces can lead to loss of information, such as cell spatial organization, inter-cell activities, etc. However, while the video-based model appears to outperform the trace-based model in short temporal context-length conditions (though similar performance in longer context length), it is unclear whether or not this is due to the additional information that exists in the raw data and that the video-based model is indeed taking advantage of such information.\n- MAE might not be the most intuitive metric for getting a sense of how the (video and traced-based) models are performing. For instance, I am not sure if an MAE value of 0.02 is good or bad, or how big of a difference is an MAE of 0.02 to 0.04? In particular, I believe ZAPBench is a new dataset and we don’t have any other models to compare against these MAE values, other than the single trace-based model provided.\n- Unclear trade-off in computation cost between video-based and trace-based models." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1.On line 308, the authors refer to a \"segmentation mask,\" yet on line 68, they claim that their method can directly process 4D data. The authors are requested to clarify the apparent contradiction in their narrative.\n\n2.In the fourth contribution (line 108), the authors state that their proposed method is the \"only approach that consistently benefits from multivariate information.\" However, I did not encounter any experimental justification related to multivariate information in the main text. If such experiments were conducted, please direct me to the relevant sections within the paper.\n\n\n3.The description of temporal dimension processing on line 200 is unclear. I would like to confirm whether the authors' approach involves merging the temporal dimension with the batch dimension, such as transforming the data shape as follows: (batch, 2048, 1152, 72, T) --> (batch*T, 2048, 1152, 72). If not, please provide clarification on their methodology." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.This paper employs a deep learning approach based on U-Net to directly process 4D neural activity recordings, circumventing complex preprocessing methods that may introduce performance degradation.\n\n2.A series of ablation studies have revealed practical insights into model pre-training and hyperparameter tuning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript proposes the utilization of deep learning techniques for the prediction of neuronal activity recordings with fluorescent calcium, asserting superior performance over previous baselines. A series of ablation studies have revealed practical insights into model pre-training and hyperparameter tuning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The authors fail to elucidate the significance of this study anywhere within the main text, which raises questions about the necessity of forecasting whole-brain neuronal activity. The authors are encouraged to provide additional context in the abstract or introduction section.\n\n2.This paper primarily utilizes the initial frames of neuronal activity recordings with fluorescent calcium indicators to predict subsequent frames, representing an application of U-Net in a specific domain. It does not offer additional insights or novel perspectives to advance the field of artificial intelligence. Therefore, this manuscript would be more appropriately submitted to conferences or journals focused on neuroscience or medical image processing, as it does not align with the thematic scope of ICLR.\n\n3.The paper presents a limited comparison with only one baseline, namely the \"trace-based model,\" as shown in Figure 6. It raises the question whether ZAPBench, as a benchmark, evaluated only this single model type. The authors are encouraged to include additional baselines for comparison to substantiate the superiority of the proposed method." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a new method for forecasting neuronal activity using volumetric videos, leveraging spatial relationships between neurons and outperforming traditional trace-based methods." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024forecasting,\ntitle={Forecasting Whole-Brain Neural Activity from Volumetric Video},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4UXIGATUTj},\nnote={under review}\n}" }, "abstract": { "value": "Large-scale neuronal activity recordings with fluorescent calcium indicators are increasingly common, yielding high-resolution 2D or 3D videos. Traditional analysis pipelines reduce this data to 1D traces by segmenting regions of interest, leading to inevitable information loss. Inspired by the success of deep learning on minimally processed data in other domains, we investigate the potential of forecasting neuronal activity directly from volumetric videos: we design a model to handle the high resolution and large receptive fields necessary for capturing spatio-temporal dependencies in volumetric whole-brain recordings. We explore effects of pre-training and perform extensive model selection, analyzing spatio-temporal trade-offs for generating accurate forecasts. Our model outperforms trace-based forecasting approaches on ZAPBench, a recently proposed benchmark on whole-brain activity prediction in zebrafish, demonstrating the advantages of preserving the spatial structure of neuronal activity." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "neuroscience", "forecasting", "video", "lightsheet microscopy", "zebrafish", "calcium imaging", "neuron activity" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/d295120254e9b0dc0a9f6797374cb7e4326166b4.pdf" }, "presentation": null, "primary_area": { "value": "applications to neuroscience & cognitive science" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Forecasting Whole-Brain Neural Activity from Volumetric Video" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the weakness above" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "HRVMamba demonstrates competitive or superior performance on COCO, Cityscapes, and PASCAL-Context benchmarks, often with fewer parameters and reduced computational load compared to similar models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes HRVMamba, a High-Resolution Visual State Space Model designed for dense prediction tasks, such as human pose estimation and semantic segmentation. The paper addresses limitations in existing Mamba-based models, particularly Mamba’s low-resolution output and challenges in retaining long-range dependencies. To overcome these issues, the authors introduce the Dynamic Visual State Space (DVSS) block, which leverages multi-scale and deformable convolutions to improve inductive bias and mitigate long-range forgetting. By integrating these innovations within a high-resolution, multi-resolution parallel structure, HRVMamba achieves competitive results across dense prediction benchmarks compared to CNN, ViT, and SSM models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Limited Novelty: In my view, this paper incorporates techniques from CNN networks, such as DCNv4 and multi-resolution structures (from FPN and HRNet), into the Mamba block to enhance network performance. I am somewhat skeptical about whether such an innovation alone is sufficient for a publication at ICLR." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the weakness part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Empirical results indicate that HRVMamba outperforms contemporary CNNs, ViTs, and SSMs on benchmarks, delivering competitive performance with fewer computational resources.\n- The figures in the paper are clean and aesthetically pleasing, which enhances readability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces HRVMamba, a high-resolution visual state space model tailored for dense prediction tasks. It builds on the Mamba framework, a hardware-efficient state space model (SSM) known for linear computational complexity. The authors highlight limitations in previous visual Mamba models—namely, insufficient inductive bias, long-range forgetting, and low-resolution outputs. To overcome these, HRVMamba incorporates the Dynamic Visual State Space (DVSS) block, combining multi-scale convolutional kernels and deformable convolution for enhanced local and long-range feature extraction. The HRVMamba model employs a multi-resolution parallel structure inspired by HRNet, preserving high-resolution representations and facilitating multi-scale feature learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Limited Novelty: The HRVMamba model mainly combines existing methods, including the VSS block, DWConv, DCN, and the HRNet architecture. This integration-based approach may not meet ICLR’s high standards for innovation.\n- Concern for `Limitation 1`: While the paper addresses the lack of 2D inductive bias in previous visual Mamba models, it raises concerns about whether introducing such bias could restrict the **scaling ability** of HRVMamba. Vision Transformers (ViTs) have demonstrated that reduced inductive bias can facilitate better scaling, so incorporating strong inductive biases might limit HRVMamba's scalability and performance on larger-scale models.\n- Concern for `Limitation 2`: The paper uses Deformable Convolutions (DCN) to mitigate the long-range forgetting issue observed in previous visual Mamba models. However, there is a concern about whether DCN can effectively address this problem as the sequence length scales up. The efficacy of DCN for maintaining high-level feature relationships over significantly longer sequences remains uncertain, raising questions about its robustness as a scalable solution for long-range dependencies.\n- The paper references preprints and arXiv versions of significant works, such as Mamba (COLM), Vision Mamba (ICML), and VMamba (NeurIPS). The authors should update these citations to their final published versions to reflect the current state of the literature." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The proposed HRVMamba add a multi-scale DW block and I'm concerned how about the performance of dropping the FFN block." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper presents the Dynamic Visual State Space (DVSS) block, which combines multi-scale convolutional kernels and deformable convolutions.\n2. This paper proposes the High-Resolution Visual State Space Model (HRVMamba) based on the DVSS block and the architecture of HRNet.\n3. The proposed HRVMamba obtains improvements compared to previous approaches on several dense prediction masks" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces HRVMamba, a High-Resolution Visual State Space Model designed for dense prediction tasks such as human pose estimation and semantic segmentation. HRVMamba addresses the limitations of previous Mamba models by incorporating a Dynamic Visual State Space (DVSS) block, which uses multi-scale convolutional kernels to enhance inductive bias and deformable convolutions to mitigate long-range forgetting. The model is based on a multi-resolution parallel design, preserving high-resolution representations throughout the network to facilitate effective multi-scale feature learning. Extensive experiments demonstrate HRVMamba's competitive performance against existing CNN, ViT, and SSM benchmark models on various dense prediction tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper's novelty is limited. It incorporates the VMamba block[1] into the HRNet architecture[2], including DCNv4, making it lack sufficient novelty and insights. In addition, similar ideas have been explored in HRFormer[3,4]. The novelty of this paper is below the threshold for ICLR publication.\n2. The motivations for using DCNv4 and DVSS block in this paper are unclear. For example, this paper lacks a specific analysis of the long-range forgetting problem of Vmamba in vision and why DCNv4 can solve such a long-range forgetting problem of mamba. Whether from the theoretical or experimental analysis perspective, the authors need to provide exact evidence to present.\n3. Lack of comparisons with recent vision mamba works, such as MambaVision[5].\n4. In Fig.1, why are there neat blocks in the activation map? How can this be explained? Is it related to the scans in different directions of VMamba? Image activation is usually continuous, and I'm confused about it.\n5. How do the feature maps of different resolutions fuse, downsample, or upsample?\n6. How about the inference latency of the proposed HRVMamba, which includes VMamba blocks, multi-scale depthwise convolution blocks, and DCNv4 blocks?\n7. In Tab.7, adding a 3x3 convolution has no effect. However, adding larger depth-wise convolutions, such as 5x5, 7x7, or 9x9, improves a little (0.3 AP), but this also introduces many additional parameters. It's unclear here whether the effect comes from extra parameters, larger convolutions, larger receptive fields, or multi-scale convolutions.\n8. The performances of baseline methods (such as HRFormer) on Cityscapes and PASCAL Context are too low, which are far inconsistent with the original paper[3], for example, HRFormer-B obtains 82.6 mIoU (Cityscapes) and 58.5 mIoU (PASCAL Context) while achieving 77.3 mIoU (Cityscapes) and 42.6 mIoU (PASCAL Context) in this paper. I think a fair comparison is very important. However, the results of the current comparison methods are obviously much lower than those of the original methods. \n\nReferences\\\n[1] Liu et al. VMamba: Visual State Space Model. NeurIPS 2024.\\\n[2] Wang et al. Deep High-Resolution Representation Learning for Visual Recognition. TPAMI 2020.\\\n[3] Yuan et al. HRFormer: High-Resolution Transformer for Dense Prediction. NeurIPS 2021.\\\n[4] Gu et al. Multi-Scale High-Resolution Vision Transformer for Semantic Segmentation. CVPR 2022.\\\n[5] Hatamizadeh et al. MambaVision: A Hybrid Mamba-Transformer Vision Backbone. NeurIPS 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Why does the performance of HRFormer-B on the segmentation datasets significantly lag behind the reported results? For example, HRFormer-B + OCR achieves a mIoU of 82.6 on Cityscapes and 58.5 on PASCAL-Context datasets, respectively. However, the performance drops to 77.3 and 42.6 in Table 5, Lines 349-350." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper first introduces Vision Mamba to a multi-resolution branch architecture. Additionally, several techniques are introduced to alleviate the limitations of Vision Mamba, including deformable convolution and multi-kernel size convolution. These techniques provide improvements over the vanilla Vision Mamba baseline. The proposed methods are easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper summarizes the current challenges encountered when applying vision mamba for dense prediction tasks, including insufficient inductive bias, long-range forgetting, and low-resolution output representation. Subsequently, the authors propose corresponding solutions, including multi-scale convolution kernels, deformable convolution, and the HRNet framework to alleviate these issues." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper feels more like a technical report instead of an academic paper. The challenges of Vision Mamba models are obtained from previous works, and the solutions are also based on current techniques. Besides, some observations and methods are not properly cited. For example, the problem where multi-way scanning approach disrupts 2D spatial dependencies has been proposed in MambaIR[1], and the solution of Multi-scale Depthwise block has also been introduced in SMT[2]. Given these considerations, this paper lacks introducing any new insights or techniques to the community.\n\n2. The experiments are not valid enough. For example, for the semantic segmentation, only results on Cityscapes and PASCAL Context are reported. The results on the widely used ADE20K are missing. The authors may consider reporting results with the same framework like Uppernet on ADE20K and compare with publicly available results like VMamba. Complementing these results and comparing with publicly available results will make this paper more solid.\n\n3. Some experimental data lack further explanation. In Table 7, the authors report the performance on the COCO val set. As the COCO dataset has many benchmarks, this \"val\" set is ambiguous. If it refers to the pose estimation, the best results here are 74.2, which is not coherent with the results in Table 3. Additional explanation for the setting differences is needed. Besides, there are some very weird data such as the one in L412. The authors mentioned the poor performance of LocalVMamba on the PASCAL-Context dataset, but no further explanation for the reason is provided.\n\n4. The detailed efficiency comparisons are missing, including the inference speed and memory cost on different datasets.\n\n5. Some minor errors: In L213, Fig.3 refers to Fig.2 by mistake. Besides, the contents from L445 to L450 are not appropriate for the ablation section.\n\n6. This work does not give any discussion about the limitations.\n\nReference:\n\n[1]. Guo, Hang, et al. \"MambaIR: A Simple Baseline for Image Restoration with State-Space Model.\" arXiv e-prints (2024): arXiv-2402.\n\n[2]. Lin, Weifeng, et al. \"Scale-aware modulation meet transformer.\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024hrvmamba,\ntitle={{HRVM}amba: High-Resolution Visual State Space Model for Dense Prediction},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4UxXe3JZta},\nnote={under review}\n}" }, "abstract": { "value": "Recently, State Space Models (SSMs) with efficient hardware-aware designs, \\ie, Mamba, have demonstrated significant potential in computer vision tasks due to their linear computational complexity with respect to token length and their global receptive field. However, Mamba's performance on dense prediction tasks, including human pose estimation and semantic segmentation, has been constrained by three key challenges: insufficient inductive bias, long-range forgetting, and low-resolution output representation.\nTo address these challenges, we introduce the Dynamic Visual State Space (DVSS) block, which utilizes multi-scale convolutional kernels to extract local features across different scales and enhance inductive bias, and employs deformable convolution to mitigate the long-range forgetting problem while enabling adaptive spatial aggregation based on input and task-specific information. By leveraging the multi-resolution parallel design proposed in HRNet, we introduce High-Resolution Visual State Space Model (HRVMamba) based on the DVSS block, which preserves high-resolution representations throughout the entire process while promoting effective multi-scale feature learning.\nExtensive experiments highlight HRVMamba's impressive performance on dense prediction tasks, achieving competitive results against existing benchmark models without bells and whistles.\nWe will make the source code publicly accessible." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Mamba", "Dense Prediction", "Human pose estimation", "Semantic segmentation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/43b8895c02f64825cd898821c60038dc64c54e1f.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "HRVMamba: High-Resolution Visual State Space Model for Dense Prediction" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "None" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Could STRAP be combined with fine-tuning of MT policies on the retrieved dataset to potentially achieve even better performance than domain-specific fine-tuning alone?\n\n- How does STRAP's performance compare against standard fine-tuning approaches when controlling for the total amount of data used?\n\n- What are the memory and computational requirements for deploying STRAP on very large trajectory datasets like Droid?\n\n- Can you add the average performance for LIBERO-10 results to the main table?\n\n- Can you provide a few more real world tasks with required baselines?\n\n- Real world retrieveal is conducted with the same robot embodiment and gripper. How does STRAP Perform when retrieving similar data from other robot datasets like BridgeV2 that does not share the same robot and scenes?\n\nSTRAP presents an novel and well designed approach to few-shot learning, that tackles several drawbacks of prior methods through sub-trajectory retrieval with dynamic time-warping. However, more comprehensive comparisons against fine-tuned baselines and clearer analysis of computational requirements would strengthen the paper's contributions. Thus, I recommend weak reject pending addressing the following concerns: (1) comparisons against fine-tuned baselines, (2) clearer analysis of computational requirements, and (3) better justification of parameter choices." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The proposed method is well motivated and achieves strong results across multiple experiments, both in simulation and real-world settings\n\n- The use of Dynamic Time Warping for sub-trajectory matching is novel and well-suited for the problem domain\n\n- Comprehensive evaluation against recent retrieval baselines demonstrates the method's effectiveness\n\n- Thorough ablation studies on different pretrained encoders provide valuable insights into architecture choices\n\n- The paper is well written and includes several illustrative figures that enhance the text" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces STRAP, a novel method for trajectory retrieval to find similar sub-trajectory segments in large-scale datasets for efficient policy learning in few-shot settings. The method's key contribution lies in combining pretrained visual encoders with Dynamic Time Warping to encode sub-trajectories of variable lengths for improved retrieval. The proposed method is tested against several retrieval baselines and BC ones on LIBERO-10 simulation and real world pick and place tasks and achieves good performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The baseline comparison against multi-task policy appears weak, as it only uses pretrained weights without fine-tuning. This seems like an artificially weak baseline since fine-tuning is standard practice for all MT-policies. \n\n- The paper's argument that retrieval is more efficient than expensive pretraining needs stronger empirical support, especially given that the robotics community regularly fine-tunes general policies for downstream tasks\n\n- The computational cost of STRAP's retrieval process on large-scale datasets like Droid is not adequately addressed, raising questions about real-world scalability. Some more clarity is necessary here\n\n- The choice of K for constructing D-retrieval lacks sufficient explanation and ablations. The paper should explore how different K values affect both retrieval quality and computational overhead and policy performance, as this parameter likely presents a trade-off between performance and efficiency. A discussion about the retrieved data quantity would provide valuable insights and strengthen then paper. \n\n- Small number of tested tasks in real world setting and missing baselines of MT policy and finetuned MT policy" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Baseline: how does STRAP compare to the pretrain-then-finetune setup? (Pretrain on the prior dataset, then fine-tune on the few-shot target demonstrations?)\n- Baseline: how does STRAP compare to a multitask policy trained on all available data?\n- Generalization: to what extent does STRAP generalize across environments or embodiments?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Sub-trajectory retrieval for the few-shot demo behavioral cloning setting is a well-motivated and novel idea.\n- The method is clear and straightforward to implement.\n- Results show that matching with DTW on vision foundation model features are robust to variations and capture task semantics.\n- Real and simulated environments show that STRAP outperforms other retrieval methods and pure behavioral cloning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the setting of generalizing a policy to an unseen task with few-shot demonstrations. Instead of deploying zero-shot, the paper proposes STRAP, training a model on task-relevant data augmented by retrieval. STRAP first retrieves sub-trajectories from a large pretraining dataset that are similar to the new task demonstrations, then combines them with the few-shot demos to train a policy. Results on sim and real environments show that STRAP outperforms other retrieval methods and pure behavioral cloning. Ablations show that STRAP is compatible with various vision encoders and justify each of its component." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- STRAP requires few-shot demos and model training at test time for a new task.\n- It would be good to see more sim and real environments for evaluations.\n- It would be more convincing to see a behavioral cloning baseline that uses all available data." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Compared to the prior framework (Behavior Retrieval, Du et al., 2023), STRAP seems to have three main differences in retrieval system. (a) use non-parametric retrieve vs. VAE (b) use sub-trajectory wise retrieve vs. single state-action pairs and (c) use DTW. Among these, what gives the most / least performance gains?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- To deal with potentially variable length during retrieval, STRAP use dynamic time warping (DTW) to match the sub-sequences\n- STRAP shows improved performance compared to the prior framework (Behavior Retrieval, Du et al., 2023) which retrieves single state-action pairs using VAE." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a task-specific robot learning framework using pre-collected datasets. Unlike the many robot learning methods that train the generalist policy model with multi-task expert data, the proposed method (STRAP) train a task-specific policies, which can yields better performance on single task. When the few-shot target demo, in addition to prior dataset is given, STRAP filters the task-relevant data from prior data and use it with target demo to train the model. One of the key features of STRAP is that it retrieves the data measuring the similarity between sub-trajectories, rather than the whole trajectories. Also, it utilize subsequence dynamic time warping (S-DTW) to match between the data. As a result, the proposed method shows improved performance compared to the previous methods, generalist policy models, and specialist models that only use the target data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The idea of using only data relevant to the target task, rather than learning a generalist policy through multi-task data, is interesting. However, retrieving new data from prior dataset and training a policy each time a new scene is encountered is highly computationally costly.\n- Comparing the entire prior data with the target data one-to-one to measure similarity is not scalable with the dataset size. Moreover, since this retrieval process requires computationally intensive neural network operations, such as DINO, it raises questions about whether this process can be performed at test time. In particular, there is no mention of how to handle an increase in offline dataset size, nor are there any discussions about limitations in this regard.\n- There is no discussion about computational cost.\n- STRAP uses a top-k retrieval dataset. Increasing this k could bring in more data but might reduce relevance, whereas a smaller k would provide more refined data but with a smaller amount. However, there is a lack of analysis on how changing this k value affects performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. Would you please update the sites in the paper?\n2. Could you please add some experiments with other imitation learning methods?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. **Innovative Approach**: The authors present a compelling intuition, demonstrating robustness in solving multitask generalization challenges.\n2. **Efficient Data Usage**: The method shows improvement in the way data is leveraged for robotics tasks, particularly in sub-trajectory retrieval.\n3. **Thorough Experiments**: The experiments are detailed and show promising results for sub-trajectory retrieval and policy creation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "## Paper Review Summary\n\nThis work advocates for training policies dynamically during deployment, utilizing the encountered scenarios to improve model performance. Instead of relying on pre-trained policies to tackle new problems in a zero-shot fashion, the authors propose a non-parametric approach that retrieves relevant data and trains models directly at test time. The paper introduces SRTAP, a method built on pre-trained Vision-Language Models (VLM) and dynamic time wrapping, which combines sub-trajectories into a policy. The approach involves some training on a set similar in language to the test set, and has been demonstrated in both real-world and simulated environments.\n\n### Strengths:\n1. **Innovative Approach**: The authors present a compelling intuition, demonstrating robustness in solving multitask generalization challenges.\n2. **Efficient Data Usage**: The method shows improvement in the way data is leveraged for robotics tasks, particularly in sub-trajectory retrieval.\n3. **Thorough Experiments**: The experiments are detailed and show promising results for sub-trajectory retrieval and policy creation.\n\n### Weaknesses:\n1. **Project Incompleteness**: There is no accessible website or supplementary information, suggesting the project might still be unfinished.\n2. **Visual Readability**: The images in the paper are difficult to interpret, potentially detracting from the clarity of the results.\n3. **Writing Quality**: The paper's writing needs improvement, especially in terms of clarity and readability.\n4. **Generalization**: Some imitation learning methods, such as [Sparse Diffusion Policy](https://forrest-110.github.io/sparse_diffusion_policy/), [HPT](https://liruiw.github.io/hpt/), [RDT-Robotics](https://rdt-robotics.github.io/rdt-robotics), and [Humanoid Manipulation](https://humanoid-manipulation.github.io/), appear to show more generalization in similar settings.\n5. **Formatting**: The paper's formatting is problematic, with certain sections being hard to read, affecting the overall readability of the work.\n\nIn conclusion, while the proposed method shows strong intuition and detailed experimentation, there are concerns about project completeness, readability, and potential improvements in both writing and generalization when compared to existing work. \n\nI would like to change the rate after discussion, but at least you should finish the site you provide." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Project Incompleteness**: There is no accessible website or supplementary information, suggesting the project might still be unfinished.\n2. **Visual Readability**: The images in the paper are difficult to interpret, potentially detracting from the clarity of the results.\n3. **Writing Quality**: The paper's writing needs improvement, especially in terms of clarity and readability.\n4. **Generalization**: Some imitation learning methods, such as [Sparse Diffusion Policy](https://forrest-110.github.io/sparse_diffusion_policy/), [HPT](https://liruiw.github.io/hpt/), [RDT-Robotics](https://rdt-robotics.github.io/rdt-robotics), and [Humanoid Manipulation](https://humanoid-manipulation.github.io/), appear to show more generalization in similar settings.\n5. **Formatting**: The paper's formatting is problematic, with certain sections being hard to read, affecting the overall readability of the work." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Subsequence-DTW for sub-trajectory retrieval to augment few-shot policy learning" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024strap,\ntitle={{STRAP}: Robot Sub-Trajectory Retrieval for Augmented Policy Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4VHiptx7xe},\nnote={under review}\n}" }, "abstract": { "value": "Robot learning is witnessing a significant increase in the size, diversity, and complexity of pre-collected datasets, mirroring trends in domains such as natural language processing and computer vision. Many robot learning methods treat such datasets as multi-task expert data and learn a multi-task, generalist policy by training broadly across them. Notably, while these generalist policies can improve the average performance across many tasks, the performance of generalist policies on any one task is often suboptimal due to negative transfer between partitions of the data, compared to task-specific specialist policies. In this work, we argue for the paradigm of training policies during deployment given the scenarios they encounter: rather than deploying pre-trained policies to unseen problems in a zero-shot manner, we non-parametrically retrieve and train models directly on relevant data at test time. Furthermore, we show that many robotics tasks share considerable amounts of low-level behaviors and that retrieval at the \"sub\"-trajectory granularity enables significantly improved data utilization, generalization, and robustness in adapting policies to novel problems. In contrast, existing full-trajectory retrieval methods tend to underutilize the data and miss out on shared cross-task content. This work proposes STRAP, a technique for leveraging pre-trained vision foundation models and dynamic time warping to retrieve sub-sequences of trajectories from large training corpora in a robust fashion. STRAP outperforms both prior retrieval algorithms and multi-task learning methods in simulated and real experiments, showing the ability to scale to much larger offline datasets in the real world as well as the ability to learn robust control policies with just a handful of real-world demonstrations." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "dynamic time warping", "few-shot imitation learning", "retrieval", "foundation models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/624af64bda2bf4edc4459d3c8fc251d06f58b723.pdf" }, "presentation": null, "primary_area": { "value": "applications to robotics, autonomy, planning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "STRAP: Robot Sub-Trajectory Retrieval for Augmented Policy Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "It is recommended to include the full term \"Online Adapter (OA)\" the first time OA is mentioned in the abstract." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Model flexibility: Accepts both monocular and multi-view data and supports any combination of various prompts (object, box, point, and novel visual prompts).\n2. Clarity of writing: The paper is well-written, logically structured, and easy to read.\n3. Extensive experiments: The main text and supplementary materials provide ample experiments to validate the effectiveness of TTC.\nPractical feasibility: The authors explain real-world application scenarios, achieving immediate error rectification through user-friendly prompts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce the Test-time Correction (TTC) system, an innovative online 3D detection framework designed for correcting test-time errors in real-time through human feedback. TTC demonstrates the capability for immediate error rectification. Extensive experiments show substantial improvements in real-time error correction over pre-trained 3D detectors, even in challenging scenarios involving limited labels, zero-shot detection, and adverse conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. OA is one of the core modules of the model, serving as a bridge between prompts and offline-trained 3D detectors. However, the explanation of OA in the method section is somewhat abstract; adding simple illustrative diagrams could aid understanding.\n2. In the Related Work section, the Online 3D Detection System subsection discusses online 3D detectors. Expanding on offline 3D detectors would help readers better understand the development of offline versus online 3D detection.\n3. There are some minor typos in the text that need correction." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. The paper does not clearly present how the proposed modules are trained, especially for the two layers MLP in key online adapter module." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The designed validation experiments are comprehensive.\n2. The proposed method is training-free, which makes it can be broadly applied." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a 3D object adaptation method that could adapt the human feedback online. The system could work with various visual prompts including reference images, box in the image and click in the image. The proposed methods is validated on both in domain and out of domain dataset, and demonstrating its effectiveness in these scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed can only handle missing objects. It seems it cannot reduce the FP in test time.\n2. Although the authors explain the differences between proposed TTC and single object tracking, the explanation is unconvincing. The visual and object prompt can be easily used in SOT setting. Such discussion and at least comparisons with bbox annotations are must. This is the key concern in my evaluation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How are the visual prompts used during testing obtained？I'm not sure if it's just adding the corresponding areas of undetected targets to the visual cue buffer. \n2. Could the TTC method be combined with LLM-based 3D detection approaches to enhance generalization for novel object categories and domain shifts?\n3. How does the TTC method handle potential noise and latency in user feedback?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Clear and well-structured writing, and easy to understand.\n2. Significant performance improvements across multiple 3D detectors and comprehensive ablation studies validate module effectiveness.\n3. The OA module is simple but effective." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a Test-time Correction (TTC) method that leverages human feedback to correct errors in real-time during testing. The core component, the Online Adapter (OA) module, enables existing 3D detectors to use visual prompts for continuously detecting previously undetected 3D objects." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Re-entering the regions of undetected targets as visual prompts introduces human knowledge, which may lead to potential biases and affect the fairness of comparative experiments.\n2. The TTC method needs to maintain a buffer of visual cues and solve the matching problem between cues and target objects, which increases the complexity.\n3. The experimental section lacks a description of how the visual prompts used during testing are obtained." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "\"To achieve such TTC system, we equip existing 3D detectors with OA module, an online adapter with prompt-driven design for online correction.\" However, the acronym \"OA\" is not defined in the abstract." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The broader impact of this paper may lie in inspiring the research community to further investigate the online rectification approach in autonomous driving systems. This crucial technology has the potential to significantly enhance the safety and reliability of safety-critical applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the Test-time Correction (TTC) system, a novel online 3D detection framework designed to correct test-time errors through human feedback. This approach aims to enhance the safety of deployed autonomous driving systems.\n\nThe key idea of the TTC system is to improve existing 3D detectors with the Online Adapter (OA) module, a prompt-driven design that facilitates real-time error correction. Central to the OA module are visual prompts—images of missed objects of interest that guide corresponding detection and subsequent tracking. These visual prompts are stored in a visual prompt buffer to enable continuous error correction in subsequent frames. This approach allows the TTC system to consistently detect missed objects in real-time, thereby effectively reducing driving risks.\n\nExperimental results show that the proposed method, through test-time rectification, enhances the performance of offline monocular detectors (Zhang et al., 2022a), multi-view detectors (Wang et al., 2023c), and BEV detectors (Yang et al., 2023a) without the need for additional training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper lacks self-containment. For example, in lines 217-231, where the authors describe various visual prompts, they mainly reference numerous other methods without offering sufficient detail. This heavy reliance on external sources renders the paper somewhat incremental, as it fails to clearly articulate the novel contributions and context of the visual prompts within the proposed framework. Furthermore, this lack of clarity results in the use of many notations, such as \"visual features\" and \"image features,\" without providing clear definitions.\n\n\nRather than referring to it as \"visual prompts,\" the pipeline developed in this paper essentially provides a template containing location and size information in the buffer, enabling generic tracking during test time without any additional training. Therefore, the authors are encouraged to clarify whether this pipeline fundamentally differs from a single-object tracker. Additionally, it would be beneficial to include an experiment comparing state-of-the-art (SOTA) trackers for test-time correction as part of the evaluation" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper is well written, and easy to understand.\n\n2. Enhancing 3D detection is important task for autonomous driving.\n\n3. The performance gain is impressive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the Test-time Correction (TTC) system, an online 3D detection framework designed to correct test-time errors through real-time human feedback. Unlike conventional offline static 3D detectors, TTC aims to learn real-time error rectification by incorporating user feedback (e.g., clicks or bounding boxes). This approach allows for immediate updates to detection results for subsequent streaming inputs, even when the model operates with fixed parameters. The TTC system is achieved by integrating an OA module, an online adapter with a prompt-driven architecture, into existing 3D detectors for real-time correction. The key is visual prompts, specifically images of missed objects, which guide both current detection and future tracking. These visual prompts, representing objects missed during inference, are stored in a buffer to support continuous error correction across frames. Extensive experiments reveal substantial improvements in immediate error rectification compared to pre-trained 3D detectors, even under limited labeling, zero-shot detection, and challenging environmental conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Motivation\n\nPretrained static 3D detection modules have clear limitations due to issues like domain gaps, and I agree with the goal of improving these. However, I find it difficult to fully empathize with the motivation behind TTC. If issues in the predictions of the 3D detection module indeed pose a significant impact on safety, as stated, would there realistically be an opportunity to perform online correction in such scenarios?\n\nAdditionally, I am skeptical about the feasibility of interventions like visual prompting during driving. Operating devices such as navigation systems manually while driving is likely a legal violation in most countries, and in practice, the difficulty level for performing such tasks during driving seems exceedingly high.\n\n2. Comparison with TTA or TTT\n\nIn this field, there are various approaches for online improvement of pre-trained static models, such as test-time adaptation (TTA) and test-time training (TTT). Notably, most of these methods function without the need for human feedback. A thorough methodological and performance comparison with these approaches is essential. Additionally, while TTT may be somewhat challenging, in the case of TTA, it seems feasible to utilize human feedback as a direct learning guidance. I would appreciate a more in-depth discussion on this aspect as well.\n\n3. Robustness\n\nIt is unrealistic to expect that user corrections will always be accurate. Depending on the situation, incorrect user interventions could potentially worsen the proposed TTC. It would be beneficial to model the noise that might exist in visual prompting and demonstrate that TTC can operate robustly even when this noise becomes more pronounced." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Please refer to the Paper Weaknesses mentioned above.\n2. The experimental section only conducts multiple tasks on MonoDETR; however, multi-view 3D detection is currently the mainstream approach in autonomous driving solutions. It is recommended to include experiments on mainstream multi-view 3D detectors across various tasks.\n3. The proposed method focuses solely on correcting missed detections, yet false positives are also a significant issue in autonomous driving. Is there scalability for correcting false positives?\n4. The experimental section lacks comparisons with existing instruction-based 3D detection methods typically utilize text, boxes, or clicks as prompts; it is recommended to include such comparisons." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper is clearly structured and written-well.\n2. This paper focuses on an interesting issue, namely that a online 3D detection system designated for online correction of test-time errors via human feedback." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Test-time Correction (TTC) system, a online 3D detection system designated for online correction of test-time errors via human feedback. The proposed TTC system includes two components: Online Adapter (OA) that enables 3D detectors with visual promotable ability, and a visual prompt buffer that records missing objects. Experiments were conducted on the nuScenes dataset, focusing on the TTC system across various 3D detectors and in out-of-training-distribution scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The rationale behind this task setup requires further discussion. Since the proposed task involves human feedback, this process is often uncontrollable, making it challenging to ensure real-time performance. This limitation affects the feasibility of applying the task in real-world autonomous driving scenarios.\n2. The rationale behind the EDS evaluation metric requires further discussion. Classification accuracy is also crucial for autonomous driving, and focusing solely on localization performance while neglecting classification performance is not realistic.\n3. The proposed method is only applicable to vision-based autonomous driving solutions, limiting its generalizability to LiDAR-based autonomous driving systems." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper introduces Test-time Correction (TTC), a novel online 3D detection system designated for swift correction of test-time errors via human feedback to ensure the reliance on safety-critical autonomous driving systems." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024testtime,\ntitle={Test-time Correction with Human Feedback: An Online 3D Detection System via Visual Prompting},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4VNfufHtoS},\nnote={under review}\n}" }, "abstract": { "value": "This paper introduces Test-time Correction (TTC) system, a novel online 3D detection system designated for online correction of test-time errors via human feedback, to guarantee the safety of deployed autonomous driving systems. Unlike well studied offline 3D detectors frozen at inference, TTC explores the capability of instant online error rectification. By leveraging user feedback with interactive prompts at a frame, e.g., a simple click or draw of boxes, TTC could immediately update the corresponding detection results for future streaming inputs, even though the model is deployed with fixed parameters. This enables autonomous driving systems to adapt to new scenarios flexibly and decrease deployment risks reliably without additional expensive training. To achieve such TTC system, we equip existing 3D detectors with OA module, an online adapter with prompt-driven design for online correction. At the core of OA module are visual prompts, images of missed object-of-interest for guiding the corresponding detection and subsequent tracking. Those visual prompts, belonging to missed objects through online inference, are maintained by the visual prompt buffer for continuous error correction in subsequent frames. By doing so, TTC consistently detects online missed objects and immediately lowers down driving risks. It achieves reliable, versatile, and adaptive driving autonomy. Extensive experiments demonstrate significant gain on instant error rectification over pre-trained 3D detectors, even in challenging scenarios with limited labels, zero-shot detection, and adverse conditions. We hope this work would inspire the community to investigate online rectification systems for autonomous driving post-deployment. Code would be publicly shared." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Autonomous Driving", "3D Object Detection", "Test-time Error Correction" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/85f7cb7993f0642feaaff0c7af8c5c4327e9d3aa.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Test-time Correction with Human Feedback: An Online 3D Detection System via Visual Prompting" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I have no further questions beyond those outlined in the weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The identification of scale shift as a specific domain shift challenge in crowd localization, and the introduction of Scale Shift Domain Generalization, bring attention to an under-explored issue with significant real-world implications. ScaleBench provides a standardized benchmark, adding practical value for the research community.\n2. The authors provide a clear theoretical explanation linking scale shift to diversity and correlation shifts, elucidating why DG models struggle with this issue. This rigorous analysis adds depth to the understanding of scale shift and its implications for DG.\n3. The paper is well-organized, with each section following logically from the last. The clear delineation between problem identification, analysis, and solution makes the contributions easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a significant contribution to domain generalization (DG) for crowd localization by addressing the challenge of scale shift, where differences in head size distributions between training and testing data impact model performance. The authors introduce ScaleBench which categorizes datasets based on scale distributions. They also propose Semantic Hook, an algorithm designed to mitigate scale shift by reinforcing the association between semantic features and task predictions. Through testing 20 state-of-the-art DG algorithms on ScaleBench and conducting theoretical analysis, the authors highlight the limitations of current approaches and introduce Scale Shift Domain Generalization as a novel research direction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The introduction lacks highlighting core contributions and findings.\n2. Although the authors indicate that the paper does not primarily focus on introducing a new method, the experiments in the main text feel somewhat limited. \n3. The appendix contains several formatting issues, particularly with tables. Inconsistencies include varying font sizes, tables floating in the middle of pages, and some tables exceeding the page width. These layout problems affect readability and detract from the paper's presentation quality." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Explain how Semantic Hook handles the scale shift on domain generalization. From the given formulation, the semantic difference will highlight the effect of the perturbation, and the decoder is now learning to map noise to task-specific outcomes. So, how does Semantic Hook reduce generalization risk?\n2. The conditional probability derived in Eq. 6 is incorrect at the first integral. The conditional probability P(y|x) does not equal integrating P(y|z) over the domain of Z. Please provide the correct derivation.\n3. The scale shift is more prevalent in crowd images under perspective projection; however, the scale is more uniform throughout the scene for aerial views. How does the proposed method handle different projections?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper addresses an under-explored issue of scale shift in domain generalization for crowd localization. In terms of contributions, the paper delivers manually annotated bounding boxes for crowd localization on existing public crowd benchmarks. The paper is well-structured and provides a good analysis of the problem, resulting in a novel solution method called Semantic Hook. Further, the authors take an analytical route for the scale shift under domain generalization connecting other attributes present in datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates the impact of scale shift on domain generalization in crowd localization. Models experience performance degradation due to head scale distribution shifts in training and testing datasets. To address this, the authors provide a theoretical analysis of the scale shift under domain generalization and introduce a novel method to mitigate the effect of scale shift, called Semantic Hook. The paper proposes a new benchmark called ScaleBench and provides bounding box annotations for existing public crowd benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper needs more detailed explanations regarding how Semantic Hook mitigates the scale shift in domain generalization. It also needs to clarify which variables or attributes are being generalized from the perturbation added during training. Additionally, the improvements from the proposed method on the ScaleBench benchmark are marginal compared to the baseline method. Furthermore, the mathematical formulations (Eq. 6) used for the theoretical analysis need to be corrected." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. What does global feature mean in Table 5? Figure 2 shows that semantic features are extracted from the encoder. How to extract global feature?\n2. Is the proposed method sensitive to the choice of gamma?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. A controlled benchmark is established to study scale variance in crowd localization.\n2. This paper proposes SemanticHook to handle scale shift.\n3. Comprehensive analyses are presented to quantify the influence of scale shift." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to study the effect of scale shift in crowd localization. To this end, a benchmark, dubbed ScaleBench, is first established to quantify the influence of scale shift. Next, SemanticHook is proposed to tackle scale shift. The key idea is to enhance the association between semantic features and targets by perturbing the input image with Gaussian noise. Empirical analyses on ScaleBench justify the effect of scale shift." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The rationality of ScaleBench is questionable. First, while the perspective effect often occurs in crowd localization, there exist images that are captured from different angles (e.g., top view). In such scenarios, image distribution regularization may fail to partition the images correctly. Second, in the real world, scale shift is often coupled with other factors, such as occlusion, weather, and appearance. For example, when the object suffers from significant appearance variations, the counting model may fail to localize objects even if training and testing data yield the same scale distribution. Third, dividing images into patches will inevitably result in incomplete objects, which could affect the localization results. Therefore, evaluations on ScaleBench may not rigorously reflect the influence of scale shift. \n2. The proposed SemanticHook does not exhibit superiority over existing methods. As shown in Table 1, the simplest baseline ERM already achieves good results. The proposed method is not necessarily better than ERM.\n3. Following the previous comment, the rationale of SemanticHook is not entirely convincing. Eq. 6 suggests that p(s, c, …) can lead to a spurious association between the output y and scale c. This term is a joint distribution of semantic s and scale c. However, the authors merely try to enhance the semantic association between semantic s and output y. Experimental results demonstrate that such a technique does not address scale shift effectively. Additionally, perturbing image is not a new idea, which is widely used in adversarial attack.\n4. It appears that the influence of image interpolation is not rigorously quantified in Table 4. First, the implementation of Random Augmentation shall be modified according to different domains, i.e., the range of random scaling should be customized based on domain Tiny, Small, and Normal. Second, it is necessary to train the model using different source domains to identify the effect of image interpolation. The results on domain Big are insufficient to conclude that the benefits of image interpolation are modest.\n5. Regarding training details. In practice, random scaling is commonly used to alleviate scale variations. As the authors use this technique to train the model, the reported results may not correctly reveal the effect of scale shift, because random scaling already simulates different scales.\n6. The paper lacks evaluations on previous methods featuring multi-scale architecture, e.g., STEER. Evaluations on these methods are helpful in revealing whether previous methods can handle scale variations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1) How does \"scale shift\" in crowd localization fundamentally differ from other types of domain shifts?\n2) How does Semantic Hook compare with simpler baseline methods, such as multi-scale training or augmentations?\n3) Can the spurious association between scale and the output be quantified?\n4) How would ScaleBench and Semantic Hook perform in real-world crowd localization scenarios with continuous scale distributions?\n5) What are the limitations of ScaleBench in generalizing to diverse crowd analysis tasks?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) The development of ScaleBench is a major contribution, offering a curated dataset specifically designed to study scale shift effects on domain generalization.\n2) The paper introduces the Semantic Hook as a novel approach to reduce the impact of scale shift in domain generalization tasks.\n3) The paper is well-structured and logically organized. Offering theoretical insights and comprehensive empirical evaluations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on the impact of scale variations on crowd localization models' ability to generalize across datasets. To tackle this, the authors introduce ScaleBench, a new benchmark dataset specifically curated to study scale shift, and evaluate 20 existing domain generalization algorithms, showing that many struggle with this type of shift. They also propose an approach, Semantic Hook, aimed at mitigating scale shift by strengthening the association between semantic features and predictions, rather than relying on scale information. While the improvements are modest, the paper offers valuable insights into scale-based generalization challenges." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The paper frames the issue of scale shift as a new challenge within domain generalization for crowd localization, but this framing seems overstated. Scale shift, where different head sizes (scales) impact model performance across datasets, is not fundamentally new. Previous works have already explored the impact of scale variation on domain adaptation in crowd analysis (albeit under different terminologies), suggesting that this issue is more a subset of a well-studied generalization problem rather than a novel concept. Claiming it as the \"first study\" on \"scale shift domain generalization\" could be seen as an attempt to rebrand existing challenges without sufficient justification.\n\n2) The proposed \"Semantic Hook\" technique to mitigate scale shift claims to enhance the association between semantic features and task predictions, but its practical effectiveness remains questionable. This method involves adding Gaussian noise to \"hook\" relevant features, yet the theoretical rationale behind this approach is underdeveloped. How \"Semantic Hook\" contributes to decoupling scale-related biases from semantic content is unclear. Additionally, the improvement in F1 scores presented in Table 2 is marginal, suggesting that the Semantic Hook might not be a robust solution.\n\n3) While the paper provides a comparison of 20 domain generalization algorithms, there is little discussion about the practical differences in their robustness against scale shifts. The Semantic Hook’s performance is only marginally better than ERM, raising doubts about its practical value. Furthermore, the experiments rely heavily on F1 scores across ScaleBench domains but do not include additional evaluation metrics (e.g., precision, recall) that could provide a fuller picture of model performance under scale shift.\n\n4) ScaleBench, with its scale-based domain partitions, may not accurately reflect real-world applications where scale distributions are more complex and continuous rather than discretely defined. The Leave-One-Out approach used for evaluation also artificially simplifies the generalization challenge. Real-world scenarios often involve more nuanced and diverse shifts between training and deployment environments, suggesting that the paper’s evaluation may lack external validity." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1-What is the difference between semantic hook and other methods that also purturb the image for crowd counting/localization?\n\n2- What other methods specifically in crowd counting and/or localization exist that have addressed the scale variance? Have any of these methods been implemetned in this paper?\n\n3- Why does each of the previous methods fail in addressing this issue? What is the authors insight in this matter?\n\n4- How were the hyperparameters for each model set? Did the authors use grid search?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1- The paper is well-written and engaging, and the ideas flow smoothly. \n\n2-The field of crowd counting/localization would benefit from an analytical work focused on the issue of scale variance, as scale shifts present a significant challenge for model generalization across diverse domains. This paper addresses this gap, and provides both a theoretical framework and a practical benchmark." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper analyzes domain generalization under scale shift in crowd localization, where object scales vary across domains. To address the lack of benchmarks for studying scale-related shifts, the authors introduce Scale Bench. This benchmark divides data into domains based on scale and evaluates models on their ability to generalize to unseen scales. They propose Semantic Hook, a training method that uses noise perturbations to reduce scale reliance and strengthen semantic associations in model predictions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1- There is limited mention (with brief explanation for each) of papers explicitly in crowd counting/localization fields that tackle the issue of scale variance.\n\n2- While the paper provides a comprehensive benchmark for scale-related domain generalization, it lacks coverage of crowd counting/localization methods specifically designed for domain generalization. How many of the methods in Table 2 discuss scale variance for crowd counting specifically?\n\n3- The paper does not clarify what has been done to prevent overfitting, particularly given the possible complexity of the models relative to the training data provided. \n\n4- Although Tables 6-18 and a brief discussion for each are included in the appendix, the paper lacks an in-depth analysis explaining why certain methods outperform others in specific cases. A discussion of these results would add valuable context to understand the strengths and limitations of each approach under different scale conditions. what could be the issue that each of these methods fail in generalizing to new domain?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present a ScaleBench to benchmark scale shift domain generalization, and propose rigorous theoretical analysis for this issue, which further motivates an algorithm called SemanticHook to support the following research on this issue." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024understanding,\ntitle={Understanding Scale Shift in Domain Generalization for Crowd Localization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4VfPLTqdrq},\nnote={under review}\n}" }, "abstract": { "value": "Crowd localization plays a crucial role in visual scene understanding towards predicting each pedestrian location in a crowd, thus being applicable to various downstream tasks.\nHowever, existing approaches suffer from significant performance degradation due to differences in head scale distributions (scale shift) between training and testing data, a challenge known as domain generalization (DG). This paper aims to comprehend the nature of scale shift within the context of domain generalization for crowd localization models.\nTo this end, we address three key questions: (i) how to quantify the scale shift influence on DG task, (ii) why does this influence occur, (iii) how to mitigate the influence.\nSpecifically, we first establish a benchmark, ScaleBench, and reproduce 20 advanced DG algorithms, to quantify the influence. \nThrough extensive experiments, we demonstrate the limitations of existing algorithms and highlight the under-explored nature of this issue.\nTo further understand its behind reason, we provide a rigorous theoretical analysis on scale shift. \nBuilding on this analysis, we further propose a simple yet effective algorithm called Semantic Hook to mitigate the influence of scale shift on DG, which also serves as a case study revealing three significant insights for future research. Our results emphasize the importance of this novel and applicable research direction, which we term $\\textit{Scale Shift Domain Generalization}$." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Crowd Localization", "Domain Generalization", "Scale Shift" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/8047c3457bc46a3ef69ecd143d0616e779227bcc.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/7f35bf90ff7898c77c950093b61c8bd4c3180831.zip" }, "title": { "value": "Understanding Scale Shift in Domain Generalization for Crowd Localization" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How many parameters do those two tokenizers have respectively?\n2. What are the ablation study results on other tasks such as property prediction and chemical reaction prediction? \n3. What are the input and output of the molecular property prediction task and other tasks? The performance gain mainly comes from hierarchical graph tokenization, and it has nothing to do with the new tuning dataset, right?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The Hierarchical Graph Tokenization (HIGHT) technique is a major advancement. By incorporating hierarchical structure at multiple levels (node, motif, and graph), the paper addresses a crucial gap in previous molecule-language alignment methods, which typically rely only on node-level information. This hierarchical approach captures the functional groups and structural motifs inherent in molecules, improving the model’s ability to represent complex biochemical properties accurately.\n\n2. The introduction of HiPubChem, an augmented molecular instruction tuning dataset enriched with motif and functional group information, enhances model training by aligning molecular structural details with language descriptions. This contribution is valuable for future work in molecular and biochemical language model alignment.\n\n3. The effectiveness of each of the two methods was verified through simple ablation studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new approach to aligning molecular graph representations with language using a method called Hierarchical Graph Tokenization (HIGHT). Traditional graph-language alignment models primarily focus on node-level information, often neglecting the inherent hierarchical structure of molecules, which leads to alignment issues and hallucination in large language models (LLMs).\n\nThe authors introduce HIGHT, which utilizes a hierarchical graph tokenizer to capture information at the node, motif, and entire molecule levels. This tokenizer incorporates both atom-level and motif-level tokens, which are then used to improve alignment with language models. To address the alignment of hierarchical molecular data with textual descriptions, the authors also develop an enhanced molecular instruction tuning dataset called HiPubChem, which provides detailed motif information." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The introduction of the hierarchical graph tokenizer seems to make the tokenizer larger compared with the ordinary node-level tokenizer. It should be discussed that whether the performance gain comes from the larger tokenizer.\n\n2. There should be more detail descriptions and discussions about the evaluation tasks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Listed in Cons." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "An improved tokenization of molecular graphs that enriches molecule's description with motif tokens." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors study Large Graph Language Models (LGLM). Drawing inspiration from Multimodal LLMs, authors focus on the task of incorporating graph data as a separate modality with a GNN encoder and an adapter. Authors conclude that node-centric tokenization of molecules leads to LLM hallucinations when asked about the presence of specific fragments. To overcome this issue, the authors propose to enrich the molecule's description by adding the tokens corresponding to BRICKS-fragments that are present in the molecule. The experimental results demonstrate that such a tokenization scheme reduces the amount of motif-related hallucinations and improves performance on other tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Specifically, most existing LGLMs directly take the node tokens from GNNs as inputs to LLMs (Cao et al., 2023): \nThe paper cites InstructMol as a previous approach that utilizes node-centric tokenization. However, if I understand correctly, InstructMol takes the embedding of the whole graph along with the SMILES representations of the molecule. Moreover, it is not clear which previous models use the node-centric tokenization and whether there are such models at all.\n\nSection 4.3 describes the fine-tuning approach that involves two stages, where the second stage is the finetuning on MoleculeNet, CheBI-20 and Mol-instructions specialized datasets. In my opinion, this implies that the resulting model is specialized. Please, provide better explanation for specialist and generalist models.\t \t \t \t\t\n\nTaking into consideration that the difference between specialist and generalist models is not clear, the resulting model does not demonstrate performance superior to baselines in most of the experiments.\n\nThere is no comparison with [1] in Table 4. The results in [1] are superior to all the models from Table 4.\n\nIn Table 5, the Mol-instruction has the highest MACCS FTS for the retrosynthesis task. However, a smaller number is balded.\n\nThe comparison on MotifHallu is not complete. Please provide comparison with SMILES-based approaches. Moreover, the improvement on the MotifHally benchmark is expected, as the proposed approach was explicitly designed to better solve this task.\n\n[1] Srinivas, S. S., & Runkana, V. (2024). Crossing New Frontiers: Knowledge-Augmented Large Language Model Prompting for Zero-Shot Text-Based De Novo Molecule Design. arXiv preprint arXiv:2408.11866." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. At line 273, the authors said \"attach positional encodings $p$ to all of the tokens\", How are position encodings of motifs obtained?\n2. If the input graphs use the positional encodings, then, should the original positional encodings in the LLMs be diabled? e.g, the widely used ROPE for the graph input part? \n3. What is the papameter count to be tuned?\n4. Besides the vicuna-v-1.3-7B, can the authors provide experimental resutls for other LLM backbones? Since different backbones may have a big impact on the performance.\n5. How is the proposed model performance for zero-shot or few-shot scenarios?\n6. In table 2, llama 13b has wrose performance than llama 7b on most of datasets. Also, Galactica-120B has a sharp performance drop on BACE. Any explanations on these results?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper proposes to incorporate hierarchical graph information into LGLMs, and the authors achieve this with new architecture and instruction tuning dataset HiPubChem.\n2. To address hallucination issue, the paper creates MotifHallu, the first hallucination benchmark based on the existence of common functional groups.\n3. The paper includes extensive experiments with 14 real-world molecular and reaction comprehension benchmarks. The results show that HIGHT significantly reduces the hallucination on MotifHallu and demonstrates significant improvement on a number of tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes HIerarchical GrapH Tokenization (HIGHT), which tries to improve how LLMs understand and process molecular data. The key idea of HIGHT is to introduce a hierarchical graph tokenizer that extracts and encodes information at multiple levels: atoms, motifs, and the overall molecule. The paper demonstrates that HIGHT can reduce hallucination and improve performance across various molecular tasks, including property prediction, molecular description generation, and chemical reaction prediction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The hierarchical graph tokenization process, which involves the use of multiple adapters, is likely to be more computationally expensive than traditional node-centric tokenization. The paper does not discuss the computational complexity. Also, LLM is tuned using LORA, and the number of parameters tuned should be discussed.\n2. One motivation of applying LLMs for graph data is to utillize the generalization capability of LLMs. However, this paper do not provide experimental results on zero-shot or few-shot scenarios of the proposed model. I think it will greatly strength the paper if HIGHT has good performance under such cases.\n3. The performance of HIGHT will largely depend on the backbone LLMs, and only vicuna-v-1.3-7B is evaulated." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- In tables 3, 4, 5, are all baselines also fine-tuned with the same dataset as HIGHT?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The overall presentation is clear and the paper is easy to follow. The work proposes a complete pipeline to build a model with stronger motif/functional group querying ability. Using motif tokens is a straight-forward solution to enhance such ability. Various experiments are conducted to validate the model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework named HIGHT to align molecular data with LLMs. It identifies a shortcoming of LLM on learning functional groups, and proposes to extend the graph tokenization to motif level. Specifically, its input to the LLM includes node/atom embeddings as well as motif embeddings. The model is fine-tuned with motif prediction tasks on a dataset constructed using RDKit. The model shows good performance on molecule properties prediction compared to language models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- From the novelty and contribution perspective, taking motif representations/tokens is not new. By simply searching on Google, I found several papers that extract motifs for graph modeling [1, 2] (as the author also mentioned in the paper). This work is a simple extension of these techniques to align the motifs to the LLM.\n\n- If I understand correctly, the motif tokenization algorithm, BRICS, will break the molecule in a very chemistry-aligned way. For example, a \"OH\" functional group will be tokenized into a motif. The downstream task of identifying the functional group will be very easy (simply aligning a single motif token with the text description of the function group, and the task is like asking \"does a -OH motif have -OH functional group\"). The author should justify how this is a helpful task besides simply telling the LLM that \"there is such a functional group.\" For example, the author should show that the method has better downstream performance than simply telling the LLM the existence of functional groups.\n\n- The distinction between specialist model and generalist model is arbitrary to me. Methods like MolFM and Text+Chem T5-augm-base have the same functionality as the proposal, yet they achieved better performance than HIGHT. I think the HIGHT is more specialized, as it requires explicit and specialized atom and motif tokenization. Can you be more specific about the distinction, and what's the advantage of a generalist model?\n\n- Even without the motif tokens, many models achieved stronger performance. Can you explain why a better motif prediction ability does not lead to better downstream performance? Link back to weakness 1, does this also mean that the proposed task is too easy for the motif tokenization, preventing the model from learning meaningful/molecule-property-related from the pretraining process?\n\n[1] Zhang, Zaixi, et al. \"Motif-based graph self-supervised learning for molecular property prediction.\" Advances in Neural Information Processing Systems 34 (2021): 15870-15882.\n[2] Chen, Xuexin, et al. \"Motif graph neural network.\" IEEE Transactions on Neural Networks and Learning Systems (2023)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present a new strategy that incorporates hierarchical graph information into supervised finetuning and instruction datasets for a better alignment of graph and language modalities." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024improving,\ntitle={Improving Molecule-Language Alignment with Hierarchical Graph Tokenization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4VmagzA2Tp},\nnote={under review}\n}" }, "abstract": { "value": "Recently there has been a surge of interest in extending the success of large language models (LLMs) to graph modality, such as molecules. As LLMs are predominantly trained with 1D text data, most existing approaches adopt a graph neural network to represent a molecule as a series of node tokens and feed these tokens to LLMs for molecule-language alignment. Despite achieving some successes, existing approaches have overlooked the hierarchical structures that are inherent in molecules. Specifically, in molecular graphs, the high-order structural information contains rich semantics of molecular functional groups, which encode crucial biochemical functionalities of the molecules. We establish a simple benchmark showing that neglecting the hierarchical information in graph tokenization will lead to subpar molecule-language alignment and severe hallucination in generated outputs. To address this problem, we propose a novel strategy called HIerarchical GrapH Tokenization (HIGHT). HIGHT employs a hierarchical graph tokenizer that extracts and encodes the hierarchy of node, motif, and graph levels of informative tokens to improve the graph perception of LLMs. HIGHT also adopts an augmented molecule-language supervised fine-tuning dataset, enriched with the hierarchical graph information, to further enhance the molecule-language alignment. Extensive experiments on **14** molecule-centric benchmarks confirm the effectiveness of HIGHT in reducing hallucination by **40%**, as well as significant improvements in various molecule-language downstream tasks." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "molecular-language alignment", "large language models", "hierarchical graph neural networks", "tokenization", "biomolecular studies", "molecule" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/11d66ed0808199b68dea9e77505357d3444ff669.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Improving Molecule-Language Alignment with Hierarchical Graph Tokenization" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See Weakness." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The task of 3D urban generation is important.\n2. The method is reasonable and looks to have better quantitative results than previous models.\n3. The writing is clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a generative urban world model that can automatically create a customized, realistic, and interactive 3D urban world with flexible control conditions. The code of this work was released." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Claim of World Model. This work belongs to the 3D urban generation. It is over-claimed to be a world model and barely related to AGI. Authors should precisely identify the task and topic. Then, focus on the specific topic and make it comprehensive rather than claim some large topics.\n\n2. Technical contributions. The motivation of the generation pipeline is unclear. Why do you need a vision language model? What are the special designs in your work different from others, and why do you need them? What is the special challenges that lead you to design the method? So far, the pipeline looks like a combination of recent advanced techniques, i.e., diffusion model and vision language model.\n\n3. Visual results. The visual results are insufficient. From only a few images, it can not be convinced that the visual quality is better than other models. Also, Figure 6 and 4 have some reduplicate results.\n\n4. Evaluation of interactive environments. The evaluation of interactive environments is coarse. The navigation tasks are not really evaluated. From an image, nothing can be evident. What are the quantitative results, and what are the video results? How do you make the simulation of physics? What is the physics engine? What is the training speed? What is the model, RL or IL? What are the evaluation metrics?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is written fluently and is easy to understand.\n2. The proposed method shows relatively better results in generating city scenes with assets that have new appearances.\n3. The authors effectively showcase various capabilities of the pipeline." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a method for 3D urban scene creation called UrbanWorld, which facilitates customized and interactive 3D urban world generation. UrbanWorld uses Blender to create untextured 3D layouts from 2D maps and incorporates Urban MLLM to generate textual descriptions for assets. A diffusion-based method is then applied to generate and refine the geometry of the 3D assets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the authors state that the method achieves “customized, realistic, and interactive 3D urban world generation,” the results appear more simulation-style and fall short of true realism. The texture quality, as seen in Fig. 3 and 4, is not particularly impressive, and there are no significant improvements over CityDreamer.\n2. The absence of video results is notable. For a 3D generation task, video demonstrations would better illustrate the quality and realism of the generated scenes.\n3. Fig. 4 includes scenes with humans and vehicles, but the method of incorporating these assets is unclear. Details on how these elements are introduced and animated within the scene are missing.\n4. Most visual results focus on limited, local areas. For a city-level generation, it would be beneficial to include bird’s-eye-view results covering larger spatial regions, similar to CityDreamer.\n5. Including a user study comparison would provide a clearer assessment of the visual quality of the generated scenes.\n6. Although the authors claim the ability to create new assets, this appears limited to the level of appearance, with geometry remaining unchanged from the asset library. Given the importance of geometry in 3D generation, this aspect should be addressed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In Figure 3, texture refinement only shows marginal improvement for buildings. Authors should provide more examples, including other objects and views.\n\n2. In Figure 4, the authors shows existence of human and vehicles, how are these generated? Are they also assets generated at some stage? Or done by manually post-processing? It is not mentioned anywhere in the paper, and this indicate the visual quality comparison with other methods is completely unfair.\n\n3.Since the framework generate 3D scenes, I suggest the authors to submit videos or at least multi-views of the same scene to demonstrate quality and view consistency of the generated scenes." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "UrbanWorld introduces a pipeline that integrates generative diffusion models with an urban-specific MLLM to achieve realistic urban scene creation. This combination allows for controlled generation of 3D assets and adaptive urban design." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces UrbanWorld, a generative model designed for the automatic creation of realistic, customizable, and interactive 3D urban environments. UrbanWorld employs a progressive, four-stage generation pipeline: flexible 3D layout creation, Urban Multimodal Large Language Model (Urban MLLM)-based scene design, diffusion-based asset rendering, and MLLM-driven scene refinement." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors claimed section A is flexible urban layout “generation”. However, this is not like generation methods where the distribution of urban layouts are learned from real-world data [1][2][3]. It seems like the authors are just using OSM’s GT data (AIGC-layout is not explained anywhere in the paper). No detail is given on how did the authors transform the OSM data or AIGC data into untextured 3D urban environment. Is there any generation models or other networks involved? In short, if you are just using GT data and Blender add-on to import it, you can’t call the process “generation”.\n\n2. In section 3.2 and the Appendix A.2, the authors shows a general urban generation prompt is converted into prompts for different categories of urban objects. However, the same prompt is generated for all objects of the same class. Doesn’t that indicate they would have exact same style and appearance? For example, if there were 50 buildings in the scene, and they all share the same T2I prompt, they end up looking the same. Meanwhile, the authors introduced descriptions for all categories are generated by an MLLM, but did not explain where does the reference image comes from.\n\n3. For a single asset, the authors generated textures from different views conditioned on the same text and reference image, then merged all textures. This approach cannot guarantee consistency between textures as no 3D condition has been used to guide the 2D diffusion model. Meanwhile, it cannot be called “3D diffusion renderer”, since the authors are only inferencing iteratively from pretrained 2D diffusion models. \n\n[1] Infinicity: https://arxiv.org/abs/2301.09637\n[2] CityDreamer: https://arxiv.org/abs/2309.00610\n[3] CityGen: https://arxiv.org/abs/2312.01508" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the weakness part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-written and easy to follow. \n2. The code has been released. \n3. The overall framework is technically sound. Based on a pre-defined set of common objects in the urban scenario, the framework bridges the gap between the 3D world and 2D views via pre-trained diffusion models. The pipeline is interesting. \n4. The framework achieves controllable and customizable scene generation, which can support tasks that require agent-environment interactions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces UrbanWorld, a generative model designed to automatically create realistic and interactive 3D urban environments, addressing the challenges of manual labor in urban scene design. UrbanWorld employs a progressive diffusion-based rendering method and a specialized urban multimodal large language model (Urban MLLM) trained on street-view image-text data to guide the generation process. The model consists of four key stages: flexible 3D layout generation, urban scene design using Urban MLLM, controllable asset rendering, and scene refinement. Extensive evaluations demonstrate that UrbanWorld achieves state-of-the-art realism and interactivity, outperforming existing methods like Infinicity and CityGen in generating diverse urban environments suitable for embodied agents." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Even though superior quantitative results are reported, the generated images are not realistic enough based on the demonstration in the paper. \n2. It would be better if the authors could provide more diverse qualitative results generated by the proposed method. The proposed system is claimed to be for 3D city generation. It would be good if a sequence of images/video captured with a moving camera is included to show the scene-level generation capability. \n3. I am confused about the UV unwrapping and UV wrapping parts. How can you ensure the wrapping process can align the texture perfectly to the mesh model? For objects of different types and shapes, I believe this process can be hard to model by the diffusion model. The UV unwrapping is usually not unique. Is there any mechanism to enforce the equivariance to different unwrapping manners? \n4. I noticed that the Position-awareTexture Completion module is applied to refine the texture map. Can you provide some qualitative results (visualization) to compare the results before and after the refinement?\n5. The section 4.4 is a little bit vague. How does your generated environment support navigation? How far is the longest distance your navigation can achieve? It could be better to show a bird-eye-view of your navigation environment." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose UrbanWorld, the first generative urban world model that can automatically create a customized, realistic and interactive 3D urban environments." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024urbanworld,\ntitle={UrbanWorld: An Urban World Model for 3D City Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4W1wTg7q9o},\nnote={under review}\n}" }, "abstract": { "value": "Cities, as the essential environment of human life, encompass diverse physical elements such as buildings, roads and vegetation, which continuously interact with dynamic entities like people and vehicles. Crafting realistic, interactive 3D urban environments is essential for nurturing AGI systems and constructing AI agents capable of perceiving, decision-making, and acting like humans in real-world environments. However, creating high-fidelity 3D urban environments usually entails extensive manual labor from designers, involving intricate detailing and representation of complex urban elements. Therefore, accomplishing this automatically remains a longstanding challenge. Toward this problem, we propose UrbanWorld, the first generative urban world model that can automatically create a customized, realistic and interactive 3D urban world with flexible control conditions. Specifically, we design a progressive diffusion-based rendering method to produce 3D urban assets with high-quality textures. Moreover, we propose a specialized urban multimodal large language model (Urban MLLM) trained on realistic street-view image-text corpus to supervise and guide the generation process. UrbanWorld incorporates four key stages in the generation pipeline: flexible 3D layout generation from OSM data or urban layout with semantic and height maps, urban scene design with Urban MLLM, controllable urban asset rendering via progressive 3D diffusion, and MLLM-assisted scene refinement. We conduct extensive quantitative analysis on five visual metrics, demonstrating that UrbanWorld achieves state-of-the-art generation realism. Next, we provide qualitative results about the controllable generation capabilities of UrbanWorld using both textual and image-based prompts. Lastly, we verify the interactive nature of these environments by showcasing the agent perception and navigation within the created environments. We contribute UrbanWorld as an open-source tool available at https://github.com/Urban-World/UrbanWorld." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Urban world model", "3D city generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/8748fc45a71987389d0007bd5025742987995b68.pdf" }, "presentation": null, "primary_area": { "value": "applications to computer vision, audio, language, and other modalities" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "UrbanWorld: An Urban World Model for 3D City Generation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1.Why is the accuracy for the Base model in Table 3 not 0? In my humble opinion, the Acc and Loc should be 0 and 100 respectively, similar to the results presented in [1]. \n2. Why were sequential editing experiments conducted on OKVQA instead of MMEdit and OKEDIT, as proposed in this paper?\n3. Previous studies have indicated that the MEND method can produce NaN values [2] during sequential editing; however, this issue does not appear in Table 4. Are there differences in the sequential editing settings between this study and [2]?\n4. IMHO, if the weights in the language module are edited, it is essential to measure text locality and compare it with other methods.\n5. The paper states that black images are used as negative samples across various visual recognition tasks. It would be beneficial to include citations to support this approach.\n6. Some proprietary terms, such as MiniGPT-4 and BLIP-2 OPT, are used inconsistently throughout the text.\n\n[1] Can We Edit Multimodal Large Language Models？\n[2] VLKEB: A Large Vision-Language Model Knowledge Editing Benchmark" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The issues addressed in this paper are of considerable significance. Existing editing methods affect the performance of the edited model on samples related to the edited ones. This paper proposed a new method to adjust the influence radius dynamically. The innovative approach of using positive and negative samples to estimate the influence radius of each knowledge edit is particularly commendable. Additionally, the paper clearly articulates the above issues and presents corresponding solutions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Existing knowledge editing methods often overlook the influence scope of a knowledge edit, leading to limited generality and locality about samples similar to the edited ones. This paper proposes a novel method, BalancEdit, to optimize the trade-off between generality and locality. To assess this trade-off, this paper constructed a new dataset, OKEDIT. Experimental results demonstrate that BalancEdit outperforms existing methods in both single and sequential editing settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method builds upon Grace [1], with the key differences being using positive and negative samples to estimate the influence radius and the fine-tuning of the transformation layer. However, the paper does not include an ablation study to evaluate the contributions of these two modules.\n2. The proposed dataset OKEDIT employs GPT-4 and a diffusion model to generate rephrased images for assessing image generality. However, previous studies have noted that generated images may shift in content, leading to inconsistencies with the original images [2]. \n3. The use of the harmonic mean (HM) is questionable, as the presence of a minimum can result in a lower harmonic mean. In Table 3, the FT method applied to BLIP2-OPT shows a performance of less than 1% on the locality.\n\n[1] Aging with GRACE: Lifelong Model Editing with Discrete Key-Value Adaptors\n[2] VLKEB: A Large Vision-Language Model Knowledge Editing Benchmark" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "As those mentioned in weakness:\n\nDoes the visual reasoning goal in this approach offer substantial differentiation from MMEdit, given the test data's similarity in the form of QA? \n\nCould a more sophisticated method replace black images as negative samples to better define the balance between generality and locality?\n\nHow significant is the impact of using diffusion model-generated images for testing generality and locality, considering their variable quality? Do you verify the image quality by any means (especially human verification), check if the generated images could be used for test?\n\nWould using more recent model architectures, like those in the LLaVA series, yield different results in these experiments?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The motivation is good, particularly the attention to balancing generality and locality in model edits. \nThe approach for setting the influence radius is straightforward, requiring no additional training, which enhances usability. \nAdditionally, the model demonstrates good efficiency in terms of both time and data requirements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "BalancEdit introduces a new model editing approach, addressing the limitations of traditional model editing techniques. Unlike existing methods, which often ignore the distinct influence of different facts, BalancEdit strikes an optimal balance between generality and locality. By using a codebook of localized edits and generating both positive and negative samples, it accurately assesses each fact's impact without altering the model's core structure. Tested on the newly developed OKEDIT dataset, BalancEdit demonstrates robust editing performance with minimal trade-offs, marking a significant advance in multi-modal model editing." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The method’s visual reasoning goal is limited, offering little differentiation from MMEdit, especially as the test data format remains similar and is based on question answering. \n\nUsing a black image as a negative sample is simplistic and may fall short in defining an \"optimal balance between generality and locality.\" Consequently, the hyperparameter alpha is fixed, potentially limiting flexibility.\n\nImages in the generality and locality tests are generated by a diffusion model, which offers limited advancement over MMEdit due to inconsistent image quality.\n\nThe study uses Blip2-OPT and MiniGPT-4 as baseline models, which are somewhat outdated and limited. Architectures like LLaVA and related models may yield different results.\n\nWriting issue:\nThere is a major issue on page 10, lines 489-514, where two paragraphs convey the same information, likely due to an unintentional oversight.\nTypo: Line 723: “labelis”\nLine 862: missing reference\nTable 3: some bold texts are not best results\nThe example in figure 4 is confusing, because the first image and the rest two has significant difference, and the main subject is two people rather than a church." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No Ethics Concerns" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "As those mentioned in weakness.\n\nAdditions:\n\nHow many locality test image for each editing case? If only one image, this can be imbalanced because generality test has 10 images for each.\n\nDo you verify the quality of generated images? How do you verify them?\n\nWhy don’t you present the results of SERAC method?\n\nWriting issue:\n\nTable 3: Misuse bold texts, some are not best results.\n\nA critical issue exists on lines 489-514, where two paragraphs redundantly convey the same information. This appears to be a significant oversight of the content.\n\nMissing reference in ine 862" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The motivation of balancing generality and specificity in model editing is good. \n\nThe method for determining the influence radius is simple and requires no extra training, which improves usability. \n\nMoreover, the model shows good efficiency in terms of both time and data usage.\n\nIntroducing more generality test for each editing case is beneficial." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "BalancEdit presents a new solution for updating large multi-modal models by achieving a balance between generality and locality in model edits. By introducing the OKEDIT dataset, this approach evaluates and addresses the generality-locality trade-off, a challenge overlooked by other methods. BalancEdit showcases minimal compromise in model performance, offering a robust and efficient approach to knowledge editing without altering the model's core weights." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Using a black or white image as a negative sample is straightforward but may not achieve an optimal balance between generality and locality.\n\nThe editing method involves finetuning a layer, which may be simplistic. Additionally, the experimental results lack comparison with the SERAC method.\n\nAbout image quality, I have some doubts on diffusion generated images, which are used as tests. From fig 4, the second image is totally different from the first image. From fig 6, the first and third examples of generality test are entirely different from the editing sample, making the test results questionable.\n\nThe experiments involve Blip2-OPT and MiniGPT-4. However, considering the fast development of MLLMs, the newer models like LLaVA series, which are widely recognized, should be tested." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024balancedit,\ntitle={BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4WsHgA8EG1},\nnote={under review}\n}" }, "abstract": { "value": "Large multi-modal models inevitably decay over time as facts change and previously learned information becomes outdated. Traditional approaches such as fine-tuning are often impractical for updating these models due to their size and complexity. Instead, direct knowledge editing within the models presents a more viable solution. Current model editing techniques, however, typically overlook the unique influence ranges of different facts, leading to compromised model performance in terms of both generality and locality. To address this issue, we introduce the concept of the generality-locality trade-off in multi-modal model editing. We develop a new model editing dataset named OKEDIT, specifically designed to effectively evaluate this trade-off. Building on this foundation, we propose \\textbf{BalancEdit}, a novel method for balanced model editing that dynamically achieves an optimal balance between generality and locality. BalancEdit utilizes a unique mechanism that generates both positive and negative samples for each fact to accurately determine its influence scope and incorporates these insights into the model's latent space using a discrete, localized codebook of edits, without modifying the underlying model weights. To our knowledge, this is the first approach explicitly addressing the generality-locality trade-off in multi-modal model editing. Our comprehensive results confirm the effectiveness of BalancEdit, demonstrating minimal trade-offs while maintaining robust editing capabilities. Our code and dataset will be available." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multi-modal learning", "Model editing" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/a92fd32526a55eab2c246a388e5aa9ece626237e.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "BalancEdit: Dynamically Balancing the Generality-Locality Trade-off in Multi-modal Model Editing" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Some questions are in the previous section on the weaknesses of the submission; other questions are given below.\n\n1. What are alternative approaches to constructing optimal preconditioner parameters? Please add a paragraph to place your work in the context of learning preconditioners from data or similar techniques. For example:\n- https://proceedings.mlr.press/v202/li23e.html \n- https://sc18.supercomputing.org/proceedings/workshops/workshop_files/ws_lasalss102s2-file1.pdf\n- https://arxiv.org/abs/2405.15557 \n- https://arxiv.org/abs/2401.02016\n- https://arxiv.org/abs/1806.06045\n\n2. What are the spectrum properties of the preconditioned matrix with the generated preconditioner? It would be interesting to observe whether they only reduce the condition number or additionally increase spectrum clustering. Condition numbers for the preconditioned matrices are presented in Tables 3 and 6, but only for limited types of PDEs.\n\n3. What was the $\\epsilon$ parameter used in experiments, and does it significantly affect the training runtime/performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Disclaimer: I am not an expert in symbolic regression, so advances in the study from this perspective could not be well-identified. \n\nThe manuscript's main strength is its attempt to apply the general symbolic regression technique to the preconditioner construction problem. The presented pipeline looks non-trivial, although the objective function is standard. In experiments, the presented approach generates preconditioners that establish faster convergence of the linear solver. In addition, the SyMMap could reconstruct the optimal expression for the $\\omega$ in SSOR for positive definite matrices." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The manuscript presents a novel approach to quasi-optimal preconditioner construction based on symbolic regression. The authors perform extensive numerical simulations to identify the optimal parameters for given linear systems through grid search. The pairs (a parameter for a linear system and the corresponding optimal preconditioner parameter) compose the training dataset. Then, the combination of RL trainer and RNN for symbolic regression fits the analytical expression for the optimal preconditioner parameter. Experimental results demonstrate that the presented pipeline gives such a preconditioner that, on average, the runtime for linear solvers is smaller for different PDE classes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main weakness of this study is that it misses the crucial step of using the derived symbolic expression to generate the optimal preconditioner. I have carefully checked Figure 2 and do not find an explicit connection between the expression $\\tau$ and the compiled preconditioner in the library. Some remarks are given in Section 5.3; however, the presented expressions depend on unclear variables $x_1, x_2, x_3$, so how to use them in preconditioner construction is unclear. \n\nIn addition, I can list the following weaknesses:\n1. The manuscript does not explicitly present the parametrizations of considered PDEs and how these parameters are passed as input to RNN. Also, the authors ignore details on how training and testing sets are prepared.\n2. I did not find the name of the linear solver used to evaluate the preconditioners, e.g., CG, GMRES, or smth else.\n3. The incomplete Cholesky/LU preconditioner is not included in the comparison, although it is among the most powerful.\n4. The authors do not report the runtime for training the presented pipeline (although for Darcy, the runtime is given in Table 6). Moreover, they do not discuss how many linear systems are needed to solve with the generated preconditioner to pay off the training costs compared to classical approaches like SSOR with the optimal parameter or ILU. Tables 1 and 2 show a gain in runtime, which is good, but how much time does the training of symbolic regression require? \n5. For unknown reasons, the authors include a comparison with MLP. However, a more interesting comparison is replacing RNN with Transformer architecture and analyzing the results in the performance of linear solver and training runtime. \n6. No theoretical guarantees on the performance of such an approach or motivation of the presented pipeline are presented, so the robustness of this approach remains unclear." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The proposed framework involves defining the optimal preconditioning parameters, subsequently searching for symbolic expressions, and integrating these expressions into the modern solver." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce Symbolic Matrix Preconditioning (SymMaP), a method that identifies symbolic expressions for efficient preconditioning parameters. These generated expressions can be seamlessly integrated into modern solvers with minimal computational overhead." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Although the application of the proposed approach in this context may be novel, the reasoning behind why it could outperform a class of preconditioners based on graph neural networks (https://proceedings.mlr.press/v202/li23e.html, https://arxiv.org/abs/2405.15557) is not evident.\n* The authors test their framework using three datasets (Biharmonic, Darcy Flow, and Elliptic PDE) with minimal variation in matrix size. It would be beneficial to validate their approach using the SuiteSparse Matrix Collection (https://sparse.tamu.edu/).\n* I found some sections a bit challenging to follow. Could the authors consider reorganizing the paper or providing more detailed explanations for each step of SymMaP to enhance clarity?\n* The values in the columns labeled \"SymMap 1\" and \"SymMap 2\" in tables 1, 2, 3 are not clear and would benefit from additional information to clarify the results.\n* A comparison of the performance in a CPU environment using an MLP is not adequate to make a definitive assertion `symbolic expressions possess equivalent expressive capabilities to neural networks in this scenario, effectively approximating the optimal parameter expressions`." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- The authors use the condition number as the evaluation metric for the AMG experiments, which is hard to compute for large matrices. Why not use computation time instead, which is the most straightforward metric? Would the conclusion be different depending on which metric to use? Note that condition number does not necessarily correlate with time, since coarsening generally poses tradeoffs between preconditioning time and convergence speed.\n\n- Continued from the above question: When constructing the training set for AMG, did the authors use time, iteration, or condition number to determine optimal parameters? \n\n- How do the following three times compare: time of generating the training data, time of training the symbolic regression, and time of one preconditioned solve?\n\n- How many data points are needed to train symbolic regression?\n\n- For the interpretability analysis in 5.3, do the expressions come from SymMaP 1 in Tables 1 to 3? Are there trade-offs between the preconditioning performance of a symbolic formula and the interpretability?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper addresses practical problems.\n\n- The symbolic approach has the potential to reveal the relationship between PDE parameters and preconditioner parameters, leading to new mathematical discovery and analysis.\n\n- The symbolic regression performance is competitive with neural network regression." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the problem of matrix preconditioning, which is a crucial ingredient in the iterative solution of linear systems. In particular, the authors focus on the parameters of some preconditioners and propose a machine-learning approach to determining these parameters. They base their approach on parameterized PDEs, such that it will predict preconditioner parameters given PDE parameters. The authors construct a training dataset and use it to learn symbolic regression formulas for these parameters. They argue that symbolic regression is more efficient and interpretable than other regression techniques." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The empirical evaluation of the proposal is conducted on either too simplistic but less effective preconditioners (SOR and SSOR), or a widely used preconditioner (AMG) with only one single parameter. It would be more informative if the authors experimented with more AMG parameters (such as the choice of the smoother and other coarsening parameters) and conducted a sensitivity analysis.\n\n- The generation of training data is costly: (# data points) * (# grid searches) * (cost of one preconditioned solve)\n\n- The training of symbolic regression can also be costly because of the sample efficiency of reinforcement learning.\n\n- It needs to be clarified of the relationship between genetic programming in section 2.2 and the rest of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024symmap,\ntitle={SymMaP: Improving Computational Efficiency in Linear Solvers through Symbolic Preconditioning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=4WvCoXU2dF},\nnote={under review}\n}" }, "abstract": { "value": "Matrix preconditioning is a crucial modern technique for accelerating the solving of linear systems. \nIts effectiveness heavily depends on the choice of preconditioning parameters. \nTraditional methods often depend on domain expertise to define a set of fixed constants for specific scenarios. \nHowever, the characteristics of each problem instance also affect the selection of optimal parameters, while fixed constants do not account for specific instance characteristics and may lead to performance loss.\nIn this paper, we propose **Sym**bolic **Ma**trix **P**reconditioning (**SymMaP**), a novel framework based on Recurrent Neural Networks (RNNs) for automatically generating symbolic expressions to compute efficient preconditioning parameters. \nOur method begins with a grid search to identify optimal parameters according to task-specific performance metrics. \nSymMaP then performs a risk-seeking search over the high-dimensional discrete space of symbolic expressions, using the best-found expression as the evaluation criterion. \nThe resulting symbolic expressions are seamlessly integrated into modern linear system solvers to improve computational efficiency.\nExperimental results demonstrate that SymMaP consistently outperforms traditional algorithms across various benchmarks. The learned symbolic expressions can be easily embedded into existing specialized solvers with negligible computational overhead. Furthermore, the high interpretability of these concise mathematical expressions facilitates deeper understanding and further optimization of matrix preconditioning strategies." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Matrix Preconditioning", "Symbolic Learning", "Linear System Solver" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1045762bf83eedff7572c541e7db818a4a3046e7.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/851136c5b9bf2bb155ad87b89090e31abf5c9e55.zip" }, "title": { "value": "SymMaP: Improving Computational Efficiency in Linear Solvers through Symbolic Preconditioning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]