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": 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. Some questions regarding the evaluation process:\n(1) The results in Table 3's \"All\" setting do not match those in Table 1. Can you explain the reason behind this gap?\n(2) In Table 3, why not compare ChuLo with other baselines used in Table 1?\n(3) Why do GPT4o and Gemini1.5pro only have results on the \"2048\" setting?\n(4) The NER task prompt used in Figure 8 might not be optimal. Please refer to some related research in this area, such as [1].\n2. Although Algorithm 1 provides some details about the keyphrase extraction process, it would be better if more explanations could be added. For example, the meaning of the regex used (for extracting noun phrases), and the effect for the position penalty. Certain notations are unexplained, like $h$ in line 8.\n3. The proposed method has a lot of hyperparameters: $a, b, n, \\alpha, \\gamma$, to name a few. How did you decide the value for them, and what are the values you used?\n4. Do you have any explanations for why RoBERTa underperforms BERT in Table 8?\n5. Why only emphasize the noun phrases instead of emphasizing key sentences that contain facts about the key phrases?\n6. Some minor mistakes: In Algorithm 1's Line 8, $l_k$ should be $l_{k_i}$. In line 216, add a space within \"key phrases\".\n\n[1] Dhananjay Ashok and Zachary Lipton. PromptNER: Prompting For Named Entity Recognition. arXiv preprint arXiv: 2305.15444." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The proposed method introduces a novel combination of unsupervised keyphrase extraction and chunk-based representation, which benefits encoder models for text classification.\n2. The paper presents a thorough empirical evaluation across several datasets, demonstrating clear performance improvements compared to traditional baselines and SoTA API-based models.\n3. The performance analysis across different document lengths provides useful information for similar research in the future." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces ChuLo, a chunk-level key information representation method aimed at enhancing the efficiency and effectiveness of Transformer-based models for long document processing. ChuLo employs unsupervised keyphrase extraction to create semantically meaningful chunks, prioritizing important tokens while reducing input length. The authors argue that this approach better preserves semantic and contextual information compared to existing techniques such as truncation or sparse self-attention. The method is validated on multiple document classification and token classification tasks, showing competitive performance improvements over baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. From the writing perspective, the structure of certain sections is repetitive and confusing. For example, in Section 3.2, the idea that extracting keyphrases is important is repeated multiple times throughout the paragraph. The same idea is repeated in Section 3.4 as well.\n2. The proposed keyphrase extraction method has some strong inductive bias without explanation, like the position penalty, which is neither explained nor verified through ablation studies. I suppose this design assumes that the noun phrases appear earlier in the text are more likely key phrases. The effect of such a design is not discussed and might limit the use case for the proposed method.\n3. There are some doubts regarding the evaluation process. More details in the Questions part." }, "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": "See 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 method of dividing long documents into manageable chunks is reasonable.\n2. The performance is good particularly in token-level classification." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Chulo, a model that enhances transformer-based approaches for long document-level and token-level classification tasks by effectively integrating chunk-level information. The method of dividing long documents into manageable chunks is reasonable, resulting in good performance especially in token-level classification tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation of this work—\"… can handle long documents efficiently while retaining all key information from the input…\" (lines 55-56)—appears unaddressed. As I understand, the proposed model maintains the same sequence length as other BERT-like models and integrates additional information, such as chunk-level details with key phrases, which in fact increases computational load. The paper would benefit from a dedicated section thoroughly discussing the motivation of the work, or detailing the method’s potential cost savings (e.g., in terms of FLOPs, model size, etc.).\n\n2. The comparison with LLMs appears unfair, as Chulo is fine-tuned on the downstream dataset. To make the comparison more balanced, it would be beneficial to fine-tune some open-source LLMs, such as LLaMA or Qwen, on the same dataset.\n\n3. The design is not novel; similar to hierarchical-BERT [1], it organizes sentences into chunks.\n\n[1] Lu, J., Henchion, M., Bacher, I. and Namee, B.M., 2021. A sentence-level hierarchical bert model for document classification with limited labelled data. In Discovery Science: 24th International Conference, DS 2021, Halifax, NS, Canada, October 11–13, 2021, Proceedings 24 (pp. 231-241). Springer International Publishing." }, "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": "1. I would not recommend using the \"long document\" concept here, as many LLMs like LLaMA have already extended the context length to 131k, whereas this paper handles only up to 10k.\n2. The comparison with GPT-4o is commendable; however, how would LLaMA perform on this task if fine-tuned directly? I don't believe this experiment can be avoided.\n3. If comparing with fine-tuned LLMs or GPT models, I would expect the authors to include inference speed comparisons, which might be one of the method's advantages." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. ChuLo outperforms GPT-4o on certain tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces \"ChuLo,\" a chunk-level key information representation method designed to improve long document processing in Transformer-based models. Traditional models face limitations handling extensive texts due to high computational demands, often resulting in information loss from truncation, sparse attention, or simple chunking methods. ChuLo uses unsupervised keyphrase extraction to identify and emphasize core content within each chunk, enhancing document and token classification accuracy without losing critical details. Experimental results demonstrate ChuLo's superior performance across various datasets, especially for lengthy documents, making it a scalable solution for tasks requiring comprehensive text analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty of the method is limited, as keyphrase extraction is already widely used.\n2. The title and experiments do not align, as \"Long Document Processing\" has a broader scope than classification.\n3. The baselines used for comparison are somewhat outdated, with most being from 2022 or earlier." }, "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. According to line 72, how does the paper determine the performance of the ChuLo method on other types of NLP tasks?\n2. What are the specific details of the model training process described in the paper?\n3. What are the scientific explanations for the significant performance differences demonstrated by ChuLo in Sections 5.4 and 5.5?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Originality: The combination of unsupervised key phrase extraction with chunk representation to improve long document understanding is uncommon in previous research.\n\nQuality: The paper aims to address the practical and significant issue of computational limitations faced by Transformer models when processing long documents. Experimental evaluations conducted on multiple document-level and token-level classification tasks demonstrate the feasibility of the proposed method.\n\nClarity: The paper is structured clearly, with a logical progression from problem statement to methodology, experiments, and conclusions. The detailed description of the SKP algorithm in the paper aids readers in understanding its working principles.\n\nImportance: The proposed ChuLo method enhances the efficiency and performance of long document processing, holding potential for application in long document classification tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a method named ChuLo, designed to address the computational limitations encountered by Transformer-based models when processing long documents. ChuLo extracts key phrases through an improved PromptRank to preserve the core content of the document while reducing the input length. The model is trained using enhanced chunk representations of key information, enabling it to effectively integrate the core semantic content of the document. The paper supports its claims through multiple document-level and token-level classification tasks, providing both qualitative and quantitative analyses. Experimental results demonstrate that ChuLo achieves competitive results across multiple datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The ChuLo method proposed in the paper focuses on long document processing, particularly in document classification tasks. It is stated on line 72 that the contributions of this method include its applicability to various NLP applications, but you have not experimentally confirmed the generalization ability of your method. Therefore, we cannot determine its performance on other types of NLP tasks, such as long document question answering and summarization.\n2. The description of the model training process in the paper is not detailed enough, lacking specific steps of the training. In Section 3.4 of the paper, only the selected model for training is introduced, with no mention of data sources, data processing, optimization algorithms, parameter configurations, or other relevant details.\n3. In Sections 5.4 and 5.5, ChuLo demonstrates significant performance differences compared to existing methods. Therefore, providing scientific explanations for these differences is very important. The lack of analysis of such significant differences in the paper is confusing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024chulo,\ntitle={ChuLo: Chunk-Level Key Information Representation for Efficient Long Document Processing},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=37mG1vvEKf},\nnote={under review}\n}" }, "abstract": { "value": "Transformer-based models have achieved remarkable success in various Natural Language Processing (NLP) tasks, yet their ability to handle long documents is constrained by computational limitations. Traditional approaches, such as truncating inputs, sparse self-attention, and chunking, attempt to mitigate these issues, but they often lead to information loss and hinder the model's ability to capture long-range dependencies. In this paper, we introduce ChuLo, a novel chunk representation method for long document classification that addresses these limitations. Our ChuLo groups input tokens using unsupervised keyphrase extraction, emphasizing semantically important keyphrase based chunk to retain core document content while reducing input length. This approach minimizes information loss and improves the efficiency of Transformer-based models. Preserving all tokens in long document understanding, especially token classification tasks, is especially important to ensure that fine-grained annotations, which depend on the entire sequence context, are not lost. We evaluate our method on multiple long document classification tasks and long document token classification tasks, demonstrating its effectiveness through comprehensive qualitative and quantitative analyses." }, "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": [ "Long Document Processing", "Long Document Classification", "Long Document Tagging" ] }, "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/5198a2aecd15fa75c07f9e00c2d7651ca77a9fb1.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": { "value": "/attachment/c2575b7c7023541aebc19568e132ba6f90fe94eb.zip" }, "title": { "value": "ChuLo: Chunk-Level Key Information Representation for Efficient Long Document Processing" }, "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": null, "code_of_ethics": null, "comment": { "value": "Thank you for sharing your comprehensive review on the manuscript. We are working on addressing each comment and improving the manuscript accordingly. We will share the updates with you soon.\n\nMeanwhile, we have a question on one of the comments:\n\n_“In section 4.4, it is mentioned that the integration of the SDE in the training of the autoencoder follows previous work [Raissi 2017], but it is not sufficiently described to make the paper self-contained.”_\n\nThe SDE is directly made part of the neural architecture, with the mathematical operators in the SDE are direct nodes in the computation graph. We can view this as a non-trainable physics layer in the neural network architecture. Should we reflect this better in Figure 1?" }, "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": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "Initial Response" }, "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": { "value": "Thank you for sharing your comprehensive review on the manuscript. We are working on addressing each comment and improving the manuscript accordingly. We will share the updates with you soon.\n\nMeanwhile, we have a question on one of the comments:\n_“The SDE formulation in Section 3.2 assumes specific forms for the drift and diffusion terms. The justification for these choices comes from prior work, but the implications of these modelling choices should be discussed. What happens when these assumptions are violated?”_\n\nOur formalization of NEE in terms of an SDE is based on the two papers referred to in line 165 (White & Luo 2008; Weng (2011)), with a drift term and a noise term. According to both of them, the noise term can be defined as a Gaussian process. The drift equations in Equation 5 are derived from the model defined in Equations 1,2 and 3. The parameters inside the drift term are predicted by the decoders in the main PIAE model, described in section 4 later. \nIs your question specifically around the parameters σnight and σday in the noise term and the drift? Do you mean we need to do more precise study on each term? Are you referring to the state-of-the-art assumptions?" }, "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": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "Initial Response" }, "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": { "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. The SDE formulation in Section 3.2 assumes specific forms for the drift and diffusion terms. The justification for these choices comes from prior work, but the implications of these modeling choices should be discussed. What happens when these assumptions are violated?\n2. The two-phase training procedure using MSE then MMD requires more theoretical grounding:\n 1. Why this specific sequence? How is convergence of the first phase determined before switching to MMD?\n 2. Were other training strategies considered?\n3. The choice of MMD kernels isn't discussed - how sensitive is the method to this choice?\n4. How sensitive is the model to SDE parameter initialization?\n5. What's the computational overhead versus RF/XGBoost?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Introducing a stochastic differential equation for NEE measurements combining daytime\nand nighttime models with Gaussian noise.\n2. Demonstrating that PIAE improves gap-filling robustness compared to state-of-the-art\nmethods, handling gaps from months to years.\n3. Better Maximum Mean Discrepancy (MMD), Wasserstein distance, and Kullback-Leibler (KL) divergence validated significant improvements in NEE distribution learning.\n4. Achieving better fit to NEE measurements validated by lower MAE and higher R2 scores.\n5. Accurately predicting SDE parameters, enhancing interpretability.\n6. Consistent improvement in nighttime predictions across metrics\n7. Strong performance on distribution-based measures (MMD, Wasserstein, KL)\n8. Ability to capture unusual events (e.g., downward NEE spikes)\n9. Effective parameter estimation for both day and night models" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Physics-Informed Autoencoders (PIAEs) to address gaps in CO2 emission measurements from agricultural fields. The method combines autoencoder architectures with physical Net Ecosystem Exchange (NEE) models, integrating equations that describe CO2 exchanges between the atmosphere and carbon pools (i.e., utilizing\nthe SDE defined as a Wiener process). Their main contribution is extending standard autoencoders with a stochastic differential equation framework that models NEE changes over time, particularly addressing nighttime measurement gaps. Their method also provides forecasting capabilities and enhances performance on NEE gap-filling by accurately learning the NEE distribution and associated parameters. They evaluate their approach on 8 years of flux tower data from East Anglia, showing improvements over current state-of-the-art methods, especially for nighttime predictions, where they achieve a 22% higher R2 score than Random Forest approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the supplementary material adequately explains the SDE derivation and diffusion coefficient determination, key points should be summarized in the main text. A brief note about how σnight and σday are derived from empirical error distributions would help readers understand the transition from Eq. 5 to 6 without requiring supplementary material consultation\n2. AE is better than PIAE for all model parameter estimation across all metrics, contrary to their claim that their method enhances performance on NEE gap-filling by accurately learning the NEE distribution and associated parameters.\n3. The computational requirements compared to simpler approaches like RF are not discussed. \n4. The two-phase training procedure (MSE then MMD) has no convergence guarantees.\n5. The claimed 22% improvement in R2 score lacks context - no variance was reported (Error bars or confidence intervals for the reported metrics would help). The hyperparameter selection process for PIAE and baseline models (including random forest) is not described. A fair comparison requires careful tuning of all methods.\n6. Missing critical details:\n 1. How were hyperparameters selected for PIAE and baselines?\n 2. What are the network architectures (layer sizes, activation functions)?\n 3. Where are the error bars and statistical significance tests?\n 4. How does computational cost compare to simpler methods\n7. The implementation details are insufficient for the reproduction\n8. The comparisons in Figures 2 and 3 show selective periods without justification for their choice\n\nMinor comments:\n1. Section 4.5's description of the loss function uses inconsistent notation compared to earlier sections. \n2. There are some writing clarity issues, like in lines 50 and 98.\n3. The paper shows results across different timescales but doesn't systematically evaluate performance as a function of gap length. This would be valuable for understanding the method's practical utility.\n4. The NEE parameter estimation details might fit better in methods" }, "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": 1 }, "primary_area": null, "questions": { "value": "1. Could the work be applied to other physical systems? Would that require knowledge of the DFE governing the system?" }, "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 applies ML techniques to enhance NEE measurements, which has the potential to improve the estimation of Co2 emissions, resulting in reduced uncertainty in our projections. **This is an important problem with high societal and environmental impact.**" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the application of autoencoders for the problem of imputing missing values in Co2 Net Ecosystem Exchange (NEE) measurements. The autoencoder takes in several covariates, such as temperatures and radiations, at a given timestep $t$ and predicts the next-step NEE, along with several variables of a Stochastic Differential Equation that models changes in NEE." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **The presentation of the paper is convoluted, and requires a degree of familiarity with the NEE problem that is uncommon in the ICLR community**. It is then hard for me to judge the significance, originality and potential impact of the work.\n More precisely, these are some points that are not sufficiently explained or that make the paper hard to read and understand:\n- In the introduction, it is mentioned that missing NEE values are due to e.g. power shortages. I assume that in such scenarios, the values of the covariates (temperature, radiation, etc) are also missing due to the same issue. However, the proposed model requires having access to all covariate values at a given time. How can the model be applied in practice without these values?\n- In the introduction, the first highlighted contribution is the introduction of a SDE for NEE measurements. Put it that way, it sounds like the SDE is novel also in the physics. However this point is not stressed again later on, so I wonder whether the SDE is known and the novelty is in its use as supervision for learning ML models.\n- Line 157, it is mentioned that the $E_0$ parameter is estimated with the nighttime model and used in the daytime one, but it is not explained why.\n- In section 4.4, it is mentioned that the integration of the SDE in the training of the autoencoder follows previous work [Raissi 2017], but it is not sufficiently described to make the paper self-contained.\n- The related work is not sufficiently described. In particular, it is not clear whether the reported baselines RFR and XgBoost variant based on the work of [Moffat 2007] are also physics-informed or only statistical.\n- Second and third lines of Equation 5: do the second (from the left) commas separate two different definitions or do they indicate the continuation of the variable suffices?\n\n2. The tables do not report standard errors, which makes impossible to judge the significance of the improvements.\n\n3. The paper does not discuss limitations nor 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": 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": 1 }, "primary_area": null, "questions": { "value": "Why is latent heat (L) excluded? Having a high correlation to the target variable would seem to be a good thing when the goal is to predict missing values?\n\nThe notation in equation 4-5 is difficult to read. Would it not be more clear to write this in terms of partial derivatives?\n\nFor improved readability, consider to use italics for variables and roman (upright) type for named functions, as subscripts in equations, and for units of measurement. Consider that multi-letter abbreviations can be confusing in equations: For example, it can be unclear if rb is a single variable or the product of r and b.\n\nrb (night/day) is not defined in the main text as far as I can see. rb is mentioned in the text in the appendix but not in the mathematical derivations.\n\nIn equation 9, should is there not a difference between dt on the left and right hand side? On the left hand side, it seems to denote an infinitessimal element, and on the right side it is 30 minutes?\n\nI am not sure how this approach is an autoencoder. As I understand the written description, the model predicts one timestep ahead with a latent encoding, and thus does forecasting rather than reconstruction. However, Figure 1 does seem to imply that the decoders predict for the same timestep.\n\nIs there something wrong with the linebreaks in Algorithm 1, step 4?\n\nWhat is the reason for the choice of the two loss phases?\n\nI am not familiar with the literature on physics-informed autoencoders, but I would like to ask whether this paper introduces any technical contributions to the framework itself, or if the contribution is primarily the application of an existing modeling framework to significant applications." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed method seems a good fit to the application." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the problem of forecasting CO2 emission from agricultural fields based on measurement data. In particular, the problem of predicting missing data is addressed. The authors present a set of stochastic differential equations that govern the net ecosystem exchange (NEE) that are used in a physics-informed autoencoder for data imputation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper was challenging to follow, primarily due to unclear notation and insufficient definitions of certain terms. Additionally, some design choices, such as the two-phase loss, are described but lack clear justification.\n\nThe main contribution appears to be a relatively straightforward application of an existing methodology to a specific domain. The novelty largely lies in application-specific details, which may not align closely with the primary interests of the ICLR community." }, "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": "The use of MMD seems a little bit odd. MMD is essentially a two-sample statistical test to identify of those samples are from the same probability distribution. We usually expect the two samples are from two independent realizations. But, based on the loss function, two samples are from the same realization, just one is the data and the other is a model prediction. If they are from the same realization ($\\omega^j$ in author's notation), minimizing the distance would make more sense, like the first phase of the training. In the end of day, for two samples from the same realization, minimizing MMD corresponds to minimizing MSE. But, all the hyperparameters (like the RBF kernel) makes it much less straightforward." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "It is shown that the proposed method outperformed some of the baseline methods that are used in the domain. It seems to suggest a potential of replacing the conventional machine learning models with the PINN models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this study, the authors utilized physics-informed neural network framework (PINN) to develop an auto-encoder for the forecasting of carbon dioxide emission. First, the overall physical process is modeled by using a set of ordinary differential equations and PINN is used to train a neural network. The authors proposed a two-stage training method, that in the first stage the neural network is trained by minimizing the mean absolute error and, then, in the second phase, the maximum mean discrepancy score is minimized. It is shown that the proposed method outperforms some of the naive baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "First of all, the study is mainly focused on the application of the widely used PINN to a physical process for a specific domain. It does not look like there are novel algorithms or problem setups that can be of interest to a broader machine learning community. I would like to suggest the authors to submit this manuscript to a more domain specific venue.\n\nThe paper is not very well written. It is unclear how the SDE formulation is treated in the modeling, how the SDE and model are used for uncertainty quantification, how the evaluations were made by using what variables as inputs and predict how long in the future, and so on. I assume that this is due to the page limitation. It would have been better if the authors had put all the domain specific modeling sections in the appendix and focused more on the generic problem set up in the main body." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a Physics-Informed Autoencoder in order to fill the gaps and improve the training of the Net Ecosystem Exchange measurements at farm scale." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024physicsinformed,\ntitle={Physics-Informed Autoencoder for Enhancing Data Quality to Improve the Forecasting Reliability of Carbon Dioxide Emissions from Agricultural Fields},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=381rZinzJE},\nnote={under review}\n}" }, "abstract": { "value": "Missing values in measurements for carbon dioxide emissions on drained peatlands remains an open challenge for training forecasting techniques to achieve net zero. At the field scale, existing methods struggle to model $\\ce{CO_2}$ emissions to fill gaps, especially in nighttime measurements. We propose robust Physics-Informed Autoencoders (PIAEs), which combine the generative capabilities of Autoencoders with the reliability of physical models of Net Ecosystem Exchange (NEE) that quantify $\\ce{CO_2}$ exchanges between the atmosphere and major carbon pools. Our method integrates equations describing the physical processes and associated uncertainties to fill gaps in NEE measurements from eddy covariance (EC) flux towers. In the PIAE, various sensor measurements are encoded into the latent space, and a set of decoders is then used to approximate the ecosystem parameters and the optimal NEE forecast, directed by dynamics described by a stochastic differential equation. These decoders utilize nighttime and daytime NEE models that describe carbon transfer as a Wiener process. Finally, we use a two-phased training routine with two loss functions describing each phase: Mean Squared Error (MSE) and Maximum Mean Discrepancy (MMD) between the measurements and the reconstructed samples. PIAE outperforms the current state-of-the-art Random Forest Robust on the prediction of nighttime NEE measurements on various distribution-based and data-fitting metrics. We present significant improvement in capturing temporal trends in the NEE at daily, weekly, monthly and quarterly scales." }, "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": [ "physics-informed machine learning", "autoencoders", "gap-fillling", "net ecosystem exchange", "noise", "stochastic differential equation" ] }, "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/3bc3a2bd321ccef5a0204682c5aff4f38b0071e0.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/8007fd71add63cb21f09410c31255e4f7d82850c.pdf" }, "title": { "value": "Physics-Informed Autoencoder for Enhancing Data Quality to Improve the Forecasting Reliability of Carbon Dioxide Emissions from Agricultural Fields" }, "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": null, "code_of_ethics": null, "comment": null, "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": null, "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": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "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": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.This paper is easy to follow and well-written.\n2.The paper introduces the Consistency Diffusion Model, which incorporates 3D structural priors as constraints within a diffusion framework. The topic is interesting.\n3.The model employs a Bayesian framework, incorporating a new constraint term that introduces 3D structural priors into the variational process. This improves the model's consistency by raising the Evidence Lower Bound (ELBO), reducing uncertainty, and enhancing overall stability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel Consistency Diffusion Model (CDM) designed to improve single-image 3D reconstruction. The proposed model leverages both 2D and 3D prior information to enhance the consistency of the reconstruction process, yielding promising results in single-view scenarios. Experimental evaluations demonstrate that CDM outperforms existing methods in certain cases, underscoring its potential." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The inclusion of multiple priors, complex rotation matrices, and depth mapping computations increases the computational burden during training. There is a lack of detailed information on the training time and computational efficiency.\n2.Additional experiments incorporating different types of 3D structural priors and 2D image priors, as well as testing on a broader range of datasets, would help to validate the model’s generalizability and robustness across diverse conditions.\n3.The paper notes that inconsistent conditions between the training and sampling phases can lead to \"model drift,\" causing learning biases and unstable results. This could result in a performance gap between the training and deployment phases, affecting the model's real-world reliability. However, potential methods for mitigating or addressing the issue of model drift are not 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": 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": "Few questions that the reviewer has for the paper:\n1. How does the model perform on non-object centric scenes? (i.e, more extrapolating views)\n2. How are random camera parameters sampled? How did the authors make sure that there is no bias brought in on the sampling procedure of the camera parameters?\n\nComment: \nTypo in Figure 2 (a)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Few strength of the papers that the reviewer appreciates are:\n1. The paper is fairly well written and easy to follow. \n2. The contributions are clearly isolated, from 2D and 3D, making it easy to justify the quality of each contributions in the ablation in the tables, e.g, Table 4\n3. Numerous technical ablations conducted to demonstrate how each components and their respective small deviations work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new way to generate 3D point clouds from a single image. The main contribution of the paper is to use 2D and 3D priors as a way to nudge diffusion models for robustness in the ill-posed problem domain of single view 3D point cloud estimation. \n2D priors are extracted from DINOv2, as depth and contour (as well as features). 3D priors are extracted as random camera transformations around an object of interest. \nThe results demonstrate that the method can outperform existing methods for single view point cloud estimation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Main weakness of the paper is that the contributions are not as well novel. Usage of 2D features and derivatives such as depth and contours are easily justifiable. However, because it is so clear and evident that use of depth and contour as a way to regularize 3D point cloud reconstruction will help, the reviewer does not find it as fundamental contribution to the community.\nIn other words, yes we know that the usage of depth and contours will help, and yes the paper has re-verified it. What leaves the takeaways? The reviewer is unsure if usage of these priors as a form of augmentations are worthy of contributions to the ICLR venue. \n\nIn addition, other contributions, such as usage of consistency in the diffusion process is not new; it may be new in the realm of point cloud diffusion, but it would be application of existing approaches." }, "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.\tCould the authors provide a theoretical basis for enforcing consistency terms across all diffusion steps, rather than focusing on key steps where 3D structure becomes clearer?\n\t2.\tHow does the addition of consistency terms impact the model’s ability to generate diverse outputs? Is there a measurable loss in generative variability?\n\t3.\tCould the authors elaborate on why 3D priors, rather than other forms of regularization, are necessary for improving consistency?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper does present an effort to enhance consistency in 3D shape reconstruction by combining 2D and 3D priors within a Bayesian framework. While the theoretical foundation is weak, the authors have demonstrated some commitment to experimenting with consistency terms, attempting to tackle an important challenge in generative modeling." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes the Consistency Diffusion Model (CDM) for 3D point cloud reconstruction from a single image, integrating both 2D and 3D priors to enhance consistency in the Bayesian framework. By converting 3D structural priors into a bound term, the approach aims to increase evidence in the variational Bayesian framework, and 2D priors from the input image are projected onto the 3D point cloud for additional guidance. The authors claim this framework offers a consistency advantage over existing methods and improves performance on synthetic and real-world datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tInsufficient Justification for Consistency Terms: The paper lacks a solid theoretical foundation for enforcing consistency terms at each diffusion step, which may not align with the iterative nature of the diffusion process. This raises concerns about the model’s validity.\n\t2.\tImpact on Variance of Generative Model: Enforcing consistency terms across all steps could reduce the model’s generative diversity, possibly resulting in outputs that lack the variability expected in a robust generative framework. This could push the model towards a U-Net-like structure, potentially sacrificing the inherent variability necessary for effective 3D generation.\n\t3.\tExperimental Limitations: The experiments do not convincingly demonstrate that this approach generalizes well. The benchmarks and comparative studies are limited, and it is unclear if the performance improvements observed are due to the proposed model’s consistency terms or other factors.\n4. I believe equation 5 is incorrect. The variational bound term should have the joint probability p(x_0 : x_t) as the numerator." }, "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.How is the camera matrix selected? If only random sampling is used, the images rendered from adjacent views will be very similar.\n\n2.Please refer to the questions and suggestions in the Weaknesses part." }, "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 integrates 3D and 2D priors to the reconstruction task. The results achieve SOTA performance in ShapeNet and Co3D dataset." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to reconstruct point cloud from a single image. First, they convert 3D structural priors derived from the initial 3D point cloud as a bound term to increase evidence in the variational Bayesian framework. Second, they extract and incorporate 2D priors from the single input image, projecting them onto the 3D point cloud to enrich the guidance for diffusion training. The results show SOTA performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The paper introduces 3D priors constraint to refine the reverse process. Does this part increase the training time of the model? It is better to give a comparison of the training time of the model with and without the 3D prior.\n\n2.Visual comparison is not sufficient. Only three samples were selected in the qualitative experiment of the Shapenet dataset (Figure 5), and the differences are not quite visible in all the samples except for the middle sample where CDM showed advantages. Besides, the advantage of CDM over PC2 is not apparent from the Visual comparison on the Co3D dataset in Figure 6.\n\n3.The depth map rendered from the point cloud is a random sample of the 3D geometry, and a lot of information is lost in the sampling process. It is better to directly adopt 3D geometric representation such as SDF as 3D prior.\n\n4.The paper lacks sufficient research and comparison on relevant methods. There are a large number of methods that can reconstruct point clouds from a single image with good results. For example, TriplaneGaussian[1] can generate multi-view rendering results in addition to point clouds; Michelangelo[2] can generate point clouds corresponding to text and images; and CLAY[3] trained a large model for point cloud generation. The ability of these methods to generate point clouds from images is not limited to a few single categories, and they have good generalization ability. These methods should be discussed and compared in the paper.\n\n[1] Zou Z X, Yu Z, Guo Y C, et al. Triplane meets gaussian splatting: Fast and generalizable single-view 3d reconstruction with transformers[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 10324-10335.\n\n[2] Zhao Z, Liu W, Chen X, et al. Michelangelo: Conditional 3d shape generation based on shape-image-text aligned latent representation[J]. Advances in Neural Information Processing Systems, 2024, 36.\n\n[3] Zhang L, Wang Z, Zhang Q, et al. CLAY: A Controllable Large-scale Generative Model for Creating High-quality 3D Assets[J]. ACM Transactions on Graphics (TOG), 2024, 43(4): 1-20." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Consistency Diffusion Models for Singel-Image 3D Reconstruction" }, "_bibtex": { "value": "@misc{\njiang2024consistency,\ntitle={Consistency Diffusion Models for Singel-Image 3D Reconstruction with priors},\nauthor={Chenru Jiang and Chengrui Zhang and Xi Yang and Jie Sun and Kaizhu Huang},\nyear={2024},\nurl={https://openreview.net/forum?id=385gQZuuuR}\n}" }, "abstract": { "value": "This paper delves into the study of 3D point cloud reconstruction from a single image. Our objective is to develop the Consistency Diffusion Model, exploring synergistic 2D and 3D priors in the Bayesian framework to ensure superior consistency in the reconstruction process, a challenging yet critical requirement in this field. Specifically, we introduce a pioneering training framework under diffusion models that brings two key innovations. First, we convert 3D structural priors derived from the initial 3D point cloud as a bound term to increase evidence in the variational Bayesian framework, leveraging these robust intrinsic priors to tightly govern the diffusion training process and bolster consistency in reconstruction. Second, we extract and incorporate 2D priors from the single input image, projecting them onto the 3D point cloud to enrich the guidance for diffusion training. Our framework not only sidesteps potential model learning shifts that may arise from directly imposing additional constraints during training but also precisely transposes the 2D priors into the 3D domain. Extensive experimental evaluations reveal that our approach sets new benchmarks in both synthetic and real-world datasets. The code will be released." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Chenru_Jiang1", "~Chengrui_Zhang1", "~Xi_Yang1", "~Jie_Sun6", "~Kaizhu_Huang1" ] }, "authors": { "value": [ "Chenru Jiang", "Chengrui Zhang", "Xi Yang", "Jie Sun", "Kaizhu Huang" ] }, "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": [ "Bound", "Variational Bayesian", "3D Point Cloud", "Single-Image", "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": { "value": "jiang|consistency_diffusion_models_for_singelimage_3d_reconstruction_with_priors" }, "pdf": { "value": "/pdf/5e1dc3f92af19c8f24cf19a60c737b897228bb32.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/8c855e79f390b08aea7dc361464218cdee783eaf.zip" }, "title": { "value": "Consistency Diffusion Models for Singel-Image 3D Reconstruction with priors" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_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": 3 }, "primary_area": null, "questions": { "value": "Is it necessary to take an increasing function f(t)? Other than removing the dependence between data at step t and t+f(t), is there any other reason to use an increasing function?" }, "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 tackle is well stated, which is to stabilize off-policy learning and improve the previous algorithm ATD and GTD. Also, the paper is clearly written and easy to follow, with rigorously stated assumptions and lemmas." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper aims to develop a convergent algorithm for off-policy temporal difference learning under linear function approximation. The proposed algorithm directly minimizes the L2-norm of expected TD updates (NEU) $||Aw+b||^2$, improving the memory requirement of estimating matrix $A$ from the previous ATD algorithm. Meanwhile, the proposed algorithm reduces the number of learning rates from two to one compared to GTD, another NEU minimization algorithm. It maintains the convergent property with a convergent rate $\\tilde{O}(1/t)$. Moreover, the algorithm is tested on Baird’s counterexample and is shown to avoid divergence in the deadly triad." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The approach needs to be more motivated. GTD is known to suffer from a low convergent rate compared to TD. However, the proposed algorithm does not handle this key problem. More experiments to compare the convergence speed and some intuition on why the proposed algorithm can fasten the learning would be great.\n\nAlso, the algorithm needs to fit better into the literature. ETD, introduced by Mahmood and colleagues (2015), is another stable off-policy algorithm. Also, a target network is suggested to help convergence (Zhang et al., 2021; Fellows et al., 2023; Che et al., 2024). Che et al. (2024) compare their TD algorithm with GTD on Baird’s counterexample, showing much faster convergence.\n\nReference\n\nMahmood, A. R., Yu, H., White, M., & Sutton, R. S. (2015). Emphatic temporal-difference learning. arXiv preprint arXiv:1507.01569.\n\nZhang, S., Yao, H., & Whiteson, S. (2021, July). Breaking the deadly triad with a target network. In International Conference on Machine Learning (pp. 12621-12631). PMLR.\n\nFellows, M., Smith, M. J., & Whiteson, S. (2023, July). Why target networks stabilise temporal difference methods. In International Conference on Machine Learning (pp. 9886-9909). PMLR.\n\nChe, F., Xiao, C., Mei, J., Dai, B., Gummadi, R., Ramirez, O. A., ... & Schuurmans, D. (2024). Target Networks and Over-parameterization Stabilize Off-policy Bootstrapping with Function Approximation. arXiv preprint arXiv:2405.21043." }, "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": "Here are a few questions that might affect the evaluation:\n1. Do the choices of $f(t)$ depend on the induced Markov chain’s mixing time? If yes, where is it in the result? If not, why?\n2. What is the convergence rate for GTD? Is it the same as the canonical on-policy TD as well?\n3. In Line 182, when $f(t)$ was a constant function, what would happen by following the classical convergence results mentioned in Line 186? What’s the consequence of using such a $f(t)$?\n4. In Line 227, the paper claims the technique of using a variable interval $T_m$ has not been used in any stochastic approximation and RL literature. Has it been used or studied in other fields? If yes, then it is worth mentioning the relevant literature.\n5. In Line 482, the paper claims that the finite sample analysis **confirms** that the proposed algorithm converges reasonably fast, but this analysis is based on a variant of the proposed algorithm with a projection step. Does the efficiency of the variant with the projection step guarantee or generally suggest the efficiency of the base algorithm?\n6. Is the proposed algorithm more or less sensitive to the learning rate compared to GTD? Since GTD has two hyperparameters, comparison methods like those in Ghiassian & Sutton, 2021 might be useful here." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The strengths of the paper include its originality, quality, and clarity:\n1. The idea in this paper is novel to the best of my knowledge. It’s neat to use two samples distanced away from each other to estimate the terms involving two $A$ matrices, which are independent if their gap is large. The sublinear memory requirement also renders this idea a practical approach.\n2. The quality of the paper is also a strength. The asymptotic convergence of the proposed algorithm is novel and may bring value to the community.\n3. The paper is very well written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new variant of the gradient temporal difference (GTD) learning algorithm for online off-policy policy evaluation with linear function approximation. The idea is to use two samples distanced away from each other to address the double sampling issue encountered during the derivation of GTD. The paper shows that when the distance between the two samples $f(t)$ used to estimate the gradient increases with a proper rate (e.g., $f(t)=\\ln^2(t+1)$), the new algorithm converges asymptotically, while its variant with a projection operator has a convergent rate comparable to on-policy TD. The consequence of this new GTD variant is that 1) it reduces the need for an additional set of weights and step size, and 2) it requires an additional memory of size $O(\\ln^2(t))$. Preliminary experiment results on Baird’s counterexample show the effectiveness of the proposed algorithm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper has weaknesses in its significance and relevant work discussion:\n1. The paper may be limited in its significance. \n - On the theory side, the finite time analysis is based on a variant of the proposed algorithm with a projection step, which is absent in the actual algorithm. Thus, the comparison between its convergence rate in this case with that of on-policy TD may not be very valuable. Note that finite sample analysis of the actual algorithm is possible, as also pointed out in the paper. Obtaining such a result can strengthen the paper.\n - On the empirical side, the experiments presented in this paper only focus on Baird’s counterexample and are rather limited. Having more experiments, even in simple environments like FourRoom (Ghiassian & Sutton, 2021; Ghiassian et al., 2024), would help strengthen the claim that the proposed algorithm is effective. In addition to testing the proposed algorithm in environments like FourRoom, a comparison with other GTD algorithms can also make the paper stronger. Other researchers may find it useful to know how the proposed algorithm compares to others in terms of sample efficiency, stability, and hyperparameter sensitivity.\n2. The paper also lacks a thorough related work discussion on off-policy policy evaluation (OPPE) with linear function approximation. There have been many follow-up works on the GTD algorithm (Mahadevan et al., 2014; Ghiassian et al., 2020; Yao, 2023). While some of them are cited in the paper, the relationship between these later ideas building on GTD and the proposed method is not thoroughly discussed, which could be useful and inspire future research. Note that Yao (2023) also introduces a GTD variant with one step-size, so it may be necessary to clarify its distinction with your approach. In addition, the paper may also benefit from discussing another line of work that addresses the deadly triad, the ETD family (Sutton et al., 2016; Hallak et al., 2016, 2017; He et al., 2023). Specifically, how does the proposed approach compare to these methods in terms of the optimality of the fixed point and the convergence property? Having a more thorough discussion of these relevant works would strengthen the paper’s positioning.\n\nGhiassian, S., Patterson, A., Garg, S., Gupta, D., White, A., & White, M. (2020). Gradient temporal-difference learning with regularized corrections. ICML.\n\nGhiassian, S., & Sutton, R. S. (2021). An empirical comparison of off-policy prediction learning algorithms in the four rooms environment. arXiv preprint arXiv:2109.05110.\n\nGhiassian, S., Rafiee, B., & Sutton, R. S. (2024). Off-Policy Prediction Learning: An Empirical Study of Online Algorithms. IEEE Transactions on Neural Networks and Learning Systems.\n\nHallak, A., Tamar, A., Munos, R., & Mannor, S. (2016). Generalized emphatic temporal difference learning: Bias-variance analysis. AAAI.\n\nHallak, A., & Mannor, S. (2017). Consistent on-line off-policy evaluation. ICML.\n\nHe, J., Che, F., Wan, Y., & Mahmood, A. R. (2023). Loosely consistent emphatic temporal-difference learning. UAI.\n\nMahadevan, S., Liu, B., Thomas, P., Dabney, W., Giguere, S., Jacek, N., ... & Liu, J. (2014). Proximal reinforcement learning: A new theory of sequential decision making in primal-dual spaces. arXiv preprint arXiv:1405.6757.\n\nSutton, R. S., Mahmood, A. R., & White, M. (2016). An emphatic approach to the problem of off-policy temporal-difference learning. JMLR.\n\nYao, H. (2023). A new Gradient TD Algorithm with only One Step-size: Convergence Rate Analysis using $ L $-$\\lambda $ Smoothness. arXiv preprint arXiv:2307.15892." }, "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": "Please see my comments in Weaknesses section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper expand the idea from $A^TTD$ and proposed a new method to solve double sampling issue in off-policy learning. Compared with $A^TTD$, this methods required less memory. The authors also provided the convergence analysis of their method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a new solution to solve double sampling issue in off-policy reinforcement learning. Specifically, this paper provided another method to estimate $A^T A$ and $A^T b$ by introducing a function $f(t)$. The authors also provided finite-time analysis for their method as well as some numerical results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper is really interesting to me. However, I have several questions. \n\n1. The advantage of selecting $f(t)$ as an increasing function over a constant one is not immediately clear. The authors state in lines 186–187 that classical convergence analysis can be applied to establish the convergence rate. Thus, it seems that a constant $f(t)$ could also ensure convergence. Additionally, the experimental results suggest that setting $f(t)=2$ is sufficient to resolve Baird’s counterexample, which further supports the idea of choosing $f(t)$ as a constant.\n\n2. Relying on samples from several steps prior may introduce additional errors during policy improvement. Although this paper focuses exclusively on policy evaluation, it is important to mark that policy evaluation serves the purpose of policy improvement. In cases where the policy is continuously updated, the samples used to estimate $A^T$ may become inaccurate, introducing further errors.\n\n3. A minor issue appears in lines 206–207, where I believe the correct notation should be $\\bar{h}(\\omega_{t_m})$ instead of $h(\\omega_{t_m})$" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024revisiting,\ntitle={Revisiting a Design Choice in Gradient Temporal Difference Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=38BBWrXUhP},\nnote={under review}\n}" }, "abstract": { "value": "Off-policy learning enables a reinforcement learning (RL) agent to reason counterfactually about policies that are not executed and is one of the most important ideas in RL. It, however, can lead to instability when combined with function approximation and bootstrapping, two arguably indispensable ingredients for large-scale reinforcement learning. This is the notorious deadly triad. The seminal work Sutton et al. (2008) pioneers Gradient Temporal Difference learning (GTD) as the first solution to the deadly triad, which has enjoyed massive success thereafter. During the derivation of GTD, some intermediate algorithm, called $A^\\top$TD, was invented but soon deemed inferior. In this paper, we revisit this $A^\\top$TD and prove that a variant of $A^\\top$TD, called $A_t^\\top$TD, is also an effective solution to the deadly triad. Furthermore, this $A_t^\\top$TD only needs one set of parameters and one learning rate. By contrast, GTD has two sets of parameters and two learning rates, making it hard to tune in practice. We provide both asymptotic and finite sample analysis for $A^\\top_t$TD, where the convergence rate is on par with the canonical on-policy temporal difference learning. Key to our analysis is a novel refined discretization of limiting ODEs." }, "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": [ "gradient temporal difference 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/bc43654a3f2154c6f73b119ba3e3b9ce49ed7aff.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": "Revisiting a Design Choice in Gradient Temporal Difference 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": 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 weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Motivated by the observation that RLA methods suffer from notable loss in spatial awareness for CLIP ViTs, SCD is specifically designed to capture the spatial structure in a region. The widespread experiments show the superior performance of the proposed method across multiple open-vocabulary dense prediction benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to improve upon the Region-Language Alignment (RLA) approaches for Contrastive Language-Image Pre-training (CLIP) models. In order to not only promote the linguistic alignment but also preserve the spatial awareness, Spatial Correlation Distillation (SCD) is proposed to plug into the existing methods such as RegionCLIP and CLIPSelf. Refiner is also introduced to enhance the regional representation of teacher model. The experiments on the open-vocabulary dense prediction tasks demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The motivation to design Refiner has been implied in CLIPSelf. When K = N, the proposed Refiner is almost the same as CLIPSelf, which reduces the technical novelty. Also, the ablation study of K is absent in this work.\n\nThe application of the proposed method is restricted. Spatial Correlation Distillation (SCD) is auxiliary and is to preserve the spatial awareness when another optimization is applied (e.g. RLA). Therefore, it seems that SCD cannot be applied independently since in this case the weights of teacher model and student model are always equal. Besides, R-SC-V is only learned from the teacher model that is optimized by Refiner (similar to CLIPSelf), so it cannot be further applied to CLIPSelf based approach.\n\nIn order to showcase the generalizability and scalability of the proposed method, the experiments with data scaling up are expected to be provided, which is missing in the current version." }, "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": "1. In the SC-RLA, how to ensure the correspondence between the RoI regions from teacher and student models? The student model would be trained, so the number, order and attribute of the RoI regions would become different with the frozen image encoder.\n\n2. In the line 242 -258, would the contents of embeded sampled images {X_i} affect the SPATIAL AWARENESS? Could you provide some quantitative analysis like the cosine similarity between the {X_i} and X_t ?\n\n3. Since the motivation and module is to improve spatial awareness (which can also be seen from the visualization), are there more segmentation related visualizations? Qualitative results using segmentation would be more convincing (e.g. finer edges)" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Experiments. The experiments are sufficent and solied to prove the improvement of framework.\n\n2. Significance. The framework is a convience, plug and play, and effecitive pipeline to improve the existing methods of the OV tasks.\n\n3. Motivation. The motivation is clear. The Fig.1 and intro explain the loss of dense prediction ability of RLA process." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework called Spatial Correlation Distillation(SCD) to refine the spatial awareness of CLIP. To recall the visual perception of vanilla CLIP loss in Region-Language Alignment (RLA) training, it proposes a Spatial Correlation Distillation training process and a light-weight refiner to distill the visual relationship from the vanilla CLIP. SCD can improve the existing OV methods and achieve the new sota on the ov tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Novelty. The frame work is composed with SCD and a refiner and the novelty of the SCD is limited. The SCD is a common distillation module to reserve the similarity between the RoI features of student and teacher models. \n\n2. Lack of analysis. This paper does not provide a deeper analysis of the experimental results. For example, why R-SC can effectively improve the classification accuracy in Tab.1? In terms of motivation and structure, this module improves spatial perception and does not have much gain in classified task with the annotated masks..\n\n3. Typo. What is the Fig.2.o in the line 193?" }, "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. It might be better to use \"language\" or \"text\" instead of \"linguistic\"\n\n2. How come methods specifically designed for CLIP dense prediction like CLIPSelf and RegionText work even worse than vanilla CLIP?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This pager reveals the potential problems of previous works, that \"the RLA process projects dense visual embeddings into a text-oriented domain, making them incompatible with visual-centric objectives\". To tackle this, the paper proposes to conduct Spatial-Correlation-guided Region-Language Alignment with Refiner to preserve spatial awareness. This design is novel and reasonable.\n\n2. The performance improvement is significant compared with baseline methods.\n\n3. The experiments and analysis are comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to enhance CLIP's spatial awareness. That is to say, increases the quality of dense features extracted by CLIP. It proposes the Spatial Correlation Distillation (SCD) framework, which preserves CLIP's inherent spatial structure and mitigates degradation for spatial awareness by Region-Languaeg Alignment. It also introduces a lightweight Refiner that extracts refined correlations directly from CLIP before feeding them into SCD, based on an intriguing finding that CLIP naturally captures high-quality dense features." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper writing is not so rigorous. Authors should give clear definitions of each term they are discussing. Like what is the definition of \"spatial awareness\", what it means by a better dense feature, what is \"visual-centric\", what is \"intra-feature structural relationships\", etc.\n\nAs an example, spatial awareness here is (probably) defined as performance for tasks like localization and recognition, which I think, is equivalent to increasing the quality of dense features extracted by CLIP, which means different parts in the image with different semantics should be extracted with the features that are distinguishable from each other." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We refine CLIP’s region-language alignment by enhancing its spatial awareness, improving performance from both visual-centric and vision-language perspectives." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024refining,\ntitle={Refining {CLIP}'s Spatial Awareness: A Visual-Centric Perspective},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=38No4B8sx6},\nnote={under review}\n}" }, "abstract": { "value": "Contrastive Language-Image Pre-training (CLIP) excels in global alignment with language but exhibits limited sensitivity to spatial information, leading to strong performance in zero-shot classification tasks but underperformance in tasks requiring precise spatial understanding. Recent approaches have introduced Region-Language Alignment (RLA) to enhance CLIP's performance in dense multimodal tasks by aligning regional visual representations with corresponding text inputs. However, we find that CLIP ViTs fine-tuned with RLA suffer from notable loss in spatial awareness, which is crucial for dense prediction tasks. To address this, we propose the Spatial Correlation Distillation (SCD) framework, which preserves CLIP's inherent spatial structure and mitigates above degradation. To further enhance spatial correlations, we introduce a lightweight Refiner that extracts refined correlations directly from CLIP before feeding them into SCD, based on an intriguring finding that CLIP naturally capture high-quality dense features. Together, these components form a robust distillation framework that enables CLIP ViTs to integrate both visual-language and visual-centric improvements, achieving state-of-the-art results across various open-vocabulary dense prediction benchmarks." }, "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": [ "Self-distillation; CLIP; Open-vocabulary dense prediction" ] }, "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/a5a579b871219820f18d259dcedc27c48b62edeb.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": "Refining CLIP's Spatial Awareness: A Visual-Centric Perspective" }, "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've mentioned my questions in the weaknesses section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper is mostly focused on experimental evaluation of different training strategies, and their experiments are well-detailed and reproducible. \n* The methodology proposed is well presented and easy to understand and follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a few techniques that promote faster convergence in learning modular addition with encoder-only transformers. The techniques include a slight modification of loss function, angular embedding of inputs and modifications of training distribution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I've split my concerns into major and minor ones:\n\n### Major concerns:\n* I don't understand why solving modular addition in scale is important. The other papers that the authors have cited and compared their work to use the setting of modular addition as a means of studying different behaviours of training algorithms or models. The authors mention that it is important in cryptography literature, but they never elaborate on how \"learning to solve modular addition\" with the given inputs is an important task. If we have the angular embeddings, or the integers, or even one-hot embeddings, then solving the task is straightforward. \n\n* **Same setting having different results:** In table 7, the numbers of the bold row (N=20, q=257) are different from the numbers in the first row of Table 8. Don't these represent the exam same setting in running experiments? If so, where is the discrepancy coming from? This setting appears in other tables with other (different) numbers as accuracy as well, which is confusing.\n\n* Section 5.4: If I understand correctly, Figure 5 claims to depict the PCA visualization of the outputs. IIUC, the targets are the angular embeddings of modular sums, the output dimension is 2. I don't see why PCA is needed here, since the output dim is already 2. Furthermore, when MSE is low, it's clear that the outputs must correspond to the angular embeddings of the targets and must be distributed on a circle, and when MSE is high they should not. I don't see how this tells us anything about the internal workings of the model.\n\n* Overall, I think the techniques proposed require a practitioner to know about the structure of the problem (that we're going to solve a modular addition problem) and are not general beyond modular arithmetic. On the contrary, when we know that we're dealing with a modular addition problem, there are far superior approaches to solve the task than learning a deep network.\n\n### Minor concerns:\n* IIUC, Mohamadi et al's claim regarding the need for a fraction of data only applies to the so called \"kernel-regime\" where the network is not allowed to do any feature learning, and doesn't apply to trained networks. \n* For the cryptography use case that the authors have mentioned: does partial correctness (achieving non-trivial but also not 100% evaluation accuracy) matter in the mentioned use case? If not, how can one ensure 100% evaluation accuracy on a given task?" }, "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": "- Is there a cryptanalytic application where a transformer implementing modular arithmetic, or something close to it, would be preferable to simply calling highly-optimized and accurate modular arithmetic routines?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The techniques used do improve performance on this problem, sometimes drastically, and indeed escape the symmetry-based lower bounds of Mohamadi et al. (2024) by using non-uniform sampling and representations which are not permutation-equivariant. This aligns somewhat with the results of Abbe et al. (2024).\n\nSome of the analyses of the impacts of different decisions in the training process are quite interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper designs an architecture, representation, and dataset to use to train an encoder-only transformer model to perform modular addition of a fixed number of addends modulo a fixed prime." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper presupposes that it is interesting to train an ML model to perform modular arithmetic in order to get good performance. I would vehemently argue, despite the existence of several recent paper which do train ML models to perform modular arithmetic (many of which I do think are quite interesting and with whose details I am very familiar), that this is not of any interest whatsoever. Here is a function far more interesting to cryptanalysis for this task: `lambda q, nums: sum(nums) % q`. This function achieves 100% accuracy for any `N` and `q`, probably runs _many_ orders of magnitude faster than your trained model with _far_ less memory, and doesn't require 240 GPU-hours of training.\n\nSo why is there so much recent work on training ML models to do modular arithmetic? This is _because_ it's such an easy problem, where we can understand what the network is doing when, e.g., exhibiting grokking behavior, or thinking about curriculum design, etc. The focus of these papers is not on obtaining the best learned model, but on what the process of learning on this toy problem can tell us about learning in general.\n\nThus, a paper about obtaining the best ML model to do modular arithmetic seems entirely misguided to me. A paper using modular arithmetic as a case study to investigate problems like curriculum/training distribution design, out-of-distribution generalization, etc could potentially be very interesting! There are a few parts of this paper that touch on things along these lines, and indeed the decisions about representation, the training distribution you use, etc are intriguing. But they're in service of a useless problem. I would suggest instead taking the kinds of decisions you made here to get things to work as an idea to explore in more general cases, taking modular arithmetic as a test case, rather than trying to get the best modular arithmetic network." }, "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": "- What about using active learning/sampling to generate training data? \n- The evaluation uses test data from a particular distribution (uniform). This is standard. But things can be different in applications. What if the test data (ie motivated via the cryptanalysis application mentioned in the intro) have a different distribution? How to adjust the techniques?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The work investigated in detail the existing training methods on the problem, identified potential drawbacks, and proposed corresponding techniques to address them.\n- The work provided empirical evidence that the proposed changes can help." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work considered learning modular addition via transformers at scale and proposed three changes to the modular addition model training pipeline for this purpose: 1) diversifying the training data; 2) an angular embedding, 3) a new loss function. The work showed that these changes lead to improvement for learning at scale, scaling up to N=256 elements modular q=3329. It also showed that these techniques generalize to other modular arithmetic problems (a few specific low degree polynomials modular q)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The problem addressed is quite limited: scaling up for a specific problem of modular addition tested over uniform distirbution. While the work provided some motivation, it is still unclear what's the impact of the work for future research/applications.\n- It is unclear if the technical contributions are significant. The changes proposed are natural and not surprising. Furthermore, although the work tested on a few other modular arithmetic problems, those problems are specific and the evaluation is quite preliminary. It is unclear if the techniques can help for more general learning settings eg other algebraic reasoning 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. **Local Minima at the Origin**: You mention that the model can converge on local minima like the origin of the unit circle, which hinders learning. Since the correct output for a label \\( x \\) would be represented as \\( \\cos(2\\pi x / q) \\) and \\( \\sin(2\\pi x / q) \\), could you clarify why the origin (0,0) acts as a local minimum in this context? It would be helpful to understand how this specific point prevents effective training, given the angular nature of the embeddings.\n\n2. **Digit-wise Tokenization for Modular Addition**: Have you experimented with digit-wise tokenization methods, such as representing numbers as sequences of digits, to evaluate how the model performs on modular addition tasks? It could provide insights into the model's ability to generalize on addition when individual digits are tokenized.\n\n3. **Comparison with Interpretability-focused Work**: In Table 2, many of the related works primarily address interpretability aspects rather than modeling improvements for modular addition. This focus makes direct comparison potentially less relevant. Could you elaborate on why these specific interpretability-focused works were chosen, and consider whether it might be beneficial to compare primarily with approaches that directly aim to enhance modular addition capabilities?\n\n4. **Comparison with Other Embedding Techniques**: Given that you propose a new embedding and custom loss, it would be helpful to see how it compares with existing methods designed for modular arithmetic or general embedding approaches, such as abacus embedding (https://arxiv.org/abs/2405.17399) or dice embedding (https://aclanthology.org/2020.emnlp-main.384.pdf). Have you tried these methods, and if so, how did they perform relative to your angular embedding? This comparison could add further depth to your evaluation of embedding strategies in modular arithmetic tasks." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "This paper’s key strengths lie in its innovative methodology and rigorous validation. The angular embedding and specialized loss function introduce solutions directly tailored to the demands of ML-based modular arithmetic. As modular arithmetic is foundational in cryptography, this work could help drive advancements in ML-powered cryptanalysis. The methodological rigor is enhanced by detailed ablation studies, and the inclusion of visualizations like PCA plots adds clarity, reinforcing the paper's accessibility and value." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper tackles the challenge of enabling machine learning models, specifically transformers, to handle modular arithmetic with significantly larger values for \\( N \\) and \\( q \\) than previously studied. Traditional ML models struggle with modular arithmetic, particularly with parameters like \\( N = 6 \\) and \\( q = 1000 \\). This work proposes three key modifications that together enhance the performance of transformers on modular addition tasks:\n\n1. **Enhanced Training Data Diversity**: By including a mix of simpler and rare modular addition examples, the authors aim to help the model generalize effectively.\n2. **Angular Embedding**: This technique maps integers onto a unit circle, aligning better with the periodic nature of modular arithmetic.\n3. **Custom Loss Function**: The authors introduce a specialized loss function designed to prevent convergence on local minima, ensuring that the model learns effectively.\n\nThese methods enable the transformer-based model to achieve high accuracy on modular addition tasks with values up to \\( N = 256 \\) and \\( q = 3329 \\), significantly surpassing prior results. The approach also shows potential for generalization across other modular arithmetic functions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the proposed data distribution and loss modifications are effective, they add complexity. Discussing potential simplifications or alternative approaches for less resource-intensive implementation would be beneficial. While the model performs well up to \\( N = 256 \\) and \\( q = 3329 \\), addressing potential limitations as these parameters increase would add further depth." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We make several improvements that enable transformers to do modular arithmetic with large moduli (up to 3329) and many terms (up to 256)" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024teaching,\ntitle={Teaching Transformers Modular Arithmetic at Scale},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=38hLpTVpe7},\nnote={under review}\n}" }, "abstract": { "value": "Modular addition is, on its face, a simple operation: given $N$ elements in $\\mathbb{Z}_q$, compute their sum modulo $q$. Yet, scalable machine learning solutions to this problem remain elusive: prior work trains ML models that sum $N \\le 6$ elements mod $q \\le 1000$. Promising applications of ML models for cryptanalysis$\\textemdash$which often involve modular arithmetic with large $N$ and $q$$\\textemdash$motivate reconsideration of this problem. This work proposes three changes to the modular addition model training pipeline: more diverse training data, an angular embedding, and a custom loss function. With these changes, we demonstrate success with our approach for $N = 256, q = 3329$, a case which is interesting for cryptographic applications, and a significant increase in $N$ and $q$ over prior work. These techniques also generalize to other modular arithmetic problems, motivating future work." }, "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": [ "transformers", "modular arithmetic", "math" ] }, "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/e2d48df0ee22753041ca0443c9e7fce2d3523e1d.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "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": "Teaching Transformers Modular Arithmetic at Scale" }, "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": "See the Weakness section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The proposed LEASE algorithm is novel, which introduces a new way to handle the challenge of limited preference data with a learned transition model. The selection mechanism for unlabeled data based on confidence and uncertainty is a thoughtful contribution to improving the stability and accuracy of the reward model.\n\n-Theoretical Framework: The paper attempts to provide a theoretical foundation for the algorithm, which is a step towards more principled offline PbRL methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel algorithm for offline preference-based reinforcement learning (PbRL) aimed at addressing the challenges of designing rewards and the high costs of online interaction. The LEASE algorithm leverages a learned transition model to generate unlabeled preference data, which is then filtered through an uncertainty-aware mechanism to ensure the performance of the reward model. The paper claims to provide a generalization bound for the reward model and a theoretical improvement guarantee for the policy learned by LEASE. The experimental results are said to demonstrate that LEASE can achieve comparable performance to baseline methods with fewer preference data without online interaction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the paper provides a theoretical analysis of the algorithm, it is not rigorous enough. For example, the approximation error of the reward function usually depends on a condition number that is exponential to $R_{\\text{max}}$ when learned from preference data [1], but it seems missing from the derived bound. The approximation error of reward error does not directly translate into an additional error term in the performance bound and requires careful treatment (e.g., use a pessimistic reward function [2]). The authors use a handwaving argument (i.e., the law of large numbers) to derive (A.24) from (A.23), but it is not accurate. Using a concentration inequality is necessary in the finite sample case.\n\n- The paper lacks some benchmarks and baselines to validate the effectiveness of the proposed method. For benchmarks, The D4RL benchmark is known to be insensitive to the accuracy of the reward function [3], and adding benchmarks like Meta-World would greatly strengthen the paper. Also, there are some recent works on offline PbRL that have a strong performance like [4,5], and LEASE should be compared with them.\n\n\nReferences\n\n[1] Pacchiano, Aldo, Aadirupa Saha, and Jonathan Lee. \"Dueling rl: reinforcement learning with trajectory preferences.\" arXiv preprint arXiv:2111.04850 (2021).\n\n[2] Hu, Hao, et al. \"The provable benefits of unsupervised data sharing for offline reinforcement learning.\" arXiv preprint arXiv:2302.13493 (2023).\n\n[3] Li, Anqi, et al. \"Survival instinct in offline reinforcement learning.\" Advances in neural information processing systems 36 (2024).\n\n[4] Kim, Changyeon, et al. \"Preference transformer: Modeling human preferences using transformers for rl.\" arXiv preprint arXiv:2303.00957 (2023).\n\n[5] Zhang, Zhilong, et al. \"Flow to better: Offline preference-based reinforcement learning via preferred trajectory generation.\" The Twelfth International Conference on Learning Representations. 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": 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 addition to Weaknesses 1-3, we have following questions.\n1. Some details of the algorithm are missing. In Line 6 of Algorithm 1, which policy is used together with the transition model to generate the data? In Line 8, the “reward update condition” is not clear. Additionally, there lacks an explanation for the “the reward model only update once instead of updating constantly in this process” statement.\n2. Can Theorem 1 show that using augmented data is better? When $N_u=0$, it is clear that the constant term increases, but the pseudo-labeling error becomes zero. Is it possible that the bound is even tighter with no augmentation?\n3. How well is the learned transition model? A bad transition model can lead to poor augmentation quality.\n4. In Figure 3, “the linear relationship between reward predicted by LEASE and ground truth is better” is not very clear. What is the possible reason that FRESH’s predictions are very narrow?\n5. Some minor problems:\n(1) In the “Model-based Reinforcement Learning” part in section 2, model-based RL is not always offline, and the presentation is a bit.\n(2) A ‘tilde’ is missing for $N_u$ in Equation 9." }, "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-organized, with the methodology and theoretical contributions presented clearly, making the core ideas easy to understand.\n2. Improving sample efficiency in offline PbRL is an important and challenging problem, with significant implications for many real-world applications.\n3. The paper provides contributions both in empirical algorithm development and theoretical analysis. Although there are concerns regarding the assumptions, the theoretical contributions add valuable insight into understanding the role of reward models and policy performance in PbRL.\n4. In addition to benchmark scores, the paper includes experiments that evaluate performance with varying amounts of preference data, as well as an analysis of the relationship between the learned reward model and the ground truth. This helps in understanding the effects of the different components of LEASE." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the problem of improving sample efficiency in offline preference-based reinforcement learning (PbRL). The proposed LEASE algorithm aims to address this issue by generating synthetic unlabeled segment pairs using an ensemble of learned transition models. These synthetic pairs are subsequently labeled with an ensemble of pre-trained reward models, followed by a filtering process that ensures the quality of the pseudo labels. The filtering mechanism employs a confidence principle, which requires that the models have high certainty in discriminating between segment preferences, and an uncertainty principle, which stipulates low variance in the predictions of the ensemble models. An offline RL algorithm can then be employed to learn on the augmented labeled dataset. The paper also supports its claims with theoretical analysis, providing an upper bound on the reward model's error learned on an augmented dataset and a bound on the policy improvement. The empirical results on the D4RL benchmark demonstrate that LEASE achieves comparable performance to baselines while using less preference data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed LEASE algorithm closely follows the pipeline of Surf[1], with only two major differences: (1) Surf augments data using random temporal cropping, whereas LEASE generates synthetic data with a learned transition model; (2) Surf only uses confidence for label filtering, whereas LEASE employs both confidence and uncertainty principles. Despite these differences, a direct comparison with Surf is missing, which would help clarify the effects of the introduced transition model and the uncertainty principle on the final performance. Without this comparison, it is challenging to establish the uniqueness or superiority of LEASE.\n2. The main results in Table 1 compare LEASE with URLHF, a previous baseline that uses more data than LEASE. However, it is crucial to also include a comparison with URLHF using the same amount of data as LEASE. This would allow readers to determine whether LEASE truly achieves superior performance with fewer data or if the perceived gains are simply due to the different quantities of data used. The current results are not sufficiently convincing without this comparison.\n3. The theoretical analysis relies on assumptions that may be unrealistic in practical settings, and the connection between theory and the empirical algorithm is weak. Specifically:\n - Assumption 2: Given a fixed learned reward model, it is possible to construct an adversarial unlabeled dataset such that the pseudo-labeling error \\(\\eta\\) becomes very large. More detailed analysis is needed to understand whether the specific data generation process of LEASE can mitigate such worst-case scenarios effectively.\n - Filtering Mechanism: While it is intuitive that filtering low-quality data improves labeling accuracy, there is no clear theoretical justification for why the proposed filtering mechanism (confidence + uncertainty) is superior to previous methods that only use confidence.\n\n[1] Jongjin Park, Younggyo Seo, Jinwoo Shin, Honglak Lee, Pieter Abbeel, and Kimin Lee. Surf:\nSemi-supervised reward learning with data augmentation for feedback-efficient preference-based reinforcement learning. In 10th International Conference on Learning Representations, ICLR 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": 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 weakness part." }, "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 of preference efficiency is vital and fundamental in offline meta-RL.\n2. The proposed method is sound and sensible.\n3. Experiment results are convincing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel model-based offline RL algorithm that improves the efficiency of utilizing limited preference data. LEASE utilizes a learned transition model to rollout data, and label preferences with confidence and uncertainty measures. LEASE can achieve high performance with as few as 100 queries on mujoco tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of analysis on the effect of transition model accuracy. The learned transition model can be inaccurate, and accumulate errors during rollout. Will this do a lot of damage to algorithm performance?\n2. Lack of analysis of baseline algorithms' performance with different numbers of preference data. How much data is needed for baseline algorithms to achieve comparable performance to LEASE?\n3. I would like to see results of more baseline algorithms, e.g., previous model-based offline RL algorithms." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024lease,\ntitle={{LEASE}: Offline Preference-based Reinforcement Learning with High Sample Efficiency},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=38kLrJNwaM},\nnote={under review}\n}" }, "abstract": { "value": "Offline preference-based reinforcement learning (PbRL) provides an effective way to overcome the challenges of designing reward and the high costs of online interaction. However, since labeling preference needs real-time human feedback, acquiring sufficient preference labels is challenging. To solve this, this paper proposes a offLine prEference-bAsed RL with high Sample Efficient (LEASE) algorithm, where a learned transition model is leveraged to generate unlabeled preference data. Considering the pretrained reward model may generate incorrect labels for unlabeled data, we design an uncertainty-aware mechanism to ensure the performance of reward model, where only high confidence and low variance data are selected. Moreover, we provide the generalization bound of reward model to analyze the factors influencing reward accuracy, and demonstrate that the policy learned by LEASE has theoretical improvement guarantee. The developed theory is based on state-action pair, which can be easily combined with other offline algorithms. The experimental results show that LEASE can achieve comparable performance to baseline under fewer preference data without online interaction." }, "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": [ "preference-based reinforcement learning", "sample 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/d6fbdc6f11cec95660865c5830ef43e9b1c1a310.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/5261634b1dc75e5ed12c0ad15269e442c9b1bc7b.zip" }, "title": { "value": "LEASE: Offline Preference-based Reinforcement Learning with High Sample Efficiency" }, "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 minor comments:\n\nThere seem to be several details missing, e.g. what implementation is used to produce the baseline for PPO; what is the actually benchmark that's being used (the paper generically cites Mujoco), etc.. (Apologies if I've missed these.) \n\nSome citations are messed up (e.g. bottom of page 3).\n\nThe paragraph starting in line 294 on page 6 is not clear." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is mostly clearly written. It proposes several reasonable, albeit well known, algorithmic components and integrates them into the PPO algorithm. It shows experimental results that suggest that these modifications can lead to improvements compared to a vanilla PPO baseline." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes several modifications to the PPO algorithm: entropy regularization, off-policy value function learning, and discounting of the state distribution. It shows experimental results that investigate the effect of these modifications and compares them to a vanilla PPO implementation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "None of the proposed modifications is novel, they have all been well studied in the literature. The paper dedicates a significant amount of space to reviewing these fairly well known ideas. I don't think that merely putting them together in a new combination is in itself a significant contribution.\n\nThe experiments are not conclusive since important comparisons to SOTA off-policy algorithms are missing. Since the paper introduces effectively an off-policy component into the algorithm (with the need to implement a replay buffer etc.), I would have really liked to see this comparison. Indeed the authors state (in the limitations) that the proposed combination of algorithmic components underperforms such existing algorithms which begs the question why one should use the combination proposed in the present paper. (NB, some off policy algorithms such as MPO also use a trust region and in that respect bear similarities to PPO.)\n\nFor this to be a strong paper I would have expected an insightful discussion why the specific algorithmic combination should be particularly useful / interesting, a demonstration that it clearly outperforms existing algorithms on relevant problems, and a detailed analysis why this is the case." }, "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 do not have any direct questions to the authors aside from those listed in the weaknesses section above. \nMy main concern with the paper is the rigor of the experimental evaluation which in addition with a lack of novelty for the suggested improvements leave me wanting for clear conclusions I would trust after reading the paper." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The adjustments made to PPO are reasonably well motivated and pull directly from the existing literature on off-policy methods. \n- There is a lack in the literature for good empirical evaluations of existing RL algorithms in fair comparisons; bridging the gap between on and off-policy methods (as done here) certainly fills part of this void.\n- The stability of PPO is of high practical relevance for varying applications from control to RLHF of large models and thus any improvements are relevant to the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper revisits on-policy RL, which is still one of the most predominant paradigms for learning controllers in simulation (or nowadays for RLHF of large models) since on-policy RL can give high quality (and minimally biased) policy improvement. The authors note that despite the simplicity of the theory underlying basic on-policy algorithms, in practice (partially due to the fact that on-policy algorithms have to trade-off optimization and exploration) they can be brittle/sensitive to hyperparameter settings.\n\nThe authors revisit the and robustify a popular on policy method (PPO) utilizing some of the insights from the recent literature on policy optimization; e.g. taking inspiration from recent results from the off-policy literature (i.e. SAC and others) such considering a maximum entropy formulation and learning an action-value (Q-function) critic instead of a state value function." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiments are unfortunately fairly limited in scope. Only 6 mujoco control domains are used and only two of them (and and humanoid) would be considered high dimensional in 2024. This limits the evidence that the paper can present for its suggested modifications seveerly. \n2. The presentation of the experiments is lacking:\n2a: A comparison to baseline PPO is presented on two domains in Figure 1 and 2. With PPO failing on the high dimensional domains. This doesn't inspire huge confidence in the results. What is causing this? Is the asymptotic performance fine and the main difference is just the speed-up from the Q-function and standard PPO would just need to run much longer?\n2b: Further Figure 3 ablates some choices of the algorithm but again seems lacking. We get no insight into which of the proposed modifications exactly makes things work. For example: how would standard PPO but with a Q-function do? It also seems like PPO without discounting could be fine on-policy (but we are missing those results here, i.e. the combination of on-policy and no discounting).\n2c: A the practical implementations of PPO for any domain with higher dimensional observations (or larger models) might consider computing the loss only on a trajectory snippet extracted from a full episode. It is unclear how that would affect e.g. the discounting.\n3. Out of the three proposed modifications two are already routinely considered in the literature/implementations: entropy regularization is a standard feature in many PPO implementations; using discounting for the 'policy gradient' loss has been considered multiple times in the literature (also partially noted by the authors) and not been consistently proven to make a big difference, so most implementations omit it. This leaves the reviewer thinking that the main contribution is to consider learning an action-value critic off-policy, but unfortunately the experiments do not properly ablate and compare this modification (see above).\n4. In many applications it is generally hard to learn an action-value critic (since conditioning on high-dimensional actions comes with it's own problems) especially when dealing with large models and or large action spaces so the algorithm here may not be generally an improvement in all cases (e.g. the situation might look very different for RLHF of large models or for experiments requiring vision 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": 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": "See above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Paper is well written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces PPO+. PPO+ aims to augment the current PPO algorithm with best-known practical practices from well-known algorithms such as SAC and TD3, as well as theoretical principles, to improve the performance and sample efficiency of PPO. These features are: 1) using off-policy data by introducing a replay buffer, 2) learning a Q-function instead of only a value function, 3) using an entropy bonus, and 4) discounting the state distribution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "--This paper presents some interesting ideas, but I think it could be strengthened by highlighting its contributions more clearly. For example, incorporating off-policy data with a replay buffer and learning a Q-function instead of a state-value function shifts the algorithm towards the off-policy RL domain. To really showcase the algorithm's effectiveness, it would be beneficial to see comparisons with established off-policy algorithms like SAC and TD3. This would provide a clearer picture of its performance within the broader context of off-policy RL.\n\n--It's also worth noting that one advantage of on-policy algorithms is their ability to learn by fitting only a value function, which can be simpler than fitting a Q-function. Introducing Q-learning in this context might add complexity, which seems to contrast with the authors' claim of increased simplicity. It would be helpful to see further discussion on this design choice and its potential implications in the context of on-policy RL.\n\n--Adding an entropy bonus is a well-established technique, having been introduced in the original PPO paper. The entropy weight is already a standard hyperparameter in most PPO implementations. More discussion on how the use of the entropy bonus here differs from standard PPO would be helpful. \n\n--Authors noted, reintroducing discounting to the state distribution doesn't yield significant performance improvements. A discussion on in which scenarios using a discounted state distribution would be beneficial would also be helpful.\n\n--Finally, the experimental results presented aren't entirely conclusive. In some domains, PPO performs better than PPO+. It's more fair to compare PPO+ with off-policy algorithms. However, as the authors mentioned, their method doesn't currently outperform SAC or TD3, despite incorporating many of the components from those algorithms. This raises questions about the specific benefits and potential advantages of the proposed modifications." }, "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 you elaborate on the performance gap between PPO+ and off-policy methods like SAC and TD3? What are the potential challenges and opportunities for bridging this gap within the on-policy framework?\n- Have you considered evaluating PPO+ on other benchmark environments beyond MuJoCo control tasks?\n- Given PPO+’s slight increase in complexity, do you have insights into how it compares in terms of training time relative to PPO, especially as task dimensionality increases?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper clearly identifies the limitations of PPO and motivates the need for a more principled approach.\n- The paper thoroughly reviews relevant literature on on-policy RL, maximum entropy RL, trust region methods, and actor-critic methods, effectively placing the proposed approach within the existing litterature.\n- PPO+ presents theoretically grounded modifications, such as leveraging off-policy data in on-policy settings, which could broaden the applicability of PPO and reduce the need for extensive tuning.\n- Experimental results, especially on MuJoCo environments, show consistent improvements over PPO, suggesting that PPO+ delivers better results in continuous control tasks.\n- The ablation studies strengthen the authors' claims, providing insights into how each enhancement (e.g., entropy regularization) impacts performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces PPO+, a new on-policy deep reinforcement learning algorithm that builds upon and improves the Proximal Policy Optimization (PPO) algorithm. The authors identify several key limitations of PPO, including sensitivity to hyperparameters and deviations from the theoretical foundations of on-policy RL. They propose solutions to address these shortcomings, resulting in an algorithm that is more principled, robust, and achieves state-of-the-art performance for on-policy methods on MuJoCo control tasks. PPO+ incorporates three major improvements: (1) correct discounting in policy gradient computation, (2) integration of off-policy data for critic learning, and (3) maximum entropy regularization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experiments are currently limited to MuJoCo control tasks. Evaluating PPO+ on a wider range of environments would provide more comprehensive proof of its capabilities.\n- A discussion of the performance gap between PPO+ and off-policy counterparts would strengthen the paper.\n- While the focus on PPO is understandable given its popularity, the paper would benefit from comparing PPO+ to other on-policy algorithms beyond PPO.\n- The authors acknowledge that optimizing for the discounted objective increases sensitivity to the choice of discount factor. While they present results for two different discount factors, further investigation into this sensitivity and strategies for mitigating it would enhance the practicality of PPO+." }, "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": "### Questions\n\n1. Could the authors provide more empirical evidence supporting that PPO+ is less hyperparameter-sensitive? Specifically, how does PPO+ perform across a range of hyperparameter settings compared to PPO across that same range? Additionally, does the introduction of new hyperparameters such as entropy regularisation, number of critics, replay buffer size now make it even harder to tune for new environments?\nWhat is the effect of the CrossQ modifications? Do you use them with the PPO baseline as well. Is there an ablation of PPO+ without using the crossQ modifications?\n3. Why was only MuJoCo used for evaluation? Would the authors be willing to extend their tests to additional benchmarks such as discrete action environments or grid-based tasks to validate the generalizability of PPO+?\n4. Could an ablation study comparing the separate backbone used in PPO+ with a shared backbone approach be added to verify that the performance gains are due to the proposed modifications and not just architectural differences?\n\n### Suggestions\n1. **Use of Evaluation Methodology**: Consider using evaluation methodology like [rliable](https://github.com/google-research/rliable) to present more statistically robust results.\n2. **Additional Comparisons**: Including at least one other on-policy algorithm (e.g., VMPO or SPO) would provide valuable context and strengthen the impact of the results.\n3. **Diversify Environment Tests**: Extending the evaluation to other types of environments and presenting results where the hyperparameters are consistent across these tests could better support the claims of reduced hyperparameter sensitivity.\n\nUltimately, I liked the paper but I think without an ablation on the shared torso i.e. using one for PPO baseline, and without a different environment suite of results, i am not willing to accept the paper. Additionally, the use of crossQ modifications concerns me as we dont fully know the interaction of these modifications. Its possible a lot of the results come from here as well. If my core concerns are addressed, I’m willing to raise my score." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. **Clarity and Methodology**: The paper is clearly written and thoroughly explains the methodology behind PPO+. It provides pseudocode and hyperparameter details, making the algorithm implementation straightforward and moderately reproducible.\n2. **Combination of Techniques**: Although the modifications themselves may not be entirely novel, the combination presented in PPO+ and its alignment with theoretical foundations is practical and beneficial for the RL community especially since its relatively easy to implement.\n3. **Ablation Studies**: The ablations are well-constructed, demonstrating the effectiveness of the individual components introduced in PPO+." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper aims to address the limitations of PPO by introducing a variant called PPO+. The authors propose several theoretically motivated modifications to reduce hyperparameter sensitivity and eliminate reliance on implementation-level tricks. PPO+ is designed to maintain the simplicity of PPO while aligning more closely with the theoretical principles of on-policy reinforcement learning. The paper evaluates PPO+ on MuJoCo benchmarks and claims state-of-the-art performance among on-policy methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Limited Comparisons**: The paper's claims of state-of-the-art performance are only made relative to PPO. Including comparisons with other on-policy methods like VMPO (https://arxiv.org/abs/1909.12238) or SPO (https://arxiv.org/abs/2402.07963) would provide a more comprehensive context and strengthen the results as I do not believe you can make this claim given only PPO as a comparison.\n2. **Backbone Usage**: PPO+ uses a separate backbone for the critic and actor networks, while PPO does not. Previous work by Andrychowicz et al. (2021) suggests that separate networks generally perform better. The lack of an ablation study comparing the shared and separate backbones in PPO+ raises serious concerns about the true source of performance gains.\n3. **Hyperparameter Sensitivity Claims**: The authors don’t explicitly claim that PPO+ is less hyperparameter-sensitive than PPO however they state it as a motivating factor for the creation of it and the paper does not provide any empirical evidence to support this. Without testing the robustness across different environments or hyperparameter settings, it seems that PPO+ doesn’t necessarily address a core motivation.\n4. **Limited Evaluation**: The evaluation is restricted to Mujoco environments. While this is a common benchmark, it is not sufficient to demonstrate that PPO+ consistently outperforms PPO or is more generalised. Testing in a broader set of environments, like grid-based or discrete action spaces, would provide more robust support for the authors' claims.\n5. **Overfitting to MuJoCo**: Given the limited environment diversity and potential over-tuning for Mujoco tasks, it is unclear if PPO+ is truly an improvement over PPO or merely a set of optimisations tailored to a specific domain." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We enhance PPO by using discounted policy gradients, off-policy data to train the critic, and adding maximum entropy bonus, while simplifying the implementation." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024revisiting,\ntitle={Revisiting On-Policy Deep Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=39JM3A3KS3},\nnote={under review}\n}" }, "abstract": { "value": "On-policy Reinforcement Learning (RL) offers desirable features such as stable learning, fewer policy updates, and the ability to evaluate a policy’s return during training. While recent efforts have focused on off-policy methods, achieving significant advancements, Proximal Policy Optimization (PPO) remains the go-to algorithm for on-policy RL due to its apparent simplicity and effectiveness. However, despite its apparent simplicity, PPO is highly sensitive to hyperparameters and depends on subtle and poorly documented tweaks that can make or break its success--hindering its applicability in complex problems. In this paper, we revisit on-policy deep RL with a focus on improving PPO, by introducing principled solutions that enhance its performance while eliminating the need for extensive hyperparameter tuning and implementation-level optimizations. Our effort leads to PPO+, a methodical adaptation of the PPO algorithm that adheres closer to its theoretical foundations. \nPPO+ sets a new state-of-the-art for on-policy RL on MuJoCo control problems while maintaining a straightforward trick-free implementation. Beyond just performance, our findings offer a fresh perspective on on-policy RL that could reignite interest in these approaches." }, "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": [ "Deep reinforcement learning", "on-policy", "policy gradients" ] }, "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/a34e12c95f64b642562b9c6f6c97dcf1d1a953bc.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": "Revisiting On-Policy Deep 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": 3 }, "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": 4 }, "primary_area": null, "questions": { "value": "See details in 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 targets a very important research problem, the time series foundation model. Because the time series data has very high variance across different domains and different tasks. How to integrate them and train one foundation model remains challenging.\n2. This paper has a very high quality of preparation. The writing, the organization, and the figure are prepared nicely and with enough details.\n3. The results shown in Tab 1, 2, and 3 are competitive compared to baseline TS models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents OTiS, a pre-trained foundation model on large-scale time series data to support multi-tasks across domains. Extensive experiments are conducted to demonstrate the powerful performance of the foundation model. This paper is prepared in high quality and can be accepted." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Add a subsection to show which category baselines will be compared. For example, traditional TS model, deep learning, TS foundation model, etc.\n2. I expect a comparison with some SOTA TS foundation models. For example, https://arxiv.org/abs/2405.02358 . If this part is added, that would be great.\n3. Currently, the author use fine-tune to adapt the pre-trained model to various downstream tasks. Can you also add one more subsection to test the prompting on this TS foundation model? That would be another great point." }, "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": "Please feel free to address any misunderstandings I've stated in the weaknesses. Answers to the following questions would help me better calibrate my score:\n1. How long does it take to finetune on new tasks?\n2. How does the finetuned model perform compared to task-specific models? Are these testing datasets really good cases that need pre-trained models?\n3. How do you get the ground truth embeddings in Figure 3?\n4. Is there any intuition around what information could be shared across such different domains to make pre-training on them useful?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper has many strengths:\n* The key idea to condition the model on different variables and domains is good. Indeed many related works effectively ignore this information.\n* The paper is overall written quite well and arguments are presented clearly.\n* The experiments investigate multiple axes, including ablation of their method and different dataset and model sizes, and visualizations of the embeddings.\n* The public data and model weights will help the community build on this work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a time series model architecture and a pre-training objective. The key idea is that their architecture acknowledges that training and testing time series may have different sampling rates and variables. The authors propose a straightforward tokenization scheme to learn embeddings for different variables, which can get added onto regular patch and temporal embeddings, thereby conditioning the predictions on the measured variables. They then pre-train their model on a collection of existing datasets, and evaluate its performance by finetuning on new datasets for some forecasting, regression, and classification datasets. They find that finetuning their model on new datasets can outperform other recent methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper has weaknesses to address:\n* The major weakness of this paper is the extremely limited experiments section. There are many experiments, yet almost no explanation of how they're run or interpretation of the results. Most of the results are written like an advertisement, mostly just stating the method outperforms others. This leaves the reader unclear why the performance gains happen. Ultimately it's not clear when/why the findings would generalize. The result is that some claims appear to be quite overstated. For example, L423-L424 states *\"embeddings of domains with shared high-level semantics cluster together, as depicted in Appendix E.1. For example, embeddings of mono and stereo audio group closely, as do those of banking and economics.\"* But this is cherry-picked---Temperature is way closer to Mono and Stereo Audio than Banking is to Economics.\n* Similarly, many important experimental details are missing or relegated to the Appendix, and the Appendix also includes almost no explanations or interpretations. For example, the PCA experiments in Figures 3, 7, and 8 aren't explained.\n* It's unclear how many variables actually overlap between training/testing, which seems to be a key element to make the model outperform others. Yet this isn't analyzed. Showing that others fail by ignoring other variables should be a key element of the experiments." }, "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": "Authors use Github Link to share the code, leading to potential personal information leakage. This may require further investigation." }, "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": 2 }, "primary_area": null, "questions": { "value": "- How does the domain-specific tokenizer adapt to unseen domains with distinct variate structures?\n\n- Additionally, how does the domain-specific tokenizer generalize across different systems within the same domain? For instance, while both electrical transformers and power generators belong to the \"energy\" domain, they exhibit differing properties and produce distinct time series readings. How does the sub-domain adaptation discussed in Section 3.1 address this scenario?\n\n- A broader question, not specific to this paper: At what level of granularity should we define the domain?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper is well-written and easy to follow. \n\n- Authors spot the important fact that the variate structure is heterogeneous across domains and this structure may represent more complex relationships. \n\n- The visualization for variate embedding seems to be interesting and insightful. \n\n- A substantial portion of this research focuses on EEG signals, which presents a novel and promising approach. The authors introduce an innovative method to model a \"specific set of systems\" that, despite being observed differently—such as TDBrain and SEED with 19 channels versus LEMON with 62 channels—remain comparable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors spot the important fact that the variate structure is heterogeneous across domains and this structure may represent more complex relationships. Thus, they propose a time series pre-training pipeline called OTiS. The OTiS is composed of a specially designed tokenizer that can add domain-specific signature to the time series and a novel loss for pretraining." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- As noted in the strengths, this work addresses the challenge of generalizing across datasets that contain time series of similar systems but are recorded differently, such as variations in sampling rates and physical values. However, the claims regarding cross-domain generalization may be overstated.\n\n- From the perspective of generalized time series analysis, the primary contribution of variate-specific embedding may not be effective in other systems where the interrelationships between variates are not as straightforward as their spatial arrangement (e.g., the electrodes in EEG as depicted in Figure 3 of the manuscript). In different physical systems, two variates may exhibit complex computational relationships (e.g., voltage and current as described by Ohm's Law), complicating the direct modeling of variates as embeddings." }, "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 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 well-written, and the method is easy to understand. The authors clearly articulate how they consider the heterogeneity of different domain time series to achieve multi-domain time series forecasting.\n\n2. This paper focuses on the problem of multi-domain time series analysis, which is crucial for building generalizable foundational models for time series.\n\n3. The experimental section utilizes a large amount of data, and the model is open-source, contributing particular engineering value to the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents OTiS for multi-domain time series analysis, building on existing pre-training paradigms for time series. It allocates domain-specific variable embeddings to distinguish the heterogeneity of different variables across domains and enhances the model's ability to learn temporal causal relationships through a dual-masking strategy. Additionally, it introduces NCC loss to capture global patterns. Experimental results demonstrate that the proposed method achieves competitive performance in time series classification, regression, and forecasting tasks across multiple domains compared to SOTA methods. Visualization results further highlight the effectiveness and interpretability of the domain-specific variable embeddings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper mentions that one of the challenges of cross-domain time series models is the significant differences in temporal dynamics and sampling frequencies among different domains. However, the paper uses the same patch size for all domains when dividing patches, failing to accommodate the unique sampling rates of different domains. This oversight means the paper does not sufficiently consider the differences in sampling rates across domains. Additionally, using a shared patch projector to encode the temporal dynamics within each patch does not adequately address the differences in temporal dynamics between domains. While this approach may be common in previous works, it does not consider the temporal heterogeneity among domains.\n\n2. The method of considering variable heterogeneity through learned variable embeddings is not uncommon. In spatiotemporal prediction, some methods [2][3] have already employed learnable embeddings to explicitly distinguish heterogeneous spatiotemporal patterns by learning time-specific and space-specific parameter spaces.\n\n3. [1] proposed using textual descriptions to label different time series domains for cross-domain time series forecasting, utilizing a channel-independent strategy. In contrast, the domain-specific variable embeddings in this paper correspond to a channel-mixing strategy. I look forward to seeing a comparison between these two strategies in cross-domain time series.\n\n4. The experimental section lacks details about the baselines. How were these methods selected? Were they pre-trained and fine-tuned? If so, what data was used for pre-training and fine-tuning?\n\n5. How does the performance of the proposed method compare to conventional time series classification or forecasting methods trained on a single specific dataset?\n\n[1] UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecasting, WWW, 2024\n\n[2] Heterogeneity-Informed Meta-Parameter Learning for Spatiotemporal Time Series Forecasting, KDD, 2024\n\n[3] Adaptive Graph Convolutional Recurrent Network for Traffic Forecasting, NeurIPS, 2020" }, "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. Have you tried to pre-train separately according to different domains and then fine-tune it for domain-specific downstream tasks? As observed from Table 1, there are several discrepancies in different domains, such as the frequencies of Economics and EEG. Is it possible that separating datasets to pre-train domain-specific models works better?\n2. The proposed method uses a fixed context for pre-training. Padding a large pre-training corpus, which generally contains univariate time series, into a fixed temporal/variate dimension. Will it cause a waste of computing resources?" }, "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 researches an important question about generalizable time series understanding across diverse domains.\n2. This work presents a large pre-training corpus, which can be greatly beneficial if the datasets are released.\n3. The method exhibits promising results in handling multivariate time series analysis by leveraging variate and domain signatures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents OTiS, a deep model pre-trained on a large corpus (11B) for general time series analysis. In this paper, the authors highlight the challenge of heterogeneity when applying self-supervised pre-training on time series. An MAE-style pre-training method is adopted to obtain a general tokenizer for multivariate time series, and then different task heads are introduced to complete time series analysis tasks. The model demonstrates strong performance across 15 diverse applications, including time series classification, regression, and forecasting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My major concern is about the novelty of the proposed method: The design of the encoder/decoder is very identical to MAE. Is there any adaptation for the time series modality? For example, considering the inherent reconstruction difficulties of time series and adjusting the mask ratio compared with the vision modality?\n2. About the model design towards generalizable time series understanding: As the authors mention an important challenge of heterogeneity, I am slightly unconvinced that a shared unified patch embedding/projector can reflect different semantics among variates and domains, even if the patch is completely the same. Prior to this, Moirai adopted different patch sizes for different frequencies, will it further enhance OTiS?\n3. This work adopts learnable embeddings as variate/domain signatures. I am convinced that the signatures can \"distinguish\" them, but how can they explicitly \"capture inter-variate relationships\"? This approach may also limit the generalization scope as the learned signatures do not apply to unseen variates/domains during inference.\n4. About the experiments: Results of classification are not compared with supervised, trained deep models, for example, TimesNet and ModernTCN. For the regression rask, can you introduce some variate-centric models into this baseline, such as iTransformer? As for forecasting, the average improvement does not seem significant compared with PatchTST. Also, can you provide some explanations about Table 3 why OTiS has a significant improvement on some datasets (such as ETTh2) and a great degeneration on similar datasets like ETTh1?\n5. A minor suggestion: the name \"dual masking strategy\" can be somewhat overstated to me, which generally refers to dual or antagonistic behavior (e.g., minimax). I would prefer to simplify the contribution as a \"mixture\" (of masking modeling and generative modeling in this paper), which is a common technique in fact. Also, I would like to know how the ratios (25% - 75% in this paper) of the two strategies are determined.\n6. The pipeline of using the masked pre-trained models seems still somewhat tedious, i.e., lacking in generalization. Supervised training should be performed after large-scale pre-training. Can the author provide an overall promotion compared with training from random initialization, or try zero-shot generalization on downstream tasks?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024towards,\ntitle={Towards Generalisable Time Series Understanding Across Domains},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=39n570rxyO},\nnote={under review}\n}" }, "abstract": { "value": "In natural language processing and computer vision, self-supervised pre-training on large datasets unlocks foundational model capabilities across domains and tasks. However, this potential has not yet been realised in time series analysis, where existing methods disregard the heterogeneous nature of time series characteristics. Time series are prevalent in many domains, including medicine, engineering, natural sciences, and finance, but their characteristics vary significantly in terms of variate count, inter-variate relationships, temporal dynamics, and sampling frequency. This inherent heterogeneity across domains prevents effective pre-training on large time series corpora. To address this issue, we introduce OTiS, an open model for general time series analysis, that has been specifically designed to handle multi-domain heterogeneity. We propose a novel pre-training paradigm including a tokeniser with learnable domain-specific signatures, a dual masking strategy to capture temporal causality, and a normalised cross-correlation loss to model long-range dependencies. Our model is pre-trained on a large corpus of 640,187 samples and 11 billion time points spanning 8 distinct domains, enabling it to analyse time series from any (unseen) domain. In comprehensive experiments across 15 diverse applications - including classification, regression, and forecasting - OTiS showcases its ability to accurately capture domain-specific data characteristics and demonstrates its competitiveness against state-of-the-art baselines. Our code and pre-trained weights are publicly available at \\url{https://github.com/OTiS-official/OTiS}." }, "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": [ "Time Series Analysis", "Multi-Domain", "Self-Supervised 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/e4d2cf724e8953bd4834c764d48045f0827da18e.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/662380416ac92716a8d258d31a6b8c8e293e316d.zip" }, "title": { "value": "Towards Generalisable Time Series Understanding Across Domains" }, "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": 3 }, "primary_area": null, "questions": { "value": "- question 1: can you present a bit more concise takeaways from your benchmark, to me I feel like I was reading a lot of pieces, and no surprising results to me either. It might be good to have some failure cases to support your point\n\n- question 2: when you evaluate the prompt complexity, how do you choose the prompt formats?\n\n- question 3: Do you think complex prompts and instructions might need manyshots?\n\n- question 4: Do you think you can add some reasoning axes to your benchmark? \n\n- question 5: maybe consider adding some long-context recent models with SOTA results, not only looking at model parameter counts but also architectural differences." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- First benchmark focusing on long-form generation during the test time\n- The evaluation combines both complexities of evaluation prompts and different scenarios\n- First batch of results on 10 mainstreamed LLMs\n- The paper is easy to follow" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces LongGenBench, a benchmark for measuring LLMs' capacities, especially their long-context abilities, by generating long-form context from 16k to 32k with rather complex instructions. This new dataset departs from traditional benchmarks aiming at decoded length in four different scenarios. Preliminary evaluations are done with main streamed LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I am a little bit distracted from the main takeaways from the experimental studies, and not so convinced with failure cases. See question 1\n\nI have other minor concerns regarding the experiment setup\n\n- There has been much research showing that the prompt format matters, what's your thought? \n\n- Reasoning tasks are not well involved, as o1 seems to argue that longer decoded length is helpful with reasoning complex tasks, in your benchmark, you might want to add an axis of reasoning ability clearly or have some analysis around this topic?\n\n- Most of the evaluation focused on existing transformer-based architecture, but models are presenting SOTA results, for example, mamba-based models. Are those models, with good inference time complexity, good at benchmark, or if not, why?" }, "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": "* How do you split the long generation and match them to all the subtask instructions?\n* How do you check if every sub-instruction is satisfied? is it by prompting another LLM or by word matching, etc.?\n* What's the significant difference between STIC-1 and STIC-2? Looks they just have difference denominators. I don't quite get it although there is a paragraph in Sec. 2.4 as below for this. Is there any specific case where an LLM can get low STIC-1 while high STIC-2, or the other way around?\n\n> STIC-1 is primarily concerned with the completion rate of instructions that result in sub-scenarios,\nfocusing on whether instructions are correctly executed. In contrast, STIC-2 assesses the overall completion of the specific instruction task, including the presence of sub-scenarios and their completion\nstatus." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* This paper is overall clear and easy to understand.\n* This proposed evaluation is novel, and the generation ability it benchmarks is not covered by previous metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Existing LLMs' long-context benchmarks rely on understanding tasks. This paper proposes a benchmark targeting specifically on long-context *generation* ability of LLMs. The authors design 4 long-context instruction-following tasks, up to 16 or 32K tokens: (1) Writing Diary for a year; (2) Wrting menu for a year; (3) Design a 100/300-floor skyscraper; (4) Plan an urban layout. As a result, all existing LLMs don't work well on these tasks of the benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Some of the details are possibly missing or hard to get by readers -- see \"Questions\".\n* In the proposed benchmark, the way to form long content is to pile short answers to many sub-queries, while the sub-tasks are actually independent, to a large extent. For example, given all the demands on one-year dairies, it should be easy for LLMs to write a diary if it is assigned a specific day of the year, while this benchmark just require the LLM generate 365 diaries all at once. In this case, the challenge of this benchmark might majorly be forgetting the instruction under the influence of generated content, instead of keeping the conherence and content interaction among generated long content. That latter should be the one mostly desired by the community." }, "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. It seems that the evaluation metrics are designed for these scenarios. If there are new scenario tasks, do we need to update the evaluation metrics?\n2. What other aspects of long text generation with LLM do you think need to be evaluated? It seems that your evaluation is more oriented towards some planning tasks or well-structured text. Is it difficult to evaluate the creation task? For example, it is difficult to design evaluation metrics for novel writing?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Interesting task design, which can evaluate the long text generation ability of large models from a certain perspective\n2. The paper is well written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces LongGenBench, a benchmark for evaluating large language models' (LLMs) ability to generate long-form text while adhering to complex instructions. It features tasks in four scenarios with different instruction types and lengths (16K and 32K tokens). The study finds that even advanced LLMs struggle with long-form generation, particularly as text length increases, highlighting the need for improvements in model architecture and training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The types of task scenarios are relatively limited, and it is impossible to comprehensively evaluate the long text generation capabilities of large models.\n2. The evaluation metrics seem to be customized according to the scenario.\n3. Limited number of models evaluated" }, "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": "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": "Would an IFEval [1] style setting where the outputs or attributes of the outputs could be verified definitively using code checks be a better option for such a benchmark? For long generations, getting the model outputs to match specific output patterns in the prompt is a challenge unto itself.\n\n[1] https://arxiv.org/abs/2311.07911" }, "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 is the first attempt at long-context generation, requiring models to generate a long text that follows a combination of specific instructions as opposed to just answering questions pertaining to long prompts.\n2. The benchmark does uncover a setting that seems to be challenging to SOTA models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a benchmark to evaluate models' strength in long-span generation tasks. It constructs synthetic examples from a fixed set of scenarios and templates in three modes. The resultant instruction measures models' abilities to faithfully follow position-specific, range-specific, and periodic instructions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper is very sparse on details regarding the evaluation of correctness. How is matching and parsing done w.r.t. the templates? A model could generate several outputs matching the criteria\n2. The benchmark is limited to a few domains and scenarios, and the paper's contributions seem quite limited overall" }, "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": 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": 2 }, "primary_area": null, "questions": { "value": "* In L460-468, it is mentioned that there are significant repetitions in long-form generations. Are these repetitions correct with respect to the given instructions, or are they semantically wrong (i.e. violating the given instructions)? The former indicates that it's probably caused by how instructions in the prompt are designed, while the latter means that model struggles at producing long and meaningful generation." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper is the first to study long-form generation as opposed to long-context generation. The perspective is novel, interesting, and of practical value to unleash the potential of LLMs for more complicated tasks.\n* Problems in the benchmark are constructed in a sound and intuitive way. While evaluating the quality of long text snippets is usually challenging and complex, the smart design in this paper enables reliable and accurate evaluation for long-form capability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new benchmark for long-form generation where the model-generated content, rather than the input context, is long. Specifically, the model is asked to roll out a detailed description of certain tasks such as diary over a year or floor plan for a skyscraper, subject to a few specific instructions that can be either singular or recurrent. Evaluation metrics include main task completion to test whether generation follows the expected format, as well as the success rate for specific instructions at both micro and macro level. Results demonstrate that the proposed benchmark is much more challenging than needle-in-the-haystack tasks for long context evaluation. Models generally do not perform very well within their claimed context lengths." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One of my major concerns is the clarity of writing in the evaluation part.\n* The definition of STIC-1/STIC-2 isn't quite clear. Using the notation $T_S=(T_{S_1}, T_{S_2}, \\dots)$ in Sec 2.3, my best guess is that STIC-1 means the average success rate over $(T_{S_1}, T_{S_2}, \\dots)$, while STIC-2 counts the entire $T_S$ as successful only if all $(T_{S_1}, T_{S_2}, \\dots)$ are successful, and gives 0 score otherwise.\n* The abbreviation **CR** in Table3 isn't defined anywhere, though I can guess this is likely the Completion Rate in Main Task Completion.\n* The terms \"main task\", \"subtask\", \"instruction task\", \"specific task\" are used in a confusing way. It would be very helpful to unify them with clear definitions, and to associate the terms to symbols like $T_S$ or $T_{S_1}$.\n* Missing x-ticks in Figure 2.\n\nApart that, there are a few technical details that are unclear to me.\n* How do you determine whether the generated text at one specific point satisfies all task requirements? For example, given the generated diary of a particular day, how do you verify that all required activities (e.g. wedding, vacation, and etc.) are covered? I would imagine that a natural language inference (NLI) module should be involved but it's not mentioned in the paper.\n* In Table 3, though the length to be evaluated at is 16K/32K respectively, the actual generation length seems only around half of the max length. How are the 16K/32K defined?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024longgenbench,\ntitle={LongGenBench: Benchmarking Long-Form Generation in Long Context {LLM}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3A71qNKWAS},\nnote={under review}\n}" }, "abstract": { "value": "Current benchmarks like ``$\\textit{Needle-in-a-Haystack}$'' ($\\textit{NIAH}$), $\\textit{Ruler}$, and $\\textit{Needlebench}$ focus on models' ability to understand long-context input sequences but fail to capture a critical dimension: the generation of high-quality long-form text. Applications such as design proposals, technical documentation, and creative writing rely on coherent, instruction-following outputs over extended sequences—a challenge that existing benchmarks do not adequately address. To fill this gap, we introduce $\\textit{LongGenBench}$, a novel benchmark designed to rigorously evaluate large language models' (LLMs) ability to generate long text while adhering to complex instructions. Through tasks requiring specific events or constraints within generated text, $\\textit{LongGenBench}$ evaluates model performance across four distinct scenarios, three instruction types, and two generation-lengths (16K and 32K tokens). Our evaluation of ten state-of-the-art LLMs reveals that, despite strong results on $\\textit{Ruler}$, all models struggled with long text generation on $\\textit{LongGenBench}$, particularly as text length increased. This suggests that current LLMs are not yet equipped to meet the demands of real-world, long-form text generation. We open-source $\\textit{LongGenBench}$ to promote comprehensive evaluation and improvement in this critical area, with code and data available at ${anonymousurl}$." }, "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": [ "Long context LLMs; Long-form generation; Benchmark" ] }, "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/a865cd333adc3d3d651b4b45f1c5eb76abc16791.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/fa5e8cc402fd05efef96deea52e04b175017dc0e.zip" }, "title": { "value": "LongGenBench: Benchmarking Long-Form Generation in Long Context LLMs" }, "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": 3 }, "primary_area": null, "questions": { "value": "1. The structure of the whole model of proposed approach is not clearly mentioned, like how many HAL layers were inserted. Providing a global view of the model pipeline will make it clearer and easier to follow. \n2. In line 422, HAL is placed after the cross-attention layer. If HAL is after this layer, how does it influence the already computed cross-attention?\n3. In line 201 and Figure 3a, \"classification accuracy improves as the number of diseases increases.\" How should this conclusion be interpreted, given that a higher number of diseases might exacerbate distribution imbalance?\n4. How is the hyperparameter alpha for the high-pass filter set to 8? From Figure 4, performance appears to still improve as alpha increases.\n5. How was the decision made to train for 39 epochs, while the generalization assessment plots the training/validation curve for 20 epochs?\n6. The baseline without the HAL layer is not reported, which could illustrate the influence of the HAL layer on the model." }, "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 novel insights into visual and textual biases important for report generation and attempts to reduce them.\n- It offers a thorough examination of these biases and their impact on model performance.\n- The clear problem definition contributes to a more focused discussion in the MRG field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the issue of low frequency (global) visual and textual biases in report generation caused by visual imbalance, textual imbalance, and skewed distribution of disease labels. The authors propose a novel approach called the High-Frequency Amplification Layer (HAL), which uses DFFT on the time axis and feature axis, followed by masking to perform high-pass filtering. This method emphasizes high-frequency components in the feature and may reduce biases. The paper includes experiments on the MIMIC-CXR and IV X-ray datasets to demonstrate improved performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some experimental settings are unclear; further explanation would improve clarity.\n- Although the implementation of the HAL layer (DFFT on the time and feature axes) is simple, this layer requires further comprehensive evaluation and ablation studies." }, "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": "My questions are listed above in the “weaknesses” section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper introduces an approach for improving the quality of medical report generation, which is a high-impact problem with potential for positively impacting the field of medicine.\n2. The authors demonstrate that their proposed approach HAL leads to performance improvements over several existing methods in this domain." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors conduct an examination of visual and textual biases in medical report generation (MRG) datasets. The analysis find that global patterns, such as normal regions and findings, contribute to visual and textual biases. These biases make MRG models prone to frequency bias, where global patterns are prioritized and local patterns (e.g. abnormal findings) are ignored. In order to mitigate this issue, the authors propose an architectural modification in the form of a high-frequency amplification layer (HAL), which aims to enhance a model’s perceptiveness to fine-grained details. HAL reduces biases, leading to improved performance in MRG tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Inadequate evaluations for demonstrating the utility of HAL**: The key claim of this paper is that the proposed method HAL improves robustness of medical report generation models, which may struggle to learn fine-grained abnormal findings. However, this claim is not sufficiently evaluated in Section 7, and as a result, it is unclear if HAL is improving robustness to these biases. Only aggregate performance values are reported on the MIMIC-CXR and IU datasets.\n\n a. Does HAL improve report generation performance (i.e. NLG and CE scores) when there is a single abnormality in the image? What about multiple abnormalities? Does HAL improve report generation performance when findings are small in size? Does HAL reduce performance on normal cases? All of these questions are critical for determining whether HAL mitigates biases as claimed, but none of these are evaluated. \n\n b. Additionally, in order to demonstrate the usefulness of HAL, Tables 1 and 2 could benefit from an additional ablation using the exact same experimental setup but without the novel HAL layer. \n\n c. In Tables 1 and 2, I recommend that the authors use more recently-developed (standard) report generation metrics for evaluating report quality with respect to factuality, such as RadGraph-F1 [1] or RadCliQ [2].\n\n2. **Inadequate evaluations for demonstrating the existence of visual and textual bias in report generation datasets:** The evaluations in Section 4.1 show that a classifier $f_{Z|X}$ trained on the images demonstrates lower performance when abnormalities occupy small regions. Similarly, a classifier $f_{Z|\\hat{Y}}$ trained on generated reports demonstrates lower performance when there are more normal samples in the training data. These results show that the classifier $f$ picks up on several biases, but how do these experiments relate to the report generation task that is the focus of this work? Do report generation models learn these same biases? It is unclear to me why classification models are the focus of this analysis. \n\n3. **Presentation issues:** There are several presentation issues in this manuscript.\n\n a. First, the notation provided in Section 3.2 is overly convoluted and unclear; for instance, how can the value of an image or text report be set to 0 or 1 (Lines 136-138)? What is meant by positive and negative in this context? This notation also seems unnecessary, since most of this notation is never referenced again in the manuscript. \n\n b. Additionally, section 4.2 is critical to this paper yet does not include adequate implementation details in the main text to understand the goals of the experiments, with most of this material being relegated to the appendix instead. For instance, details on the classification task, classification model, dataset, etc. are not provided in the main text, making it difficult to understand the problem setup. \n\n[1] Delbrouck et al. \"Improving the Factual Correctness of Radiology Report Generation with Semantic Rewards”.” 2022.\n\n[2] Yu et al. “Evaluating progress in automatic chest X-ray radiology report generation.” 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": 3 }, "primary_area": null, "questions": { "value": "I have listed some questions in the “Weaknesses” section. Below are a few more.\n\nLine 267:\n\nWhy is $T$ the first dimension of $A$? I think it should be $N$ because the $U$ is $N \\times |d|$.\n\nLine 412 and figure 5:\n\nIt is unclear what “neurons” refer to here. Is it the output of the attention layer or MLP layer, or something else?\n\nFigure 4: \nHow is accuracy calculated here?\nSince the loss keeps decreasing for larger $\\alpha$, have the authors considered increasing the $\\alpha$ beyond 8?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The work presents a novel perspective on medical report generation, identifying bias towards low-frequency regions as a challenge for learning good visual representations. The authors introduce the problem with clarity, providing evidence for the correlation between signal frequency and performance. Using Fourier transforms to filter out low-frequency regions from an image is an interesting solution to the problem. The efficacy of the method is backed by empirical results: the model trains better and achieves comparable performance with the SoTA. Overall, there seems to be potential both for mitigating visual and textual bias as an area of research and this specific method for doing so." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper identifies transformer models’ bias towards low-frequency regions of an image as a potential source for their low performance on the medical report generation (MRG) task. It provides evidence for visual and textual bias, where larger abnormal regions and the number of diseases in a study leads to a higher F1 score of the generated report. Since most of an image is normal, models are biased towards classifying images as normal. The paper addresses this issue by proposing a high-frequency amplification layer (HAL) in order to filter out low-frequency regions. It demonstrates that models trained with HAL learn more discriminative representations of diseases, among other benefits, which leads to comparable performance on natural language generation (NLG) and clinical efficacy (CE) metrics to the state-of-the-art (SoTA)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Table 1:\n\nAlthough HAL achieves performance that puts it in the top 3, ultimately its F1 is still more than 8 points lower than the SoTA, which brings into question its advantage over PromptMRG. It would be interesting if the performance gain from HAL composes with gains from other methods. For instance, would RGRG + HAL result in better performance than just RGRG alone? Therefore, the authors should include a comparison with a simple transformer that does not use HAL in order to quantify the effect of HAL on model performance.\n\nTable 2:\n\nAs the authors themselves noted, this comparison is unfair because the baseline models were evaluated zero-shot on IU-Xray while HAL was trained. The authors should provide a fairer comparison, perhaps by also evaluate zero-shot a model with HAL trained on MIMIC-CXR but not IU -Xray.\n\nLine 130-131:\n\n> Each medical image is paired with a corresponding medical report… indicates the size of the vocabulary.\n\nThe notation is weird here. Why does $Y = [y_1, \\cdots, y_t, \\cdots, y_T]$ belong in $\\{0, 1\\}^{|v|}$? What does this set, $\\{0, 1\\}^{|v|}$ refer to?\n\nLine 133-138:\n\nI think it is unnecessary to use math notation here to talk about positive and negative samples. It does not add clarity to the explanation. For example, the notation $|X^{(z)}|$ does not give the reader any more information about how the size of an abnormal region is calculated.\n\nFigure 3a (left) is very hard to read. I cannot figure out which bar has the score of 0.50. Furthermore, although the discussion of textual bias is interesting, it is left unaddressed by the paper as it focuses on visual bias." }, "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. There should be more explanations about why clinical efficacy is lower than the SOTA model. The paper aims at utilizing HAL to capture more abnormal regions, but the CE metric for detecting abnormalities was not been improved. \n2. Can HAL be applied in other existing MRG models? More ablation studies of HAL should be conducted to demonstrate its effectiveness." }, "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 defines visual and textual biases, highlighting their impact on MRG model performance.\n2. The empirical analysis of visual and textual biases confirms the presence of each bias and demonstrates the existence of the frequency bias.\n3. The work introduces a high-frequency amplification layer to amplify high-frequency signals, enabling improved detection of abnormal features.\n4. The paper provides a thorough experimental analysis, including ablation studies and qualitative comparisons, to substantiate the effectiveness of the proposed methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper identifies the challenge of visual and textual biases in automated medical report generation, which stems from the overwhelming presence of normal features in both medical images and reports. The authors define visual bias and textual bias, associating these biases with *frequency bias*, where models tend to emphasize low-frequency (normal) signals over high-frequency (abnormal) signals. To counter this, they propose the High-Frequency Amplification Layer (HAL), designed to heighten the model’s sensitivity to abnormal (high-frequency) details, thus enhancing diagnostic accuracy. Validation on MIMIC-CXR and IU X-ray benchmarks shows HAL’s effectiveness through various analyses and demonstrates competitive or superior performance compared to state-of-the-art models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Many of the experiments are conducted to demonstrate the presence of visual and textual biases, but the experimental details are not clearly articulated. For example, how many samples were utilized to analyze the visual bias, and what is the text classifier? Moreover, most of the figures for analysis should have more explanations (e.g., Figure 3, Figure 9 and Figure 10). It is not clear right now.\n2. It is not clear where HAL applied. Were they adopted in all cross-attention layers? \n3. The paper lacks sufficient novelty, as it only combines HAL. However, it does not adequately explain the results of the baseline model. Exactly how HAL works in the final report generation should be further enhanced with examples of generated reports." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024debiased,\ntitle={Debiased Medical Report Generation with High-Frequency Amplification},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3AAXabeZPG},\nnote={under review}\n}" }, "abstract": { "value": "In recent years, automated medical report generation (MRG) has gained significant research value for its potential to reduce workload and prevent diagnostic errors. However, generating accurate radiology reports remains challenging due to the prevalence of normal regions in X-ray images and normal descriptions in medical reports. Despite various efforts to address these issues, the definitions of visual bias and textual bias remain unclear and there is still a lack of comprehensive analysis of how these biases affect model behavior. \nIn this work, we rigorously define and conduct an in-depth examination of visual and textual biases inherent in MRG dataset. Our analysis emphasizes that global patterns, such as normal regions and findings, contribute to visual and textual bias. Further, we discuss how these biases make MRG models especially prone to frequency bias, where models tend to prioritize low-frequency signals that capture global patterns, while neglecting high-frequency signals. To debiase the frequency bias, we propose the high-frequency amplification layer (HAL), aimed at enhancing the model's perceptiveness to fine-grained details. Our extensive experiments show that by amplifying high-frequency signals, HAL reduces both visual and textual biases, leading to improved performance in MRG 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": [ "Medical Report Generation", "Debiased Generation", "Visual Bias", "Textual Bias", "Frequency Bias", "Fourier Transform", "High-pass Filtering" ] }, "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/65983378dedc4702c4dd55a45bde72867c2e720a.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": "Debiased Medical Report Generation with High-Frequency Amplification" }, "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": "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": "How can we evaluate whether the synthesized tabular data has general applicability for downstream tasks?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+. The authors introduced a new fidelity metric based on Wasserstein distance to evaluate diverse data types, addressing the heterogeneity and high dimensionality of tabular data.\n+. The authors introduced the membership disclosure score as a novel privacy metric effectively addresses the limitations of existing privacy metrics and enhances the understanding of privacy risks in data synthesis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper reviews the current state of tabular data synthesis, an approach that balances data utility with privacy. Despite numerous proposed algorithms, a comprehensive comparison of their performance is lacking due to the absence of standardized evaluation metrics. The authors critique existing metrics and propose new ones focusing on fidelity, privacy, and utility. They also introduce a unified tuning objective that enhances the quality of synthetic data across different methods. Extensive evaluations on eight synthesizers and twelve datasets reveal insights that guide future research on privacy-preserving data synthesis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-. The related work section of the paper is relatively weak and should be systematically organized to provide a more comprehensive introduction of relevant studies.\n-. My major concern is whether the synthesized tabular data can maintain usability for more complex downstream applications. In other words, is this usability specific to certain downstream tasks, or can it apply to any downstream task?" }, "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": "W1-W4" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1. Extensive experiments are conducted to examine the performance of each tabular data synthesizer w.r.t. the three metrics adopted in this paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the performance of tabular data synthesizers w.r.t. three metrics, fidelity, privacy, and utility. The authors conduct extensive experiments to compare SOTA tabular data synthesizers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. The technical depth of this paper is limited. All the metrics are already proposed by existing work and the authors mainly conduct an empirical comparison among tabular data synthesizers. The novelty of this paper is not clear.\n\nW2. This paper discusses tabular data synthesizers. However, it seems to me that the three metrics also fit general data synthesizers. What is the unique feature of tabular data that requires the adoption of the three metrics in model evaluation? If the authors cannot elaborate on the connection between the three metrics and tabular data, the motivation for adopting the three metrics will be unclear.\n\nW3. Since the authors adopt three metrics, the tabular data synthesis task becomes a multi-objective optimization problem. What are the relationships among the three objectives? Do they contradict each other? Is it possible to maximize the model performance w.r.t. all three metrics? The authors should provide an in-depth analysis of these issues.\n\nW4. For fidelity, why only consider the marginal distribution (definition 1)? If we only consider marginal distributions, the complex relationships among columns in a table may be ignored. Note that for tabular data, we may have some (approximate) functional dependencies among columns, which are very important for data integrity and challenging to capture for tabular data synthesizers." }, "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": "**MDS**\n\n1) Def 2 isn't well defined:\n- H is supposedly sampled \"at random\" from the dataset, but the distribution of this sampling isn't defined.\n- please replace the expectation's subscript with $H \\subset D \\setminus \\{x\\}$, or clarify that the expression $x\\in H$ is an additional requirement to $H \\subset D$.\n- what is $\\mathcal{M}$ in this definition? Can it be _any_ distance? Any particular property it should have?\n\n2) I have several reasons to think that the membership disclosure score (MDS), as defined in Eq. 8, is a very poor choice:\n- It's important to note that MDS, as defined in Eq. 8, is just an estimate, and it gives no formal guarantees. For a privacy metric, this is troublesome, and I highlight one counterexample to its reliability below.\n- Here's an example where MDS suggests high privacy, but where an attack is trivial. Consider a synthesizer $s(\\cdot)$ that maps a point as follows $s: x \\mapsto -x$ . It's trivial to see how an attacker can achieve 100% accuracy. Yet, suppose the nearest neighbor to $x$ in the real dataset is $x+\\varepsilon$. Then MDS would be proportional to $|d(x, s(x)) - d(x, s(x+\\varepsilon))|$, which can be made arbitrarily small with $\\varepsilon$. Hence MDS is a metric which can be tricked, and this makes it unsuitable for any serious privacy application. NOTE: I noted that in Appendix C you acknowledge possible drawbacks, but seem to dismiss them. Unfortunately, these are _critical_ issues even for less contrived synthesizers: for example, privacy metrics are routinely used to ensure that synthesizer implementations are bug-free, and this is certainly something that MDS cannot be trusted to do. \n- MDS' value is (potentially) unbounded, and it's unclear how its value can be matched to the risk of successful attack. Note that the two main ways of measuring privacy in this context both offer this: 1) DP (its parameters can be mathematically matched to the risk of MIA) and of course 2) running a (potentially worst-case) MIA attack directly tells us this.\n- Finally, an important drawback is that MDS doesn't really capture the worst-case: it takes the average (expectation) across multiple runs of the generator. This may be fine, but it should be carefully motivated.\n\nI strongly recommend using a conventional metric (e.g., risk against state of the art MIA attack), which is empirical (similarly to MDS), but provides a better interpretation and it is well-understood by the security community.\nTogether with this MIA metric, I recommend also including a metric with theoretical guarantees; DP parameters $(\\varepsilon, \\delta)$ would be the most standard choice for this.\n\n**Utility**\n\nThe authors introduce \"Machine Learning Affinity\" (MLA) as a metric, which is defined as the average difference across various models of the performance of a model that uses training or synthetic data. This feels incremental: most works on synthetic data generation already look at the difference between the performance (e.g., (Jordon et al., 2021)), and looking at the average across models looks like the natural next-step. I would recommend downtuning the claim that this metric is novel." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Well-written, and evaluates many methods against many datasets\n- Wasserstein distance seems like an appropriate choice for measuring fidelity, and it is adequately justified; similarly, MLA and QueryError seem good measures for utility.\n- The paper offers some interesting takeaways: statistical methods work best for privacy applications, while diffusion models offer good fidelity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper offers an opinionated selection of methods to evaluate synthetic data generation methods for tabular data. In particular, they select metrics to evaluate fidelity (Wasserstein distance), empirical privacy (they introduce the Membership Disclosure Score), and utility (via Machine Learning Affinity and QueryError). The paper goes through a very extensive set of experiments, where it compares synthesizers based on these metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed privacy metric, MDS, has critical flaws that make it unsuitable (and, potentially, misleading) for evaluating privacy. [See below]\n- The use of Wasserstein distance for synthesizer isn't exactly new; for example, Singh et al. [1] used it in their optimization objective. Similarly, as argued below, MLA is incremental.\n- Besides evaluating synthesizers with an opinionated (and well-justified, in some cases) approach, this paper feels quite redundant and it's unclear what is the \"delta\" from prior work. From a quick search, there's dozens of synthetic data evaluation frameworks [2-6], and it's unclear why a new one is needed. I appreciated your comparisons between metrics, provided in the appendix, but the main question is: can you empirically demonstrate that one would be wrong in using one of the prior frameworks, and that they should use yours instead?\n\n[1] Singh Walia, Manhar. \"Synthetic Data Generation Using Wasserstein Conditional Gans With Gradient Penalty (WCGANS-GP).\" (2020).\n\n[2] https://github.com/schneiderkamplab/syntheval\n\n[3] https://github.com/Vicomtech/STDG-evaluation-metrics?tab=readme-ov-file\n\n[4] Qian, Zhaozhi, Rob Davis, and Mihaela van der Schaar. \"Synthcity: a benchmark framework for diverse use cases of tabular synthetic data.\" _Advances in Neural Information Processing Systems_ 36 (2024).\n\n[5] Livieris, Ioannis E., et al. \"An evaluation framework for synthetic data generation models.\" _IFIP International Conference on Artificial Intelligence Applications and Innovations_. Cham: Springer Nature Switzerland, 2024.\n\n[6] McLachlan, Scott, et al. \"Realistic synthetic data generation: The ATEN framework.\" _Biomedical Engineering Systems and Technologies: 11th International Joint Conference, BIOSTEC 2018, Funchal, Madeira, Portugal, January 19–21, 2018, Revised Selected Papers 11_. Springer International Publishing, 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": [ "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 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 paper is very thorough. Given the extensive appendix, I imagine it has gone through one or more review cycles before. Nevertheless, the authors clearly have discussed in details a lot of very reasonable concerns about their approach, which I quite appreciate. \n2) The paper is quite topical and there aren't that many similar papers out there." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present an evaluation framework called SynMeter to evaluate generative modeling approaches for tabular data across 3 different dimensions: i) fidelity, ii) utility, and iii) privacy. The authors introduce reasonable metrics to evaluate algorithms/datasets along these 3 dimensions. The authors then evaluate several SOTA tabular data generation algorithms using SynMeter." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Presentation: \na) The authors try to cram in too many things into the paper. All figures are too tiny. I recommend moving some of the figures to the appendix if make the remaining ones bigger. \nb) The appendix needs better structure. I'd recommend moving the experiment details ahead of discussion of limitations. \n2) Technical points: \na) The MDS metric is designed for the synthesis algorithm whereas others are for a specific synthetic dataset. This is a major inconsistency. \nb) MDS only captures privacy against MIA attacks. This is not unreasonable but please spend a few lines explaining why you chose to only focus on MIAs?\nc) One of the references [1] is quite similar to this paper in scope but uses slightly different metrics. Given the similarity, I would love to see the authors discuss the key differences between the papers and areas of novelty. \n\n\n\n[1] Qian, Zhaozhi, Rob Davis, and Mihaela van der Schaar. \"Synthcity: a benchmark framework for diverse use cases of tabular synthetic data.\" Advances in Neural Information Processing Systems 36 (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": 3 }, "primary_area": null, "questions": { "value": "Since the tuning objective includes the same metrics used for evaluation, isn’t the observed performance improvement in Table 1 simply an expected result? Can this performance improvement truly be considered a reflection of the tuning objective’s effectiveness?" }, "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 effectively identifies and addresses the limitations of existing metrics for fidelity, privacy, and utility, highlighting the need for the proposed metrics.\n2. By introducing fidelity, privacy, and utility as core evaluation dimensions, the paper offers a well-rounded framework for assessing synthetic data quality.\n3. The proposed metrics are thoroughly validated through extensive experiments on a large number of datasets, demonstrating their robustness and applicability in various contexts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper critically examines the limitations of existing evaluation metrics and introduces new metrics for fidelity, privacy, and utility, establishing a comprehensive framework for assessing tabular data synthesis. Additionally, it proposes an integrated tuning objective that consistently optimizes data quality across different synthesizers. The study demonstrates that recent advancements in generative models significantly enhance tabular data synthesis performance while also highlighting key challenges, such as privacy risks and performance disparities among synthesizers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is a lack of experimental comparison between the proposed fidelity and utility metrics and existing metrics. For example, including case studies where the proposed metrics and existing ones yield different evaluations on the same model could have strengthened the paper's claim of improved fidelity and utility assessments. \n2. While the proposed privacy metric is an innovative approach, its reliance on numerous shadow models and synthetic datasets could lead to high computational costs. This complexity might render the metric impractical for large datasets, as the evaluation process could require substantial time and resources, limiting its usability in real-world applications." }, "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": "(also see weaknesses)\n\n- Could you come up with an experimental setup and results which would compellingly show why the Wasserstein-based fidelity metric is strictly better than other deployed methods?\n- Why are MIAs against tabular data synthesis not well understood? \n- There indeed does not exist one MIA effective across all synthesizers, but this does not seem like a justification why MIAs are not useful? The ineffectiveness of the MIA might also just reflect little privacy leakage? \n- How does the MDS metric resolve your previously raised concerns regarding MIAs? To my understanding, you are in fact proposing a new MIA, but not evaluating it as such. \n- Could you implement the MIA developed by Houssiau et al, and explain why the MDS metric is superior to compute the MIA performance for records identified by Meeus et al? \n- Could authors clarify why the query error should be part of the utility and not part of the fidelity evaluation?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Comprehensively evaluating synthetic data generators is an important problem, and the paper provides a systematic, multi-dimensional evaluation framework to do so. \n2. The paper includes considers many datasets and synthetic data generators, and comparing them across metrics is valuable for the research domain as a whole.\n3. Proposes a way to pick hyperparameters across a multiple dimensions. \n4. Authors make the framework publicly available as a tool for people generating synthetic data" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new framework, called SynMeter, to assess (tabular) synthetic data generators. They focus on three dimensions: \n\n- Fidelity: \n\t- Authors argue the need for a faithful and universal metric\n\t- They propose a Wasserstein distance-based metric to evaluate complex, high-dimensional tabular data distributions\n- Privacy\n\t- Authors argue syntactic privacy scores to not be adequate\n\t- Authors argue that existing MIAs are ineffective, as they are not well understood and no MIA is effective against all synthesizers. \n\t- They propose a new metric called membership disclosure. \n- Utility\n\t- Authors state that the traditionally used ML efficacy is not adequate, as they argue that there is no consensus on which evaluator should be used. \n\t\t○ They propose two new metrics: ML affinity and query error. \n\nThe paper then includes a holistic tuning objective as a combination of all metrics, to be used for hyperparameter selection. \n\nFinally, the paper includes comprehensive experiments evaluating all metric across datasets and generators." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I understand the need for a holistic and widely agreed upon evaluation framework for synthetic data generators, as a reader, I am not convinced that the metrics proposed by the authors are novel, or particularly better than previously proposed ones. I elaborate on each of the dimensions: \n\n**1. Fidelity.**\n\nWhile I find the notion of using Wasserstein distance to compute fidelity interesting, I remain to be convinced why this would be better than existing methods. \n- Could you come up with an experimental setup and results which would compellingly show why the Wasserstein-based fidelity metric is strictly better than other deployed methods?\n\n**2. Privacy.**\n\n I agree with the authors on the shortcomings of syntactic metrics, and like the example given for DCR. However:\n- I do not follow the arguments made for why MIAs are not sufficient. \n\t- Why are MIAs against tabular data synthesis not well understood? \n\t- There indeed does not exist one MIA effective across all synthesizers, but this does not seem like a justification why MIAs are not useful? The ineffectiveness of the MIA might also just reflect limited privacy leakage? \n- I do not understand what the difference is between the MDS metric and an MIA. If I understand it correctly, you are building a shadow model setup to then compute an MIA scoring function (which you then not evaluate as an MIA). You then pick the record for which you get the best distinction for this scoring function. To me this basically comes down to compute MIA performance for all records, and use the highest MIA performance as the privacy metric. \n\t- How does this resolve your previously raised concerns regarding MIAs? \n\t- Moreover, with this, it is not clear whether this is the state-of-the-art MIA. \n- Finally, in this entire discussion, I believe authors fail to mention (and implement) important related work. Houssiau et al [1] propose a new MIA which beats the one proposed by Stadler et al, and Meeus et al [2] propose a principled way to identify most at-risk records. \n\n**3. Utility.**\n\nI agree with the authors that there is no consensus in the literature on which metric should be used to evaluate the utility of the synthetic data. My thoughts:\n\t- While the exact formulation of the MLA score is, at least to my knowledge, new, I believe its novelty to be very limited. For instance, Stadler et al (in Sec. 6.3) measure utility as a decrease in ML accuracy of a model trained on real compared to a model trained on synthetic data. The only difference with the MLA metric would be the averaging across ML models and the normalization. \n\t- Similarly, the query error seems very similar to the k-way marginals fidelity approach, which has also been studied in for instance Annamalai et al. [3] \n\t- Could authors clarify why the query error should be part of the utility and not part of the fidelity evaluation? \n\n**References**\n\n[1] Houssiau, F., Jordon, J., Cohen, S. N., Daniel, O., Elliott, A., Geddes, J., ... & Szpruch, L. (2022). Tapas: a toolbox for adversarial privacy auditing of synthetic data. arXiv preprint arXiv:2211.06550.\n\n[2] Meeus, M., Guepin, F., Creţu, A. M., & de Montjoye, Y. A. (2023, September). Achilles’ heels: vulnerable record identification in synthetic data publishing. In European Symposium on Research in Computer Security (pp. 380-399). Cham: Springer Nature Switzerland.\n\n[3] Annamalai, M. S. M. S., Gadotti, A., & Rocher, L. (2024). A linear reconstruction approach for attribute inference attacks against synthetic 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": "1) How does your work relate to existing surveys on tabular synthetic data?\n\n2) Why is Wasserstein distance an appropriate metric for categorical fields? \n\n3) How do tabular transformers like RealTabFormer compare to the baselines you have evaluated?" }, "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 interesting, I think these kinds of benchmarking studies can be very useful. \n\nThe paper tackles an important problem, there is a lot of interest in generating synthetic tabular data right now. I thought some of the metrics were interesting, particularly query error, which seems to address a common use case for synthetic data. I appreciated that the MDS score was taken as a maximum over x in D, in line with best practices in the membership inference literature. \n\nThe head-to-head comparison between models is very practically useful, and could have real-world impact." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem of evaluating tabular synthetic data generation techniques. To this end, it critiques existing metrics for fidelity, privacy, and utility, and proposes new ones. Then, it evaluates these metrics on a set of 12 datasets. The authors find that diffusion models generally outperform other model types in terms of utility and fidelity, whereas statistical methods are better suited for privacy. They also evaluate differentially-private synthesizers and evaluate the effect of DP budget." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I’m a little worried that the level of the technical contribution and the findings may not rise to the level of a top-tier conference. This is particularly true given that there exist other surveys on the quality of synthetic data generators. \n\n1)\tI don’t understand why the Wasserstein metric makes sense as a fidelity metric for categorical variables. You have defined the cost matrix as infinite between classes. So how can you find a meaningful transport map? This doesn’t seem like the right metric for categorical variables. By the way, Wasserstein distance has already been used as a fidelity metric for synthetic data over a metric space (e.g., CTAB-GAN+ (Zhao et al), DoppelGANger (Lin et al)), with total variation distance being used for categorial variables. \n\n2)\tThe paper seems not to mention a number of related works, including:\n\n-\tOn the Inadequacy of Similarity-based Privacy Metrics: Reconstruction Attacks against Truly Anonymous Synthetic Data (Ganev and De Cristofaro)\n-\tOn the Quality of Synthetic Generated Tabular Data (Espinosa and Figueira)\n-\tA universal metric for the evaluation of synthetic tabular data (Chundawat et al)\n\nFor a survey paper, these omissions worried me--I’m concerned that there may be others I’m not aware of. I would definitely want to see a more in-depth literature review. How does your work relate to these, particularly the surveys on synthetic tabular data? What does your survey add to the discussion? \n\n3)\tThe evaluation seems incomplete in terms of baselines. I thought there should at least be a tabular transformer (e.g., RealTabFormer or a successor) in the mix. GReaT is a transformer, but the tokenization is not really tailored to tabular data, and RealTabFormer is (a) more widely used in practice, and (b) not compared against in the GReaT paper. Also, the tables in the evaluation are illegible—it would be nice to make them bigger." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Proposed a set of new evaluation metrics and framework for tabular data synthesis from fidelity, privacy and utility" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024systematic,\ntitle={Systematic Assessment of Tabular Data Synthesis},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3ANoEa7roV},\nnote={under review}\n}" }, "abstract": { "value": "Data synthesis has been advocated as an important approach for utilizing data while protecting data privacy. In recent years, a plethora of tabular data synthesis algorithms (i.e., synthesizers) have been proposed. A comprehensive understanding of these synthesizers' strengths and weaknesses remains elusive due to the absence of principled evaluation metrics and head-to-head comparisons between state-of-the-art deep generative approaches and statistical methods. In this paper, we examine and critique existing evaluation metrics, and introduce a set of new metrics in terms of fidelity, privacy, and utility to address their limitations. Based on the proposed metrics, we also devise a unified objective for tuning, which can consistently improve the quality of synthetic data for all methods. We conducted extensive evaluations of 8 different types of synthesizers on 12 real-world datasets and identified some interesting findings, which offer new directions for privacy-preserving data synthesis." }, "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": [ "Tabular Data Synthesis", "Privacy", "Evaluation Metric", "Generative 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/1493fa9bb40e8023cad955b6e1048bdcbf5d120b.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/3769be1ad5e564411633f8476d83f8e43a8700b3.zip" }, "title": { "value": "Systematic Assessment of Tabular Data Synthesis" }, "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": { "value": "This paper has substantial overlaps with the paper [1] on arXiv. For example, Section 3 in this paper is almost the same as Section 3 in [1], Eq. 5 in this paper is the same as (4) in [1], The figure in line 270 is the same as the one in Page 4 of [1], Algorithm 1 in this paper is the same as Algorithm in [1], and so on.\n\nAlthough this paper might be a resubmission on top of [1] by the same authors, it is still quite weird that [1] is not properly cited and compared as a related work since the experimental results in this paper are quite different from those in [1] and the main selling points of these two papers are intrinsically different: [1] highlights the multiple importance sampling method while this paper is purely based on importance sampling with a single distribution. \n\n[1] Salaün, Corentin, Xingchang Huang, Iliyan Georgiev, Niloy J. Mitra, and Gurprit Singh. \"Multiple importance sampling for stochastic gradient estimation.\" arXiv preprint arXiv:2407.15525 (2024)." }, "flag_for_ethics_review": { "value": [ "Yes, Research integrity issues (e.g., plagiarism, dual submission)" ] }, "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 (6) is only for binary classification tasks (as indicated in Line 205)? Could $J$ be larger than 2? \n- In [1], the authors mention that ``The individual parameter derivatives vary uniquely across the data points, and estimation using a single distribution inevitably requires making a trade-off\" and advocate for the multiple importance sampling (MIS) approach. Could you please comment on this? Could you experimentally compare their MIS-based algorithm with your algorithm?\n\n[1] Salaün, Corentin, Xingchang Huang, Iliyan Georgiev, Niloy J. Mitra, and Gurprit Singh. \"Multiple importance sampling for stochastic gradient estimation.\" arXiv preprint arXiv:2407.15525 (2024)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The idea of using importance sampling weights for data pruning is interesting.\n- The plots in Figure 1 numerically verify that the learned importance sampling weights are somewhat meaningful and provide intuitions of why this method could improve upon uniform sampling." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new adaptive method of importance sampling for stochastic gradient estimation in multi-class classification. The sampling weights of this method do not depend on the costly full backpropagation on each data point. The importance sampling weights can also be used for data pruning, which can be further combined with the importance sampling for gradient estimation. The authors conducted experiments on classification tasks to verify the effectiveness of their algorithm compared to SGD with uniform sampling and previous importance sampling methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- A major concern with the importance sampling method proposed in this paper is that it remains \"loss-based\". To be specific, the weight of each data $x$ is proportional to $\\\\|\\frac{\\partial \\mathcal{L}(x)}{\\partial m(x,\\theta)}\\\\|\\_2 = \\sqrt{\\sum_{j=1}^J(s\\_j(x) - y\\_j(x))^2}$, where $s\\_j(x)$ is the predicted probability of data $x$ belongs to class $j$ while $y\\_j(x)$ is the groundtruth. Thus, the importance sampling weight of data $x$ can be viewed as its $\\ell_2$ loss on label prediction. However, it is unclear how this approach relates to the theoretically optimal importance sampling weight based on gradient norms. If the gradient w.r.t. the output does not take the specific form as in the logistic loss, does it still make sense to sample based on the norm of the gradient w.r.t the output?\n- There is no formal convergence analysis of the proposed algorithm. So the algorithm remains heuristic. \n- The experiments in this paper were not repeated with different random seeds, resulting in a lack of error bars on reported values and curves. This makes their experimental results less reliable." }, "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": "1. Can you clarify the novelty of your proposed metric and why should it be better than those used in the literature?\n2. Can you add more comparisons in the experiments with more recent works ([2] and a lot more) and also the most simple random reshuffling with a reduced number of epochs?\n3. Can you add more details regarding sampling with or without replacement? It is said within a batch, it is sampled without replacement, so the samples will not repeat. But what about for different batches from the same epoch. To me, it seems like samples can repeat within an epoch, and this can lead to inferior performance compared to those will not repeat within an epoch such as [2]." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Overall, the writing is clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose to use the derivative wrt to the network output for importance sampling. They also propose that their method can be used for online data pruning. They demonstrate their performance is stronger than some of the baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are many papers in the literature regarding importance sampling and the authors seem not to know many of them. It is true that the methods proposed in the earlier years have the problem of computation efficiency, but nowadays there are many methods with little overhead.\n\n\n1. The proposed method has limited novelty. While the authors claim that the derivative wrt to the network output is different from what is used in (Katharopoulos&Fleuret,2018), from my understanding, it is pretty similar if not identical, and there are many other works that utilize something similar like the EL2N score in [1]. Even if there is some difference, the novelty seems limited and the changes are not argued or justified. (Like the difference may just be taking the derivative to the pre- or post- activation output)\n2. The experiment comparison is weak. They do not compare with some more recent work like [2]. Additionally, they do not compare with some simple and already used baseline in practice. For example, random shuffling with a reduced number of epochs is often stronger than many of these importance sampling methods. (This is sometimes much stronger than uniform sampling as the samples won’t repeat within an epoch. Also, it is important that the learning rate schedule changes so that the learning rate decays)\n\n\n\n[1] Paul, M., Ganguli, S., & Dziugaite, G. K. (2021). Deep learning on a data diet: Finding important examples early in training. Advances in neural information processing systems, 34, 20596-20607.\n\n[2] Qin, Z., Wang, K., Zheng, Z., Gu, J., Peng, X., Zhou, D., ... & You, Y. InfoBatch: Lossless Training Speed Up by Unbiased Dynamic Data Pruning. In The Twelfth 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": 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": "See 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 evaluated their proposed framework on multiple datasets and different tasks. The results look promising and show some gains compared to the competitors that seem consistent across tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work studies stochastic gradient descent with online importance sampling and data pruning. The authors propose a practical metric that requires little computation and use it both for the importance weights and pruning scores. This metric is updated on the fly during training. The authors then evaluate their framework on multiple tasks such as classification and regression and popular benchmark datasets such as Cifar10/100 and MNIST." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) I found the writing in Sections 3 and 4 to be unclear and somewhat lacking in precision; they would need rewriting and clarifications in my opinion.\nFor example, equation (1) is confusing, as it seems different from the typical loss minimized, which is generally expressed as $\\mathbb{E}(\\mathcal{L}(m(x, \\theta), y))$, where $x,y$ follow a given data generating process. I fail to understand this renormalization and what $p(x,y)$ refers to here: is it the data-generating process or importance sampling weights? Could you explain and define $p(x,y)$.\nBesides, there are some assumptions that are not clearly stated and appear here and there in the text, for example, that the model is Lipschitz (Section 4.1). \n\n2) The proposed metric seems to be the same as the EL2N proposed in Deep Learning on a Data Diet, Paul et al. Could the authors explain the difference?\n\n3) Although the proposed method seems to outperform the other methods consistently, the differences are sometimes very small, and it is difficult to know if we can not attribute it to statistical noise. The authors indicate that they average over 3 runs; it would be interesting to quantify the variability of the results." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This work introduce a novel importance sampling algorithm for machine learning frameworks improving convergence by prioritizing crucial data points through simplified importance functions based on loss derivative with minimal computational overhead." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024online,\ntitle={Online importance sampling for stochastic gradient optimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3AQAUMObuc},\nnote={under review}\n}" }, "abstract": { "value": "Machine learning optimization commonly relies on stochastic gradient descent, where the accuracy of gradient estimation is crucial for model performance. Rather than relying on uniform sampling, importance sampling can improve accuracy by focusing on data points that have more significant impact on learning. However, existing methods for importance sampling face challenges with computational efficiency and integration into practical machine learning workflows.\nIn this work, we introduce a novel adaptive metric based on the loss derivative wrt the network output that can be used for both importance sampling and data pruning. Our metric not only enhances gradient accuracy by prioritizing influential data points but also enables effective pruning by identifying and removing data that contributes minimally to training. We propose an efficient adaptive algorithm that leverages this metric with minimal computational overhead. Our evaluations on classification and regression tasks demonstrate improved convergence and reduced training data requirements, validating the efficacy of our approach." }, "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": [ "SGD", "Importance sampling" ] }, "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/89d0fe6f51ebaf471f29eb7d0ed8812f5b27ad87.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": "Online importance sampling for stochastic gradient 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": 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 authors utilize a pre-trained sentence transformer as the prompt encoder in their training process. Do the authors have any insights into how the size of the prompt encoder influences the overall performance, as the size of the prompt encoder will affect the models' ability to understand the input prompt? \n\n2. Training diffusion models often incorporate classifier-free guidance. Is the proposed method compatible with training under this manner?" }, "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 introduces a novel approach by proposing adaptive prompt-based pruning that routes input prompts to specialized pruned sub-networks based on their characteristics. This represents a difference from conventional static and dynamic pruning methods,\n2. The empirical results training on datasets like CC3M and MS-COCO demonstrate the method’s effectiveness compared to other pruning methods. The results show that the proposed method outperforms other baselines by significantly reducing computational cost while maintaining or improving output quality as measured by metrics like FID, CLIP score, and CMMD score." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces adaptive prompt-based pruning strategy to reduce the computation cost of diffusion model. The proposed approach involves encoding input prompts into architecture embeddings, which are mapped to specialized architecture codes. These codes determine the routing of each prompt to a pruned sub-network. By training a prompt router using a combination of contrastive learning and optimal transport, the proposed method ensures that prompts are dynamically assigned to appropriate sub-networks. The results of the paper demonstrate the reduction in computational cost while maintaining FID and CLIP scores." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The major concern is the empirical evaluation of the proposed method: \n\n1. as stated in the paper, most organizations typically fine-tune pre-trained diffusion models on their target data but evaluate these models on broader benchmarks to demonstrate generalizability. In this study, however, the authors only fine-tune their model on CC3M and MS-COCO and limit their evaluation to the corresponding validation sets. Expanding the evaluation to a common benchmark would better showcase the model’s generalization capabilities. Specifically, demonstrating that the prompt router can handle prompts outside the training distribution would be more convincing.\n\n2. The paper also references other model pruning methods, such as MobileDiffusion[1], SnapFusion[2], and LD-Pruner[3]. However, it does not include quantitative comparisons with these approaches. It would be helpful for the authors to explain why these comparisons were omitted.\n\n3. In efficient inference for stable diffusion, recent papers show that one-step or few-step generation can speed up the generation. This paper does not include comparisons with methods like INSTAFLOW[4], which would have provided valuable insights into how APTP compares with state-of-the-art approaches in rapid generation.\n\n\n\n[1] Zhao, Yang, et al. \"Mobilediffusion: Subsecond text-to-image generation on mobile devices.\" arXiv preprint arXiv:2311.16567 (2023).\n\n[2] Li, Yanyu, et al. \"Snapfusion: Text-to-image diffusion model on mobile devices within two seconds.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[3] Castells, Thibault, et al. \"LD-Pruner: Efficient Pruning of Latent Diffusion Models using Task-Agnostic Insights.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.\n\n[4] Liu, Xingchao, et al. \"Instaflow: One step is enough for high-quality diffusion-based text-to-image generation.\" The Twelfth International Conference on Learning Representations. 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": 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 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 idea behind the paper is technically sound and novel.\n2. The paper is well written. \n3. The authors present some interesting interepretability experiments on expert assignment that aid the proposed concepts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose prompt-based tuning of text-to-image diffusion models, in which different sub-networks within pre-trained models are trained for different prompts/concepts. They authors performe experiments on multiple datasets to show that for a given latency, their model performs comparably to higher latency pretrained models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited experiments - Although I find the proposed ideas novel, I believe that the paper lacks extensive experimentation on \n - different types of architecture - It is currently unknown if the proposed methods is generalizable across architectures (DiT, MMDiT etc). \n - Small datasets - How does the method perform when data is limited?\n - Fine grained concepts - How does their method handle expert assignment when concepts are fine-grained (breeds of different animals)\n2. Comparison to Mixture-of-Experts (MoE) models - How does the proposed method compare to other prompt-conditinal architectures like MoE text-to-image diffusion models? Currently the competitors in Table 1 (a and b) are static structural pruning baselines, but I believe the paper's contribution is prompt-conditional pruning, which demands comparison to prompt-conditional efficient architectures like MoEs like [1].\n3. I am concerned about the 4 and 7 point drop in FID of the proposed method in Table 1. The authors have not presented any trade-off between latency and performance, which would help understand how limiting computational budget affects performance\n\n\n[1] RAPHAEL: Text-to-Image Generation via Large Mixture of Diffusion Paths, Xue et al" }, "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": "- Did the authors try their approach on other architectures? even other backbones of Stable Diffusion?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The idea of pruning different parts of the network for each prompt is non-trivial and interesting.\n- Visual results show APTP seems to use various pruning modes, and does not collapse to a single one." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new approach for accelerating the sampling from diffusion models using adaptive pruning denoted as APTP. The method is tailored for a specific scenario where the inference is on a specific known data distribution and the data for it is given (e.g. a company’s private dataset). Using this data, APTP trains a prompt router module that selects which parts of the architecture should be pruned for each given prompt. The selection is from a fixed set of reduced architecture states (denoted as architecture codes). Both the prompt router and arch. codes are trained end-to-end." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**(1) Missing Baselines of Less Diffusion Steps:** My main concern is that the paper does not compare to approaches that perform less diffusion steps. More specifically the following should be considered as baselines:\n\n- *Step Skipping Distillation of Diffusion Models [1,2,3]:* As mentioned in the paper, several methods tackled faster inference of diffusion models using this idea, mostly by knowledge distillation of the model to inference with fewer (between 1-4) steps. These methods should cut the latency by 8-25 times from the original model, while the proposed APTP was shown to cut latency by much less (not even 50% of the original model timing). These approaches also require training as APTP does, and their weights are readily available for many SoTA architectures, e.g. SDXL, SD 1.5, etc. \n\n- *Caching Based Acceleration [4]:* These approaches cache features of the U-Net architecture and reuse them in later steps. Such approaches hold a distinct advantage of being training-free.\n\n- *Less Denoising Steps at Inference Time:* A trivial training-free baseline is to simply sample images with less denoising steps. I wonder how that would compare to the proposed APTP. Can APTP further accelerate such a setting?\n\nAs step skipping approaches became more popular recently, I believe including some of these as baselines is a must for such an approach.\n\n**(2) Quality of Writing (Sec. 3):** Sec. 3 is complicated and difficult to understand. Specifically, I think Sec. 3.2,3.3 are overloaded and could be shortened to a much clearer version. These subsections are the only place in the paper that describe the actual approach of APTP (and what does arch. codes mean), therefore are especially important for the proposed approach.\n\n**(3) Visual Comparisons:** The paper offers a very limited selection of visual comparisons - having only 8 qualitative comparisons to the original baseline, and no such comparisons to previous approaches. Could the authors please supply a larger set of comparisons to the original model and baselines (including a step skipping distillation [1,2,3], caching [4] and less denoising steps). \n\n**(4) Clustering of Pruning Modes:** While this qualitative analysis was promised in the abstract and introduction, it only exists in figures at the last 2 pages of the appendix without any textual description. Given it is mentioned in the abstract I think it should be included as a part of the main manuscript.\n\n**(5) Limited Setting and Empirical Results:** Unlike other approaches, the proposed method is limited to a specific data distribution. Although the task is much more confined, the reduction in latency is not substantial: To keep performance comparable to the original model in terms of FID or CLIP score, APTP can only reduce 20% of the latency (Tab.1). \n\n[1] Liu, Xingchao, et al. \"Instaflow: One step is enough for high-quality diffusion-based text-to-image generation.\" The Twelfth International Conference on Learning Representations. 2023.\n\n[2] Sauer, Axel, et al. \"Adversarial diffusion distillation.\" European Conference on Computer Vision. Springer, Cham, 2025.\n\n[3] Meng, Chenlin, et al. \"On distillation of guided diffusion models.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.\n\n[4] Ma, Xinyin, Gongfan Fang, and Xinchao Wang. \"Deepcache: Accelerating diffusion models for free.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 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": 3 }, "primary_area": null, "questions": { "value": "* I would be curious to know how big the router-model itself is (in terms of parameters and memory footprint) and, by extension, how does the size of the router model affect the maximum batch size on an A100 GPU?\n* Did you do any additional experiments concerning varying resolutions and aspect ratios and their impact on the pruned-image quality?\n* Did you try applying this technique to transformer-based image-generation models like Pix-Art-Alpha / Sigma, do you see any major hurdles regarding the switch from ConvNets to Transformers?\n* How specific is the APTP-model to its training data? How do out-of-distribution prompts impact the quality of the generations?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Good Illustrations and concise explanation\n* The core idea is an interesting combination of known concepts in a new, but highly relevant setup.\n* A good amount of comparison to other methods using multiple datasets with different image-types (COCO being primarily real-world images and CC3M being more diverse)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a mixture-of-expert-esque strategy for efficiently, which they coin Adaptive Prompt Tailored Pruning (APTP). The methods combine the benefits of dynamic and static pruning methods and archives good generative metrics while decreasing the computational cost." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* While FID, CLIP-Score and CMMD alongside the visual examples provide a good overview, I, personally, would have preferred some human-user study (which I see is a lot of effort and for this reason, would not ask from the authors). As an alternative substitute, I propose compute the pick-scores [Kirstain et al. 2023] against the pruning metrics similar to [Pernias et al 2024] on a set of diverse prompts like Partiprompts could provide additional, easily interpretable evidence of this methods' superiority in terms of generation quality.\n\nKirstain et al. 2023 https://proceedings.neurips.cc/paper_files/paper/2023/hash/73aacd8b3b05b4b503d58310b523553c-Abstract-Conference.html\n[Pernias et al. 2024] https://proceedings.neurips.cc/paper_files/paper/2023/hash/73aacd8b3b05b4b503d58310b523553c-Abstract-Conference.html" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a prompt-based pruning method for Text-to-Image diffusion Models, which prunes a pretrained model for a target task into a set of specialized efficient models for different categories of input prompts, given a target compute budget." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024not,\ntitle={Not All Prompts Are Made Equal: Prompt-based Pruning of Text-to-Image Diffusion Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3BhZCfJ73Y},\nnote={under review}\n}" }, "abstract": { "value": "Text-to-image (T2I) diffusion models have demonstrated impressive image generation capabilities. Still, their computational intensity prohibits resource-constrained organizations from deploying T2I models after fine-tuning them on their internal *target* data. While pruning techniques offer a potential solution to reduce the computational burden of T2I models, static pruning methods use the same pruned model for all input prompts, overlooking the varying capacity requirements of different prompts. Dynamic pruning addresses this issue by utilizing a separate sub-network for each prompt, but it prevents batch parallelism on GPUs. To overcome these limitations, we introduce Adaptive Prompt-Tailored Pruning (APTP), a novel prompt-based pruning method designed for T2I diffusion models. Central to our approach is a *prompt router* model, which learns to determine the required capacity for an input text prompt and routes it to an architecture code, given a total desired compute budget for prompts. Each architecture code represents a specialized model tailored to the prompts assigned to it, and the number of codes is a hyperparameter. We train the prompt router and architecture codes using contrastive learning, ensuring that similar prompts are mapped to nearby codes. Further, we employ optimal transport to prevent the codes from collapsing into a single one. We demonstrate APTP's effectiveness by pruning Stable Diffusion (SD) V2.1 using CC3M and COCO as *target* datasets. APTP outperforms the single-model pruning baselines in terms of FID, CLIP, and CMMD scores. Our analysis of the clusters learned by APTP reveals they are semantically meaningful. We also show that APTP can automatically discover previously empirically found challenging prompts for SD, *e.g.,* prompts for generating text images, assigning them to higher capacity codes." }, "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": [ "Model Pruning", "Diffusion Models", "Inference 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/fae55d05a8cf9a2535f855c4b04da54394c51551.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": "Not All Prompts Are Made Equal: Prompt-based Pruning of Text-to-Image Diffusion 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": "1. How well the LLM-powered deductive logic engine performs on standard logical deduction problems.\n2. The reported accuracy for ReClorTeam (GPT-4-0613) on the ReClor leaderboard is 90.10, which is notably different from the numbers presented in this paper. What may cause the difference?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "**Originality: 4/5**\n\nA closely related work, SatLM, uses the Z3 solver as its logical reasoning backbone, whereas LINA leverages an LLM-prompt-based approach. While both frameworks share a similar conceptual foundation, LINA’s LLM-based reasoning backbone is more adaptable to loosely defined questions, enabling it to outperform the more rigid solver approach. This novel application of an LLM-driven deductive logic engine enhances generalizability.\n\n**Quality: 3.5/5**\n\nPros: The authors provide both theoretical proofs on complexity and robust experimental results across multiple benchmarks. One question that arises is how well the LLM-powered deductive logic engine performs on standard logical deduction problems.\n\nCons: The reported accuracy for ReClorTeam (GPT-4-0613) on the ReClor leaderboard is 90.10, which is notably different from the numbers presented in this paper.\n\n**Clarity: 3.5/5**\n\nPros: Figure 1 effectively clarifies the pipeline, and the appendix, which includes the actual prompts, further aids understanding.\n\nCons: Figure 2 is challenging to interpret without sufficient context, and it’s unclear why the Chain-of-Thought (CoT) approach does not explore additional steps.\n\n**Significance**\n\nThis work is of the interest for both neural symbolic community and NLP community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose LINA, a framework that decomposes the reasoning steps for complex questions using four main components: (1) an LLM-based logic extractor, (2) an LLM-based query extractor, (3) an LLM-powered logic deducer, and (4) a core algorithm that integrates context and derived results to analyze the correctness of the underlying answers. They also provide theoretical analysis of LINA’s properties and complexity. Experimental results demonstrate that LINA significantly improves performance on benchmarks requiring multi-step reasoning, outperforming existing methods like Chain-of-Thought (CoT) by a substantial margin." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "As shown in strength." }, "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 authors propose an agentic framework that utilizes formal logic to enhance LLMs. Could a broader comparison with other relevant approaches (in addition to LINC) [1-4] be considered to provide a more comprehensive evaluation?\n\n\n2) LLMs are generally stronger in processing natural language compared to formal logic. Could the authors clarify the advantages they see in converting logical reasoning tasks from natural language into Propositional or First-order Logic for LLM-based reasoning? If this conversion strategy offers benefits, might it be more effective to prompt LLMs with Chain-of-Thought reasoning including Propositional or First-order Logic?\n\n\n3) The authors introduce an agentic framework for symbolic reasoning without an external solver. Could they explain the rationale behind this choice in more detail? If the concern is that formal logic generated by LLMs may be unreliable for external solvers, how does the proposed framework address this issue? Additionally, since the agentic approach relies on a sufficiently capable base model for sub-task management, would this framework extend well to smaller models (such as 7-8B parameters)?\n\n\n4) Given that LLMs can struggle with self-bias [5], could the authors discuss any potential limitations in having the same LLM serve as both the deductive reasoner and supervisor/judge? Are there mechanisms in place to help mitigate self-bias and enhance the model's verification process?\n\n\n5) One challenge with deduction using formal logic can be the restricted scope, especially if the required deduction rules are not explicitly included as known information. Could the authors share any strategies to address this challenge? Additionally, do they see any potential for extending this formal logic framework to reasoning tasks that require broader expressiveness, such as math reasoning, coding, and question answering?\n\n\nReference:\n\n[1] Pan, Liangming, et al. \"Logic-LM: Empowering Large Language Models with Symbolic Solvers for Faithful Logical Reasoning.\" The 2023 Conference on Empirical Methods in Natural Language Processing.\n\n[2] Yang, Sen, et al. \"Neuro-symbolic integration brings causal and reliable reasoning proofs.\" arXiv preprint arXiv:2311.09802 (2023).\n\n[3] Xu, Fangzhi, et al. \"Symbol-LLM: Towards foundational symbol-centric interface for large language models.\" arXiv preprint arXiv:2311.09278 (2023).\n\n[4] Xu, Jundong, et al. \"Faithful Logical Reasoning via Symbolic Chain-of-Thought.\" arXiv preprint arXiv:2405.18357 (2024).\n\n[5] Huang, Jie, et al. \"Large Language Models Cannot Self-Correct Reasoning Yet.\" The Twelfth International Conference on Learning Representations, 2024." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Improving LLM-based reasoning with neuro-symbolic integration is a good research problem. The writing is well-structured and clear. Empirical results are given with details. Code and data are provided for reproducibility." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces LINA, a neuro-symbolic approach designed to enhance the logical reasoning abilities of LLMs. LINA implements a hypothetical-deductive reasoning paradigm by enabling LLMs to autonomously manage logical reasoning without external solvers. It extracts propositional logic from natural language, and performs deductive logical reasoning. Empirical results show LINA outperforms existing methods, including LINC and other prompting techniques." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The core concept of the proposed approach is an agentic framework equipped with formal logic, which is relatively common. The advantages of translating natural language into formal logic and using LLMs for reasoning remain ambiguous. The effectiveness of the agentic framework is influenced by the capabilities of base model and potential self-bias. The application scope of the method is limited." }, "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) Why did you choose to use the train set of FOLIO while using the validation set for other datasets? Is there a specific reason for this decision? Most prior work (e.g., Logic-LM) typically evaluates the test set of FOLIO, so it would be helpful to clarify the rationale behind this choice.\n\n(2) Could you provide more details about how the hypothesis is generated? Additionally, could you elaborate on how the Reasoning Process Judgment is integrated into the framework? It appears in Figure 1 but is not included in Algorithm 1, which causes some confusion. Providing more information on this would make the methodology easier to follow for readers.\n\n(3) Do you have quantitative data to support your claim?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(1)\tThe framework claims to effectively address the issue of poor generalization to different question formats and the problem of information loss when using only symbolic language by combining symbolic and natural language.\n\n(2)\tThe main experiment shows that the method surpasses the baselines across five datasets using GPT-3.5 and GPT-4o." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a framework called LINA to address the generalization problem and information loss found in existing methods. The framework consists of two main components: an Information Extraction Module and a Symbolic Reasoning Module. First, the Information Extraction Module condenses and translates the reasoning question into a symbolic format. Then, the Symbolic Reasoning Module iteratively performs one-step deductive reasoning, utilizing both symbolic and natural language, with a judgment step to verify the correctness of each reasoning step. By leveraging GPT-3.5 and GPT-4o, the paper demonstrates that LINA outperforms the baselines across five datasets. Additionally, the paper includes comparisons to ToT and SatLM, along with an ablation study and a case study." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1)\tThe level of innovation in this work raises some concerns. To the best of my knowledge, some previous work (SymbCoT [3]), also addresses the issue of information loss by leveraging both natural language information and first-order logic (FOL). The key difference is that SymbCoT conducts reasoning and verification as a linear process, whereas this work transforms the process into an iterative one. In summary, it appears that this work mainly modifies the linear process from SymbCoT into an iterative framework, which limits its novelty. From my perspective, the primary innovation lies in this framework’s adaptability to a wide range of question formats, a feature the previous work lacks.\n\n(2)\tThe Information Extraction Module requires further clarification. How is the context classified? Additionally, how do you determine the \"ease of translation\"? Upon reviewing the context classification prompt provided in the appendix, it seems more focused on simplifying logical statements rather than classification. Please clarify if my understanding is incorrect.\n\n(3)\tIn Section 4.2, you explain that the context is first classified into lengthy text and non-lengthy text, with the lengthy text then being condensed into shorter sentences. These condensed texts are further classified based on their ease of translation. Further details are needed to understand this process. For example, how many classes are used in this step? Which classes will be translated, and which will not? This is important because the paper claims an advantage in using both symbolic and natural language, so it is crucial to understand what content is represented in symbolic language and what remains in natural language.\n\n(4)\tThe Reasoning Module lacks crucial details. Firstly, there is no explanation of how the deductive process works and how information LS, NL, and H interact to reach the reasoning conclusion C. Secondly, when performing the Check() operation, is it checking for errors in the reasoning process, or is it verifying whether the information contradicts or supports the hypothesis? Third, you mention that if an error occurs, the supervisor may adjust C or reset C = H. How is this step implemented exactly? This is not explained in the main text nor in Algorithm 1, and more details are needed to help readers understand how the reasoning module operates.\n\n(5)\tThe paper lacks a detailed analysis, which hinders the reader's understanding and the transparency of the framework. For example, the paper's main claim is that it addresses information loss and improves the framework's generalizability, but there is a lack of relevant analysis to support this claim. Besides, prior work in this stream (e.g., LINC [1], Logic-LM [2], SymbCoT [3]) typically includes an analysis of accuracy per necessary proof depth in ProofWriter. Including this type of analysis would be valuable, as it could demonstrate how robust your method is with respect to increasing reasoning complexity, a common challenge in real-world applications. Furthermore, the paper lacks an error analysis, which would provide a clearer understanding of where failures occur and improve confidence in the proposed framework.\n\n(6)\tThe analysis section also lacks some details. In Section 5.4, when you state that the LLM cannot generate effective z3 solver code or easily adapt for execution, does this mean that the rule-based solver completely fails to execute the problem, or can it execute but fail to reach the correct answer? Do you have quantitative data, such as execution rates, to back up this observation?\n\nReference:\n[1] LINC: A Neurosymbolic Approach for Logical Reasoning by Combining Language Models with First-Order Logic Provers (Olausson et al., EMNLP 2023)\n[2] Logic-LM: Empowering Large Language Models with Symbolic Solvers for Faithful Logical Reasoning (Pan et al., EMNLP 2023)\n[3] Faithful Logical Reasoning via Symbolic Chain-of-Thought. (Xu et al., ACL 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": "see above" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "See below" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a pure prompt-based framework that solves reasoning problems, namely LINA. The framework first prompts LLM to convert problem into formal logic representation with natural language information; then it solves the problem as a deductive reasoning task by iteratively prompt the reasoner for deducing new facts and the supervisor for verification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "## Novelty\n\nThe proposed method is a pure prompt-based framework with a straightforward design. The specific design of performing reasoning without using external tools has been studied in several prior works [1,2]. That said the novelty of this work is minor.\n\n\n## Quality\n\nThe idea of \"removing the tool usage yields better performance for deductive reasoning\" is poorly motivated and justified\n\nL48 \"First, the process of converting logical problems into formal expressions leads to information loss\"\n- This is true for problems without FOL groundtruth, such as ReClor and LogiQA which are evaluated in the experiments.\n- However, **these problems are not meant to be solved with the traditional formal logic method in the first place**, the prior work such as SatLM and LogicLM mostly focuses on solving the NLI task with datasets that come with groundtruth FOL annotations. Also note that ReClor and LogiQA contain not only deductive reasoning but also other reasoning tasks that cannot be characterized by FOL.\n- That said, criticizing translation leads to information loss is fine, but it hardly motivates the approach proposed here if it is meant to solve problems that already fall outside of the formal logic bucket.\n\nL78 \"Second, the reliance on specific external tools results in poor generalization of these methods, limiting them to solving only certain types of problems, such as FOL propositional inference problems or satisfiability problems\"\n- This statement is problematic. Many works show tool usage increases rather than decreases the capability of LLMs in solving formal reasoning problems.\n- Formal tools such as Prover9 and Z3 can be used for not only propositional logic but also first-order logic. And sat problem is a very generic problem setting where many reasoning problems can be converted into a sat problem, and being able to solve sat problem should be not considered as a disadvantage.\n- That said, the authors should motivate their work properly.\n\n\nNot every reasoning problem in ReClor and LogiQA can be formed into deductive reasoning:\n- The authors propose to solve all reasoning problems with deductive reasoning. This is simply inappropriate for many of the problems in ReClor and LogiQA. For example, ReClor contains questions like \"which of the following most challenges/supports/aligns with the argument in the context?\" and \"which of the following arguments shares the same reasoning pattern as that in the context\", such questions do not fit into any formal logic categories and certainly cannot be solved with deductive reasoning.\n\n\nThe experiment setting misses many details and is potentially problematic:\n- It's unclear how many ICL examples are used for GPT CoT baselines. However, an accuracy of 76 with GPT-4o on ReClor seems too bad to be true. As a comparison, [3] shows that with just a few ICL examples, GPT-3.5 can achieve about 60% accuracy and GPT-4 can achieve above 90% accuracy, which aligns much better with the scores reported in the public leaderboard.\n- As mentioned above, including methods like LINC in ReClor and LogiQA benchmarks is not sensible, as these methods are designed for NLI task and not these benchmarks.\n\t\n\n\n## Clarity\n\nThe paper is generally easy to follow.\n\n\n## Significance\n\nWhile I agree with the authors that moving beyond standard NLI tasks into more \"in the wild\" reasoning problems such as that in ReClor is an interesting and important direction, it cannot justify the pure prompt-based design, as it effectively rendering the approach into yet another fancy CoT method that could hallucinate during its reasoning. From a pure performance perspective, the significance of this work is still questionable as the results from the baseline approach are too bad to be true. That said, the significance is also minor.\n\n\n[1] Zhu, Zhaocheng, et al. \"Large language models can learn rules.\" arXiv preprint arXiv:2310.07064 (2023).\n\n[2] Feng, Jiazhan, et al. \"Language models can be logical solvers.\" arXiv preprint arXiv:2311.06158 (2023).\n\n[3] Yang, Yuan, et al. \"Can LLMs Reason in the Wild with Programs?.\" arXiv preprint arXiv:2406.13764 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024lina,\ntitle={{LINA}: An {LLM}-driven Neuro-Symbolic Approach for Faithful Logical Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3BoCwZFRJX},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) have exhibited remarkable potential across a wide array of reasoning tasks, including logical reasoning. Although massive efforts have been made to empower the logical reasoning ability of LLMs via external logical symbolic solvers, crucial challenges of the poor generalization ability to questions with different features and inevitable question information loss of symbolic solver-driven approaches remain unresolved. To mitigate these issues, we introduce **LINA**, a LLM-driven neuro-symbolic approach for faithful logical reasoning. By enabling an LLM to autonomously perform the transition from propositional logic extraction to sophisticated logical reasoning, LINA not only bolsters the resilience of the reasoning process but also eliminates the dependency on external solvers. Additionally, through its adoption of a hypothetical-deductive reasoning paradigm, LINA effectively circumvents the expansive search space challenge that plagues traditional forward reasoning methods. Empirical evaluations demonstrate that LINA substantially outperforms both established propositional logic frameworks and conventional prompting techniques across a spectrum of five logical reasoning tasks. Specifically, LINA achieves an improvement of 24.34% over LINC on the FOLIO dataset, while also surpassing prompting strategies like CoT and CoT-SC by up to 24.02%. Our code is available at https://anonymous.4open.science/r/nshy-4148/." }, "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", "Logical Reasoning", "Neuro-Symbolic Approach", "Hypothetical-Deductive 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/e0af48b8cfbf91fd34ecf218d8d4003aa8e2b008.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": "LINA: An LLM-driven Neuro-Symbolic Approach for Faithful Logical Reasoning" }, "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": 3 }, "primary_area": null, "questions": { "value": "Please refer to my questions above. Overall, I feel the method shows some promising results but more evaluations might be required to assess the approach." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. I think few-shot animation of human actors is an important direction and the proposed approach offers meaningful contributions to this field.\n\n2. The paper compares a range of baselines and demonstrates notable improvements in terms of the alignment to the reference image and smoothness of the actions. \n\n3. The paper is overall well organized. Ablation studies are shown for each component." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper tackles the problem of animating human actors in the few-shot setting. To address the challenges, the authors propose the FLASH framework, which mainly consists of a Motion Alignment Module and a Detail Enhancement Decoder. The effectiveness of the method is tested on 12 atomic human actions selected from HAA500." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. For the proposed approach, I wonder if it also works on general videos beyond human actors? If not, what is the specific design in the model tailored for human actors? I think it would add more value to the paper if it also works well on general motions. Since many of the baselines are designed for general motions, I think showing its generality is important and also makes the comparison more fair.\n\n2. Experiments: a) The visual quality is still limited, in the examples shown, there’s still clear object flickering and motion jittering. b) It is only tested on HAA500, with 12 actions. I think that more tests on different datasets are required to see the effectiveness of the method. As for the metrics in Table 1, some metrics are worse than some baselines and I think more explanation would be helpful.\n\n3. If given more videos, would the method still outperforms the baselines? Showing a figure that illustrates the improvements relative to the number of input videos would be helpful. This would make it clearer to understand the range in which the method outperforms others." }, "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": "The network details and training procedures are rather unclear. Please refer to the questions listed in the Weakness." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Although video generation models have achieved impressive advances, they rely on extensive training sets and computation resources. However, human video generation, especially with large body movements, is still challenging. The idea of exploiting very few video sequences to train video generation models specifically for humans is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the problem of generating videos with static images as input. Instead of having an extremely large dataset, they trained a model over a few videos(16 videos) with similar motions or actions. The key idea is to extract consistent motion patterns or features from those few videos; afterward, the human video is generated by enforcing the motion patterns as well as the appearance from the input image. They trained the models on HAA500 dataset containing several different categories of motion actions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "First, some of the technical details and network structure are not very clear. For example, 1) in Fig. 2 of the system overview, the text prompt is injected into the Unet, but from the caption of the figure and also from the description in the method part, it is still not clear how to get the text prompt fed into the Unet, and do we need to have a text-encoder, cross attention as in stable diffusion? 2) What is the encoder and structure in Fig.2? Do we train the entire network together with the encoder and decoder as well as Unet, or do we need first to train the encoder and decoder? 3) When selecting the 16 videos for training, what are the selection criteria? Is the selection done manually or automatically? 4) From the given network structure and description in the method part, it is still unclear how to encode the first/reference frame image into the network. \n\nSecond, from my understanding, a model will be trained for each text prompt. This is rather inefficient. \n\nThird, some images and visual results need to be included: 1) since augmentation plays an important role in the overall design, it is better to show some augmented images. 2) Without any video included in the supp, it is rather difficult to find out whether the temporal consistency issue exists." }, "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 you please provides failure cases of the techniques and point out certain constraints of the work. I think this part is missing." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper investigates the compelling and essential task of learning to animate human actions from a limited number of videos.\n\nThe problem is clearly explained and well-motivated, emphasizing the need to learn generalizable motion patterns without overfitting to the appearance of training videos, while ensuring smooth transitions from the initial reference image.\n\nTo address these key requirements, the authors introduce a Motion Alignment Module that aligns motion features and inter-frame correspondence relations. Additionally, to enhance transition smoothness from the reference image, FLASH employs a Detail Enhancement Decoder." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes FLASH (Few-shot Learning to Animate and Steer Humans), which improves motion generalization\nby aligning motion features and inter-frame correspondence relations between videos with different appearances. Even with limited training data, this approach minimizes the overfitting issue to visual appearances and enhances the generalization of learned motion patterns. Experiments demonstrate that FLASH effectively animates images with unseen human or scene appearances into specified actions while maintaining smooth transitions from the reference image." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The Motion Alignment Module requires generating highly augmented versions of videos to learn motion patterns across varied appearances. This process depends heavily on the quality and effectiveness of augmentations, which, if inadequate, could fail to capture essential variances in motion or introduce irrelevant features." }, "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": "- How does the strongly augmented video mechanism work? What exactly is the \"random color adjustment\" here? After performing the \"strong augmentation\", is the video still in the data domain?\n- Is the proposed model trained from scratch or initialized from some pre-trained model? If relying on a pre-trained model, what model does it exactly fine-tune?\n- What is the requirement for the training data? How much similarity do the videos of the same motion have to share? Do they need to be in a similar view (i.e., front vs side vs back)?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The problem studied in this paper is interesting, important, and challenging. Video diffusion models indeed manifest strong hallucinations in human motion, which need to be addressed to facilitate several downstream tasks.\n- The solution of using some data augmentation techniques to solve the data sparsity issue is well-motivated. Based on this motivation, the authors design some reasonable network architecture modifications to learn from those augmented data.\n- Both quantitative and qualitative metrics are shown to have a better comparison between the proposed method and the baselines. Although not beating all baselines quantitatively, the user study shows better results for the proposed method.\n- Several ablation studies have been conducted to show the effectiveness of the proposed components." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem of learning rarely-seen human motion information for video diffusion models from sparse video clips. The authors propose a \"motion alignment module\" to solve this challenging problem, where they first conduct pixel-level data augmentation and then encourage the model to reconstruct the video based on the shared motion information. Experiments are conducted with comparisons to baselines to show superiority of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The technical writing of this paper is not clear enough to fully explain their method. For example, it's not clear how the method performs augmentation and whether the augmentation is reasonable. More questions can be found below.\n- Some potential baselines are not compared. For example, MotionDirector trains a motion LoRA to adapt the video diffusion model to rarely seen motions. How will the proposed method compare to those methods?\n- Experiment results are lacking for the comparison of the proposed method and baselines on Internet videos. It is not clear whether the proposed method is also superior on generalizability.\n- The ablation studies do not show a video animation result, so it's hard to tell whether the proposed components indeed help the generation quality. Also no quantitative user study is performed for the ablation study." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning to Animate Images from A Few Videos to Portray Delicate Human Actions},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3By4N0GAdt},\nnote={under review}\n}" }, "abstract": { "value": "Despite recent progress, video generative models still struggle to animate human actions from static images, particularly when handling uncommon actions whose training data are limited. In this paper, we investigate the task of learning to animate human actions from a small number of videos---16 or fewer---which is highly valuable in real-world applications like video and movie production. Few-shot learning of generalizable motion patterns while ensuring smooth transitions from the initial reference image is exceedingly challenging. We propose FLASH (Few-shot Learning to Animate and Steer Humans), which improves motion generalization by aligning motion features and inter-frame correspondence relations between videos that share the same motion but have different appearances. This approach minimizes overfitting to visual appearances in the limited training data and enhances the generalization of learned motion patterns. Additionally, FLASH extends the decoder with additional layers to compensate lost details in the latent space, fostering smooth transitions from the reference image. Experiments demonstrate that FLASH effectively animates images with unseen human or scene appearances into specified actions while maintaining smooth transitions from the reference image." }, "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": [ "Image Animation", "Video Generation", "Few-shot" ] }, "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/c540d3696b07d0790a15a4d48cfd5a42d248ca9f.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/d9f2e2fc1aa2c8d31fd15e45a994b3b35c7b8469.zip" }, "title": { "value": "Learning to Animate Images from A Few Videos to Portray Delicate Human Actions" }, "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": "+ The taxonomy of memorization is purely established based on the property of data, i.e., the number of duplications in the pre-training corpus and the implicit template within the data. However, memorization is also a concept and phenomenon related to model behavior and model behaviour can also be included into the taxonomy as an evidence to classify different types of memorization." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ A new taxonomy for understanding and analyzing the model memorization. \n+ Interesting findings on the dynamics of memorization during the scaling-up of data and model size.\n+ An empirical evaluation of the utility of the taxonomy based on predictability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a taxonomy for model memorization and classifies model memorization into three categories, namely recitation, reconstruction and recollection. The authors identify several data-related or model-related features and test their correlation with model memorization. To verify the effectiveness of the proposed taxonomy, the authors trained three linear regression models for three categories respectively and found that group the regression models attain better performance than the predictors trained on other features." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "+ A perplexing part of the taxonomy is the classification of repetitive or incremental sequences following a specific pattern. If the sequence duplicates more than five times, how do we know whether it is truly \"memorized\" or it is reproduced simply because the LLM learns its pattern? \n+ Why do we use more than five times duplication as the decision boundary for recitation and non-recitation. How is the hyper-parameter decided? Using a single threshold actually assumes an equality in difficulty for reciting every sequence." }, "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": "What insights from previous work on memorization mechanisms support or conflict with these findings?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed memorization taxonomy is intuitive and interesting, drawing parallels with human memorization. This taxonomy is particularly valuable as it provides a structured approach to analyzing what has typically been treated as a uniform phenomenon. \n\nThe analysis methodology is another strong point, featuring a thorough examination of dependencies between features and memorization categories, supported by effective predictive modeling to validate the taxonomy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel taxonomy for memorization in language models, categorizing it into three types: recitation (highly duplicated sequences), reconstruction (predictable templates), and recollection (other memorized sequences). The authors validate their taxonomy through predictive modeling and analysis across model scales and training time, demonstrating how different factors influence each category distinctly. The work provides valuable insights into understanding memorization as a multifaceted phenomenon rather than a uniform behavior." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main weakness of this paper boils down to two key issues. First, while the idea of categorizing memorization into three types sounds cool, the paper doesn't dig deep enough to tell us why we should care. Sure, they show that code gets memorized more than text across all categories - but why? And what does this mean for how these models actually work? How different types of memorization contribute to model capabilities. These are the kind of insights that would make the taxonomy actually useful, but they're missing.\n\nIn addition, the experimental setup is not convincing. For example, the experiments are conducted solely on Pythia models without validation of other popular models. And some of the key choices seem pretty arbitrary like picking $k=32$ for their memorization tests or saying \"more than 5 duplicates\" counts as recitation. Why those numbers? What happens if you change them?\n\nOverall, I think the paper lacks insights and the experiments are not very solid." }, "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) Can you provide empirical justification for this specific cutoff? How sensitive are your results to this choice?\n2) Could you include statistical significance tests for the reported trends across model sizes?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper exhibits several notable strengths that demonstrate its potential value to the field. The proposed taxonomy of memorization provides an intuitive and practical framework for understanding different types of memorization in LLMs. The extensive experimental validation across model scales and training time points offers valuable insights into how memorization behavior evolves. The authors' approach to validating their taxonomy through predictive modeling and dependency analysis shows methodological rigor and provides empirical support for their theoretical framework." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This comprehensive paper presents a novel taxonomic analysis of memorization in LLMs, breaking it down into three distinct categories: recitation (highly duplicated sequences), reconstruction (inherently predictable sequences), and recollection (neither duplicated nor predictable). Through extensive experimentation with the Pythia models ranging from 70M to 12B parameters, the authors demonstrate that different types of memorization exhibit distinct patterns and dependencies on factors like sequence duplication, model size, and training time. They validate their taxonomy by showing its effectiveness in predicting memorization likelihood and reveal that recollection grows disproportionately with model size and training time. The work provides valuable insights into how different factors influence memorization depending on the taxonomic category." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The template detection approach appears oversimplified. For instance, only basic patterns of \"repeating\" and \"incrementing\" sequences are considered, potentially missing more complex templates. The duplication counting relies on exact matches without accounting for semantic similarity or near-duplicates (e.g. slightly modified code or text passages).\n2) The paper insufficiently compares its taxonomy against existing memorization frameworks. For example, the relationship between these categories and counterfactual memorization, which is mentioned but not analyzed, deserves exploration. The advantages of this taxonomy over other approaches to studying memorization are not quantitatively demonstrated.\n3) The exact procedure for computing KL divergence in Fig 3 is unclear, and the methodology for computing perplexity scores used throughout the analysis lacks essential details. The robustness of results to different tokenization choices is not evaluated." }, "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. How would the predictive model perform without continuation perplexity as a feature? This would help assess how much signal the other identified factors provide to predict memorization.\n2. How novel is the introduction of statistically validated taxonomy? Do similar case studies exist in other fields? Beyond the reference to the Simpson paradox, the paper doesn't include references in this area.\n3. Are corpus statistics computed on the prompt, continuation or the full sequence? \n4. What are the characteristics of sequences with high duplicate counts (>100) that don't get memorized? Understanding these cases might provide insight into factors that inhibit memorization despite repeated exposure.\n5. How sensitive are the semantic similarity measurements to the length mismatch between the 64 (or 32)-token samples and the 2049-token sequences they're compared against? \n6. (Less important) Could the sudden increase in reconstruction at 86% of training be related to the \"induction bump\" phenomenon described in the mechanistic interpretability literature? Again, see \"In-context Learning and Induction Heads\", by Olsson et al. for the introduction of the concept.\n\n\nMinor points:\n\nThe Figure 6 would benefit from being vertically compressed without deforming the text.\n\nLikely typo lines 140-142: \"we generate document embeddings for each full sequence using SBERT and count the number of sequences with cosine similarity ≤ 0.8. These sequences are semantically similar but may not be exact token-level duplicates.\" -> I guess it should be \"cosine similarity ≥ 0.8\" instead." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- **The paper provides a methodologically sound approach to developing and validating taxonomies in ML research (and beyond).** By grounding qualitative distinctions in statistical analysis (e.g., following the example of Simpson's paradox), it offers a template for studying complex phenomena in ML beyond memorization that could be of interest for the ICLR community. \n- **The taxonomy enables discoveries about memorization dynamics.** For instance, the finding that duplicate count beyond 5 barely affects recitation likelihood challenges simple assumptions about exposure and memorization. The categorical distinctions also help align research directions with specific applications (e.g., privacy vs. copyright concerns). \n- **Analysis of the semantic properties of the sequence.** This type of statistics provides valuable insights into how models can perfectly reproduce sequences through pattern completion rather than pure memorization for the reconstruction category. This distinction, while simple in hindsight, is important for understanding the relationship between memorization and generalization, and the limit of the k-extractability definition of memorization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a taxonomy of memorization in language models, breaking it down into three categories: recitation of highly duplicated sequences, reconstruction of inherently predictable patterns, and recollection of rare sequences. The authors validate their taxonomy through statistical analysis and predictive modeling, showing that different factors influence memorization differently across categories. \n\nThey analyze how memorization patterns evolve with model scale and training time, finding that recollection grows disproportionately with model size. The paper illustrates the practical value of this taxonomy by showing how different categories matter for different applications (privacy, copyright, scientific understanding) and by achieving better predictive performance compared to both a baseline without taxonomy and an optimally partitioned model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **The predictive model's performance is relatively poor** (precision ~0.5, PR AUC ~0.5), despite including continuation perplexity as an input feature. This raises questions about the practical utility of the identified factors in predicting memorization. The heavy reliance on continuation perplexity for every category (see Figure 5.), which is closely related to the definition of k-extractability, makes it difficult to assess the independent predictive power of other factors. \n- **No clear progress in understanding the memorisation of rare sequence.** While the paper identifies recollection of rare sequences as an important phenomenon, particularly as models scale, it provides limited insight into the underlying mechanisms. This gap is particularly notable given the paper's emphasis on understanding different types of memorization.\n- **The presentation lacks clarity at times.** \n\t- When introducing the taxonomy, early concrete examples of each category would significantly improve understanding. \n\t- The paper should also better highlight the distinction between intuitive notions of memorization and the technical definition of k-extractability used in the study. This could help the reader understand why the reconstruction phenomenon (where sequence outside of the training set could be predicted perfectly) fall in the scope of the study of memorization. \n\t- The study could benefit of including reference to a broader set of references such as the study of mechanistic interpretability and training providing more insights on how and when models become able to predict simple sequences. See for instance \"In-context Learning and Induction Heads\", by Olsson et al.\n- **Methodological limitation in the computation of the corpus statistics.** \n\t- The corpus statistics are not broken down into prompt/continuation/full sequence. This could enable the isolation of sequences with a frequent prompt but infrequent continuation, or the opposite for instance. The paper doesn't clearly state which one of the three is used for the corpus statistics.\n\t- If I understood correctly, the semantic similarity measurements are made between sequences of length 64 (or 32) tokens (the memorized/non-memorized samples), and the 2049-token sequences from the Pile. This length mismatch could introduce heavy distortion as even if the small sequence is included in the large sequence, it is not clear that the cosine similarity of their embedding would be similar." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We break memorization down into a taxonomy: recitation of highly duplicated sequences, reconstruction of inherently predictable sequences, and recollection of sequences that are neither." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024recite,\ntitle={Recite, Reconstruct, Recollect: Memorization in {LM}s as a Multifaceted Phenomenon},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3E8YNv1HjU},\nnote={under review}\n}" }, "abstract": { "value": "Memorization in language models is typically treated as a homogenous phenomenon, neglecting the specifics of the memorized data. We instead model memorization as the effect of a set of complex factors that describe each sample and relate it to the model and corpus. To build intuition around these factors, we break memorization down into a taxonomy: recitation of highly duplicated sequences, reconstruction of inherently predictable sequences, and recollection of sequences that are neither. We demonstrate the usefulness of our taxonomy by using it to construct a predictive model for memorization. By analyzing dependencies and inspecting the weights of the predictive model, we find that different factors have different influences on the likelihood of memorization depending on the taxonomic category." }, "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": [ "memorization", "ontologies", "language modelling" ] }, "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/e6f4389afbf30374da98ea2d2c7a61b0533d144e.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": "Recite, Reconstruct, Recollect: Memorization in LMs as a Multifaceted Phenomenon" }, "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": "**Regarding Weakness 1**:\n\nQ1.1. Could the authors explicitly compare their task-level sparse prompting approach to that in [1], highlighting key methodological differences, and discuss any potential advantages of their approach over the method in [1]? \n\nQ1.2. Could the authors explain their motivation for sampling the task dictionary from N(0,1) rather than using dictionary learning?\n\nQ1.3. Could the authors clarify how their use of dormant neuron resetting differs from or improves upon the approach in [2], and explain why they consider it a key contribution despite its prior use in continual RL?\n\n**Regarding Weakness 2**:\n\nQ2.1. Could the authors provide clear explanation of the advantages of their \"sensitivity dormant scores\" over previous dormant score definitions, and provide an ablation study comparing their \"sensitivity dormant scores\" to other dormant score metrics?\n\nQ2.2. Could the authors provide theoretical or empirical justification for their key model choices?\n\n**Regarding Weakness 3.1**:\n\nQ3.1. Could the authors explain the discrepancy between their reproduced CoTASP results and those reported in the original paper?\n\nQ3.2. Could the authors provide generation performance results for SSDE for a fair comparison?\n\nQ3.3. Could the authors discuss how the performance gap affects their conclusions about SSDE's effectiveness compared to CoTASP?\n\n**Regarding Weakness 3.2**:\n\nQ3.4. Could the author explain why their method performs inferior to [1] when dormant neuron resetting is not used, and provide additional analysis or experiments to clarify whether their co-allocation strategy offers advantages over the method in [1]?\n\nQ3.5. Could the authors discuss potential reasons for the performance difference when dormant neuron resetting is not used and its implications for their method's effectiveness?\n\n**Regarding Weakness 4**:\n\nQ4.1. Could the authors provide a sensitivity analysis or ablation study for key hyperparameters to demonstrate the method's robustness?\n\nQ4.2. Could the authors provide a discussion of strategies for hyperparameter selection in practical applications?\n\nQ4.3. Could the authors provide a comparison of the number and sensitivity of hyperparameters in SSDE to those in baseline methods like CoTASP?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors proposed to separate sub-network sparse prompting into global and task-specific levels, which seems effective in their ablation study. \n\n2. The introduction of parameter-level masking and dormant neuron resetting techniques is beneficial for mitigating plasticity loss and keeping the learning capability of networks. \n\n3. Combining the proposed techniques, the overall framework achieves higher computation efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a new structure-based method for plasticity and stability tradeoff in continual RL. Specifically, the authors proposed (1) a sub-network co-allocation method which enhances meta sparse prompting by task-specific sparse prompting, (2) a finegrained masking method which only freeze exact parameters trained in previous tasks, and (3) a periodic resetting strategy for dormant neurons. However, this work is largely built upon previous works, which limits its technical contributions. And the reported experimental results are somewhat misleading. Thus, I lean towards reject at current stage." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper largely follows the technical parts of [1][2], which limits their own contribution. In detail,\n\n1.1. The authors claim a crucial contribution of introducing task-level sparse prompting, which has also been proposed in [1], where the task dictionary is obtained through a dictionary learning algorithm using previous tasks' optimized prompts and their embedding. Here the authors sampled the task dictionary from N(0, 1) which shares the same distribution of global dictionary. I don't see the motivation of this choice. Could the authors elaborate more on this part and also detail the difference between [1] and their method?\n\n1.2. In 4.2, since periodic dormant neuron resetting is well established in [2] for continual RL, it is inappropriate to claim this as a key contribution of their framework. In addition, the use of \"exploration\" is arguably misleading, which has various meanings in different contexts. I would suggest the authors use the specific term \"structural exploration\" throughout the paper for better clarification. \n\n2. There are many model choices made without sufficient justification. In addtion to 1.1, the proposed \"sensitivity dormant scores\" also lacks clear motivation. What is the advantage of this metric than previously defined dormant scores? And ablation study is necessary for justifying such choices. \n\n3. The experiments results seem misleading.\n\n3.1. Throughout the experiments, their reproduced results of CoTASP has a huge gap with the reported ones in CoTASP's paper (e.g. for P: 0.73 v.s. 0.92 in CW10 and 0.74 v.s. 0.88 in CW20), which has no essential difference with SSDE (CW10) or even better (CW20). And the generation performance of SSDE is not reported for a fair comparison with previous methods. \n\n3.2. In addition, the results are quite misleading given the reported P in CoTASP. In Table 4, the average success of SSDE w/o Dormant is 0.85 while that of CoTASP which also don't use Dormant neuron resetting achieves 0.92. Does this indicate that the proposed co-allocation strategy is inferior to the allocation method in [1] which also use sparse prompting? \n\n4. The overall frameworks contain a lot of hyperparameters, such as the sparsity controlling parameters in Equation (3-4), the trade-off paramter in Equation (6), and the dormant threshold in Definition 4.1, which make the framework less practical. In addition, no ablation results are provided to show the robustness of the system to these hyperparameters. \n\n[1] Yang, Yijun, et al. \"Continual task allocation in meta-policy network via sparse prompting.\" International Conference on Machine Learning. PMLR, 2023.\n\n[2] Sokar, Ghada, et al. \"The dormant neuron phenomenon in deep reinforcement learning.\" International Conference on Machine Learning. PMLR, 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": 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. If the differences between tasks are substantial, could using fixed forward-transfer parameters introduce issues that reduce flexibility?\n2. Can the method proposed by the author continue adapting to additional tasks, and can its task completion performance still outperform other methods?\n3. For additional issues, please refer to the weaknesses section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper addresses a critical issue in continual reinforcement learning: balancing plasticity and stability to mitigate catastrophic forgetting. The proposed method offers greater computational efficiency than existing approaches. Experimental results are promising, demonstrating that the proposed method achieves a higher success rate and outperforms other baseline methods on the CW10-v1 Continual World benchmark." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a structure-based method, SSDE, to enhance the plasticity-stability trade-off in continual reinforcement learning. It introduces a fine-grained allocation strategy that decomposes network parameters into fixed forward-transfer and task-specific trainable components. Furthermore, to improve the model’s exploration capability, this paper presents an exploration technique based on sensitivity-guided dormant neurons. Experiments conducted on the Continual World benchmark demonstrate that the proposed method achieves a superior success rate and outperforms current state-of-the-art methods in the CW10 task." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper can be improved in terms of writing/presentation.\n2. From Table 3, it is evident that the author’s proposed method generally performs weaker than the ClonEx-SAC method. Therefore, will the ClonEx-SAC method continue to outperform the author’s method as the number of sequential tasks increases? Additionally, why doesn’t the forward transfer ability, or generalization ability, of the author’s method improve as the number of tasks increases?\n3. The proposed sensitivity-guided dormant neurons offer limited novelty.\n4. The author does not include comparisons with the latest regularization-based methods in continuous reinforcement learning, and the multi-task methods referenced are also outdated." }, "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)Can we have more continual RL tasks? In the existing version, CW tasks may be not very convincing, especially CW20. Maybe you can refer to [a] for more RL tasks to evaluate continual RL methods.\n[a] Online Continual Learning for Interactive Instruction Following Agents, 2024\n(2)\\beta in Eq.6 is very important since it balance the stability and plasticity in CL. Could you show its sensitivity or how do you decide the optimal value." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "(1)SSDE not only achieves SOTA stability standards but also achieves competitive plasticity even when compared to strong behavior cloning baselines that benefit from data replay.\n(2)Experimental results demonstrate that SSDE outperforms current state-of\u0002the-art methods and achieves a superior success rate of 95% on the CW10-v1." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces SSDE, a novel structure-based continual RL method. SSDE formulates an efficient co-allocation algorithm that enhances sub-network allocation by increasing the capacity for trainable parameters while leveraging frozen parameters for effective forward transfer from previous policies. SSDE introduces a trade-off parameter to balance these two groups of parameters in its fine-grained inference process." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "As shown in Table 3, SSDE takes no obvious advantages in F and FT metrics. These two metrics usually represent backward and forward transfer. So why Average Performance (P) is so good while F and FT not? I am a little confused. Also, ClonEx-SAC seems to perform comparably with SSDE on CW 20, although it replays 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": 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": "* Could the authors provide an intuitive explanation for the use of sparse coding in co-allocation? I agree with the general idea of enhancing plasticity with forward-transfer parameters [1], but I do not fully understand why they can be selected using sparse coding.\n\n---\n\n[1] Lin, et al. TRGP: Trust region gradient projection for continual learning. ICLR 2022." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper is generally well written and very informative.\n* The technical contributions (fine-grained subnetwork allocation and dormant neuron re-activation) seem solid, both effectively addressing the stability-plasticity dilemma in continual reinforcement learning.\n* The experimental results on the Continual World benchmark show great improvements over the existing baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a structure-based continual reinforcement learning algorithm. To address the stability-plasticity dilemma in the previous subnetwork allocation approaches, the authors propose a fine-grained allocation strategy with two key designs: (1) The parameter space is decomposed into forward-transfer and task-specific components, which are co-allocated by sparse coding to enhance plasticity. (2) The dormant neurons are periodically reactivated to improve exploration and rapid adaptation in new tasks. The proposed method is validated on the Continual World benchmark and shows significant simprovement in performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The second contribution (dormant neuron re-activation) could be validated more carefully. There are already several works dealing with dormant neurons (e.g. [1, 2]), so a comparison with the existing literature in terms of method differences and experimental results would help to justify the contribution of this paper. Also, since the authors propose a new criterion for dormant neurons, it should be investigated how it overlaps with the original definition in [1], is it possible that it induces false positives and thus destabilizes the learning process? \n* The proposed method utilizes task embeddings from a pre-trained BERT, which limits its applicability in broader scenarios involving implicit knowledge that cannot be verbalized. For example, while BERT embeddings work fine in the Continual World benchmark involving manipulation tasks, they may be difficult to transfer to the Brax scenarios used in [3, 4] involving locomotion tasks. It would be interesting to see further experiments or discussions on this issue.\n* Regarding efficiency, the authors only discussed their allocation time, but there is no mention of total training time and model size. These are more representative metrics for evaluating the efficiency of continual reinforcement learning, and should be reported in detail in the paper. Also, it would be better if the authors could include an intuitive figure of the performance-size or performance-training time tradeoff, perhaps like Figure 1 in [3] and Figure 4 in [4].\n* There is a lack of sensitivity analysis for several hyperparameters, including $\\beta$ in Equation (6), $\\Delta$ in Equation (8) and $\\tau$ for thresholding. The paper determined these hyperparameters by grid search, but I am curious if their choice is critical to the final performance and difficult to choose in practice. The authors could present a sensitivity analysis of these hyperparameters to address my concern, and additional studies of their generalizability across different types of tasks would be appreciated.\n\n---\n\n[1] Sokar, et al. The dormant neuron phenomenon in deep reinforcement learning. ICML 2023.\n\n[2] Dohare, et al. Loss of plasticity in deep continual learning. Nature 2024.\n\n[3] Gaya, et al. Building a subspace of policies for scalable continual learning. ICLR 2023.\n\n[4] Sun, et al. Rewiring neurons in non-stationary environments. NeurIPS 2023." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "SSDE improves Continual Reinforcement Learning by balancing plasticity and stability to prevent forgetting." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024plasticity,\ntitle={Plasticity from Structured Sparsity: Mastering Continual Reinforcement Learning through Fine-grained Network Allocation and Dormant Neuron Exploration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3ENBquM4b4},\nnote={under review}\n}" }, "abstract": { "value": "Continual reinforcement learning faces a central challenge in striking a balance between plasticity and stability to mitigate catastrophic forgetting. In this paper, we introduce SSDE, a novel structure-based method that aims to improve plasticity through a fine-grained allocation strategy with Structured Sparsity and Dormant-guided Exploration. Specifically, SSDE decomposes the parameter space for each task into forward-transfer (frozen) parameters and task-specific (trainable) parameters. Crucially, these parameters are allocated by an efficient co-allocation scheme under sparse coding, ensuring sufficient trainable capacity for new tasks while promoting efficient forward transfer through frozen parameters. Furthermore, structure-based methods often suffer from rigidity due to the accumulation of non-trainable parameters, hindering exploration. To overcome this, we propose a novel exploration technique based on sensitivity-guided dormant neurons, which systematically identifies and resets insensitive parameters. Our comprehensive experiments demonstrate that SSDE outperforms current state-of-the-art methods and achieves a superior success rate of $95\\%$% on CW10 Continual World benchmark." }, "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": [ "Continual reinforcement learning", "Policy transfer" ] }, "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/6ec5fc4c7085bf4df09cdb326da3b80aff5f04a7.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": "Plasticity from Structured Sparsity: Mastering Continual Reinforcement Learning through Fine-grained Network Allocation and Dormant Neuron Exploration" }, "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": 3 }, "primary_area": null, "questions": { "value": "1. Could the authors elaborate on how they determined the weights for performance and complexity in the reward function? More detail on this could clarify the balance between the two objectives.\n\n2。 How does TCTO perform on high-dimensional datasets with over 10,000 features? Is the pruning strategy sufficient to maintain stability without compromising feature diversity?\n\n3. Were there any specific scenarios where TCTO’s backtracking mechanism was particularly beneficial in terms of model performance or feature diversity?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. While mostly clear, certain sections (e.g., cascading agent decision process) could benefit from additional details.\n\n2. The framework is well-supported by experimental evidence showing its adaptability across different datasets and improvement in downstream model performance.\n\n3. TCTO introduces a novel approach to automated feature engineering by employing a transformation-centric methodology with a graph-based roadmap, overcoming limitations of existing feature transformation methods.\n\n4. The approach’s ability to backtrack and optimize feature transformations dynamically makes it highly applicable in real-world ML tasks where feature diversity and stability are crucial." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an automated feature transformation framework designed to enhance downstream machine learning model performance. The TCTO framework leverages a reinforcement learning-based graph structure to maintain a roadmap of feature transformations, enabling efficient exploration and backtracking of transformation pathways. TCTO uses a multi-agent reinforcement learning approach, clustering and encoding transformation states to strategically apply feature transformations. Experiments on multiple datasets demonstrate TCTO's performance over existing methods by improving robustness and flexibility in feature generation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While effective on a range of datasets, it is unclear how well TCTO scales with extremely high-dimensional data or very large datasets, as the pruning strategy may require fine-tuning in these cases.\n\n2. The cascading decision-making process is intricate, and further simplification or additional visuals might aid understanding.\n\n3. The reward structure combines performance and complexity, but further discussion on how these metrics are weighted could improve transparency and replicability of the model’s efficacy." }, "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": "1.How does the computational complexity of TCTO scale with larger datasets, and are there any strategies to mitigate potential performance bottlenecks?\n2.Are there scenarios or specific types of datasets where TCTO’s performance may be limited, and if so, what adjustments might be necessary to enhance its adaptability?" }, "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 several notable strengths. Firstly, the authors present a well-motivated framework that addresses clear gaps in current automated feature transformation methods, such as the need for effective historical data utilization and robust backtracking. The proposed TCTO framework is innovative in its use of a graph-based roadmap and cascading multi-agent reinforcement learning, which enhance the flexibility and adaptability of the transformation process. Additionally, the authors provide a comprehensive experimental evaluation across diverse datasets, which convincingly demonstrates TCTO’s superior performance compared to traditional methods. This solid empirical foundation supports the framework's potential for broad applicability in feature engineering for machine learning tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors present TCTO, a graph-based reinforcement learning framework designed for automated feature transformation. The approach addresses limitations in current methods, such as the lack of historical insight utilization and insufficient flexibility in transformation exploration. By constructing a transformation roadmap with nodes representing feature states, TCTO leverages a cascading multi-agent system to dynamically select transformations, reuse effective paths, and prune inefficient ones. The experimental results demonstrate that TCTO outperforms existing methods in generating high-quality features, suggesting its potential to enhance feature engineering in machine learning tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While this paper offers a promising framework, it has some weaknesses. Firstly, the explanation of the cascading multi-agent system and its decision-making processes could benefit from more clarity and detail, as the current description may be challenging for readers to fully grasp without additional context. Additionally, the computational complexity of TCTO is not thoroughly analyzed, especially regarding scalability to larger datasets, which may impact its practical applicability. Finally, while the experimental results are extensive, the paper could further strengthen its claims by providing more insight into specific scenarios or datasets where TCTO may struggle, thereby clarifying the framework’s limitations and potential areas for improvement." }, "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": 1 }, "primary_area": null, "questions": { "value": "* How were the small uncertainties in Table 1 achieved? How often were the experiments repeated?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* It can be seen (among other things from the large number of specific illustrations) that a lot of effort was put into preparing the paper" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper deals with the automated generation of features. The generation process consists of several steps, which are represented as a graph. The graphs are to be optimized using multi-agent reinforcement learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* I find the text very badly written. Examples follow. The novelty and benefits of the method are hard for me to understand.\n* It seems to me that there is too much material for a conference paper, the number of pages is simply not enough to present it in a convincing way.\n\nDetails, examples and further comments:\n* I don't think “roadmap” is a suitable term, “schedule” or \"sequence\" would probably be better.\n* The title sounds strange. Wouldn't \"Optimization of transformation sequences for automated feature generation“ be better?\n* The abstract uses terms that are incomprehensible:\\\nmathematical feature-feature crossing\\\nthe roadmap of feature transformation\n* „Feature transformation task aims to generate high-value features and improve the performance of downstream machine learning tasks using the mathematical feature-feature crossing” needs to be reformulated.\n* \"Classic machine learning is highly dependent on the structure of the model, the activation function\" cannot be said in this way, it seems to refer exclusively to neural networks and not to classical machine learning in general.\n* A reference should be given for \"a cascading multi-agent reinforcement learning (MARL) algorithm\", because it is not generally known what “cascading multi-agent reinforcement learning” is.\n* “we present an innovative framework” -> “we present a novel framework”\n* In the loss function, Equation 8, the square is probably missing. \n* \"In this study, we introduce TCTO, an automated feature transformation framework. Our method emphasizes a transformation-centric approach, in which a transformation roadmap is utilized to systematically track and manage feature modifications.\" should be reworded. What is the information content? What should be expressed?\n* I think that the Abstract and Conclusion need to be completely rewritten." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024build,\ntitle={Build Roadmap for Automated Feature Transformation: A Graph-based Reinforcement Learning Approach},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3EeyQNgKTP},\nnote={under review}\n}" }, "abstract": { "value": "Feature transformation task aims to generate high-value features and improve the performance of downstream machine learning tasks using the mathematical feature-feature crossing. \nCurrent frameworks rely on iterative sequence generation with exploration optimization through performance feedback from downstream tasks.\nHowever, these approaches fail to effectively utilize historical decision-making experiences and overlook potential relationships among generated features, thus limiting the flexibility of the whole process.\nMoreover, the decision-making process lacks dynamic backtracking capabilities for each feature, leading to insufficient adaptability when encountering inefficient pathways, adversely affecting overall robustness and exploration stability. \nTo overcome these challenges, we present an innovative framework that employs a feature-state transformation graph to maintain the roadmap of feature transformation, with each node symbolizing a transformation state. \nDuring exploration, three cascading agents sequentially select nodes and mathematical operations to generate new transformation states.\nThis strategy leverages the graph structure's inherent properties, allowing for the preservation and reuse of sight-seen and valuable transformations. \nIt also enables back-tracking capabilities through graph pruning techniques, which can rectify inefficient transformation paths.\nTo validate the efficacy and flexibility of our approach, we conducted comprehensive experiments and detailed case studies, demonstrating superior performance in diverse 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": [ "Automated Feature Transformation", "Tabular Data", "Multi-Agent 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/741a0ae34f509e38c93795f4fd3954341ede5866.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "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/22d2afa2f51dd27c3984077c662b697447aafce0.zip" }, "title": { "value": "Build Roadmap for Automated Feature Transformation: A Graph-based Reinforcement Learning Approach" }, "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": "- How sensitive is the model to the quality and completeness of the pathway knowledge used? Have you tested with different pathway databases or subsets of pathways?\n- How does the computational complexity scale with the number of biological processes? Is there a practical limit to how many processes can be incorporated?\n- Have you explored whether the causal relationships discovered by the model align with known biological pathway interactions beyond the examples provided?\n- Can you confirm that all genes involved in the double perturbations were also present in your single-perturbation training data?\n- Has $N$ and $\\tau$ been mixed up in the Hits@N metric?\n- How does table 2 show that “both models tend to assign most interventions to a small number of latent factors”?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Clear technical contribution that bridges causal representation learning with biological interpretability while maintaining theoretical guarantees\n- The paper contributes to causal representation learning for Perturb-seq data by introducing biological interpretability through pathway information, while maintaining the theoretical guarantees of discrepency-VAE.\n- Well written and clear presentation of the method and results.\n- Thorough ablation studies\n- Demonstrates interpretability of latent factors with concrete biological examples." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the challenge of learning interpretable causal representations for Perturb-seq data (gene expression in cells). The primary contribution is the novel introduction of masking to incorporate biological process (BP) knowledge into an existing method for causal representation learning (discrepency-VAE), which is named SENA-discrepancy-VAE. The masking ensures that latent factors can be interpreted as linear combinations of the activity of BPs. Since this modification is compatible with the discrepancy-VAE, the original model's theoretical guarantees for causal representation learning remain.\n\nThe method and ablated variants are evaluated on a Perturb-seq dataset collected from one particular cell line and is set up to minimize the overlap between the BPs. The results demonstrate that SENA performs similarly to discrepency-VAE in terms of reconstruction yet results in sparser and more interpretable results. Furthermore, by studying the contrast between inferred activity levels on perturbed and control samples the authors show that the latent factors can be associated with BPs and are therefore interpretable." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Experimental validation is limited to one dataset and no baselines other than their own ablations and discrepancy-VAE. The paper would benefit from comparisons to at least one of the other listed related works.\n- No comparison with simpler approaches like post-hoc interpretation of standard discrepancy-VAE latent factors.\n- While the link between latent factors and BPs is investigated, the quality of the discovered causal graph is not.\n- Given that the latent factors group a large number of BPs into a small number of latent factors there should be a deeper investigation of the biological plausibility and practicality of this result beyond the contrasting activations.\n- Readability of several figure texts should be improved." }, "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": "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": "What is the intuition behind the $\\lambda$ hyperparameter to tune small influences of a gene on a biological process? Should this be a constant value throughout the mask matrix or would it be better to learn this influence via some type of attention weights?" }, "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 written with strong motivations behind using CRL techniques for biological applications.\n- The metrics proposed (differential activation, Hits@N) seem to be robust indicators of perturbation effects on BPs and downstream effects. I believe these evaluation metrics are one of the key interesting contributions of this work.\n- The empirical evaluation is exhaustive and illustrates some interesting observations, especially the representational capacity of the VAE-based SENA method compared to the traditional discrepancyVAE.\n- The interpretability analysis of the reparameterization layer is interesting and reveals which genes were affected the most upon perturbations. I do believe that exploring real-world applications of CRL is a very important direction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes to extend the discrepancyVAE interventional causal representation learning framework to biological processes applications. Specifically, the authors propose to embed prior knowledge about biological processes (BPs) through a framework called SENA-discrepancyVAE, which recovers latent factors that are a linear combination of a set of biological processes (pathways). The main idea presented in this work is to design a more flexible encoder class (SENA-\\delta) specific to mapping biological pathways to latent causal factors for interpretability. Empirical results show that the framework is shown to improve performance in predicting the effect of unseen perturbations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Although the application in gene regulatory networks is quite interesting, this work seems to be more of an evaluation study of the discrepancy-VAE framework proposed by Zhang et al. I do not see much of an added contribution beyond the original paper besides highlighting the application.\n- The difference in performance between the SENA variant and the original discrepancyVAE seems to be quite marginal in terms of representation in the double-perturbation scenario. For instance, in Table 2, the KL-divergence for double-perturbation prediction is only marginally better than the original MLP-based discrepancyVAE." }, "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": "Can you provide examples of BPs in the appendix? \n\nRegarding DAR evaluation: What happens if unnaffected pathways have very low action? Also, more general, how would you deal with impalanced pathways, which might lead in measuring large noise levels? \n\nIn table 2, for SENA λ=0.1, latent dim 105, the variance compared to original MLP and λ=0 is significantly lower (0.000081 vs 0.001087). Just double checking if this is corrent.\n\nL237: During filtering you end up with a (biased) set of BPs. How much do you think this can influence interpretability? Is there a risk of removing useful BPs?\n\nSuggestions\nL100: the word faithfully here gets confused with causal faithfulness. Please consider an alternative adverb if possible.\nL105: target instead of targets (remove final s)" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Clarity: The paper is well written and easy to follow.\n\n* Novel Technical Contribution: The paper successfully extends causal representation learning to incorporate domain knowledge while preserving theoretical guarantees. The SENA-δ encoder architecture is a clever solution to balance interpretability and performance.\n\n* Practical Impact: The work addresses a significant gap in current causal representation learning methods for biological data, where interpretability is crucial for scientific insights. The ability to map latent factors to biological processes makes the model more useful for domain experts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents SENA-discrepancy-VAE, an extension of the discrepancy-VAE framework that incorporates biological pathway knowledge to produce interpretable causal latent factors. The authors modify the encoder architecture to map gene expression through biological processes (BPs) while maintaining the theoretical guarantees of the original model. The approach achieves comparable predictive performance to the original discrepancy-VAE while providing biologically meaningful latent representations and interpretability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Limited Biological Validation: While the authors show statistical associations between perturbations and biological processes, there could be more validation using external biological knowledge or experimental validation of the discovered causal relationships.\n\n* Hyperparameter Sensitivity: The model introduces an additional hyperparameter λ that significantly impacts performance. While ablation studies are provided, more guidance on selecting this parameter would be valuable (this is important given that there's some large impact on the performance of the method)\n\n* Restricted Evaluation: The empirical evaluation is limited to a single dataset (Norman et al., 2019). Additional validation on different types of biological data would strengthen the claims of 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": 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 the authors clarify whether multi-label or probabilistic pathway activity labels were considered as an alternative to binary labels? Binary labels may oversimplify gene activity levels, especially when certain pathways exhibit gradations rather than discrete on/off states.\n\n2) How would the model perform if pathway relevance were dynamically adjusted based on specific task contexts or experimental conditions? Pathway importance often varies, and adapting pathway relevance could improve model flexibility. A task-specific analysis to explore whether dynamically adjusting pathway selection enhances generalizability across datasets and biological contexts could provide insights into the model’s robustness.\n\n3) In the causal graph (Figure 3), how sensitive are the edge weights to the choice of λ and latent dimension? Please provide a sensitivity analysis.\n\n4) The mask matrix M (equation 2) assumes binary gene-pathway relationships. Have you considered using weighted relationships based on pathway databases' confidence scores?\n\n5) How does D_KL evolve during training for different λ values? Training curves would help understand if this is an optimization or regularization effect. Does this phenomenon persist if you randomize the pathway annotations while maintaining the same sparsity structure? This would determine if the benefit comes from biological knowledge or just sparsity.\n\n6) Have the authors empirically validated that the expectation in Equation (10) aligns with the observed experimental outcomes? Demonstrating this match would reinforce the theoretical assumptions and provide additional confidence in the model's causal interpretability." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) The proposed model’s integration of biological pathway data as a prior in the causal representation learning (CRL) framework is interesting and practical, offering a biologically grounded solution to gene expression analysis. Through the direct alignment of the latent factors with biological pathways, SENA-discrepancy-VAE addresses the common limitation in CRL models of producing uninterpretable latent factors. \n\n2) The paper presents thorough experiments across multiple perturbation types, including single and double-gene knockouts, demonstrating the model’s robustness. Its generalization to unseen perturbations is compelling, and the ablation studies, which explore interpretability-reconstruction trade-offs, further validate the model’s design.\n\n3) The authors have done a good job communicating the importance of embedding biological processes into latent spaces, with visuals that illustrate BP-specific activity levels influencing latent factors. The use of causal graphs and the differential activation (DA) metric enhances transparency, allowing readers to trace latent factors back to biological functions. \n\n4) The model’s ability to provide interpretable predictions on cellular responses can aid in experimental design and offer insights into the potential effects of genetic or drug interventions. This approach addresses a pressing need in biomedicine for interpretable causal representation learning (CRL) models that can shed light on the intricate causal relationships underlying gene function and cellular processes." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces SENA-discrepancy-VAE, a novel model in causal representation learning (CRL) designed to make biological data analysis—especially from Perturb-seq experiments—more interpretable. A key innovation of this work is how it integrates biological processes (BPs) as prior knowledge, directly linking the model’s latent factors to known biological pathways. This approach fills an important gap in existing CRL methods, which often struggle with interpretability since they don't directly associate learned representations with actual biological mechanisms, making them less useful for real research applications.\n\nSENA-discrepancy-VAE builds on the standard discrepancy-VAE by introducing a pathway-based masking strategy within a new encoder, SENA-δ. This encoder uses a two-layer masked MLP where the first layer maps gene expression values to BP activity levels, with a tunable parameter that adjusts the influence of genes outside predefined pathways, giving the model flexibility in gene-pathway associations. The second layer models latent factors as combinations of these BP activities, which is a more realistic approach since biological interventions often impact multiple pathways. This setup stays true to the CRL assumption that each intervention targets a single latent factor but does so in a way that aligns with biological realities.\n\nThe authors evaluate SENA-discrepancy-VAE on a Perturb-seq dataset of leukemia lymphoblast cells. They show that the model performs as well as the original discrepancy-VAE on unseen perturbation combinations while providing greater interpretability by identifying specific BPs associated with each latent factor. This interpretability is validated through pathway-specific analysis, demonstrating the model’s ability to reveal biologically meaningful patterns in response to genetic interventions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) While the model effectively identifies single-point perturbations, it does not accommodate multi-step perturbations or capture the progression of cellular responses over time. In biological experiments, the cellular response often evolves in phases, with gene activity showing distinct transitions that are essential for understanding the effect of interventions.\n\n2) The model’s validation on a single dataset (K562 cell line data) restricts insights into its generalizability across different cell types or conditions. Testing on additional datasets, such as those from other cell lines or cellular environments, would offer a stronger assessment of robustness and applicability across a wider range of biological data.\n\n3) The model assumes static pathway relevance across all tasks, which may limit its adaptability in varied biological contexts where pathway importance changes with cell type or condition.\n\n4) The paper assumes that each intervention corresponds to a single latent factor, limiting the model's ability to capture complex interactions where multiple latent factors might be influenced by a single intervention. This simplification restricts the model’s interpretability in representing overlapping or interacting biological processes, which are common in gene expression dynamics.\n\n5) The model lacks a detailed exploration of how varying the number of latent dimensions impacts the interpretability and causal mapping of latent factors. Larger or smaller dimensions can influence the granularity of the factors and thus affect the biological insights the model can provide." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024interpretable,\ntitle={Interpretable Causal Representation Learning for Biological Data in the Pathway Space},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Fgylj4uqL},\nnote={under review}\n}" }, "abstract": { "value": "Predicting the impact of genomic and drug perturbations in cellular function is crucial for understanding gene functions and drug effects, ultimately leading to improved therapies. To this end, Causal Representation Learning (CRL) constitutes one of the most promising approaches, as it aims to identify the latent factors that causally govern biological systems, thus facilitating the prediction of the effect of unseen perturbations. Yet, current CRL methods fail in reconciling their principled latent representations with known biological processes, leading to models that are not interpretable. To address this major issue, in this work we present SENA-discrepancy-VAE, a model based on the recently proposed CRL method discrepancy-VAE, that produces representations where each latent factor can be interpreted as the (linear) combination of the activity of a (learned) set of biological processes. To this extent, we present an encoder, SENA-$\\delta$, that efficiently compute and map biological processes' activity levels to the latent causal factors. We show that SENA-discrepancy-VAE achieves predictive performances on unseen combinations of interventions that are comparable with its original, non-interpretable counterpart, while inferring causal latent factors that are biologically meaningful." }, "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": [ "Causal Representation Learning", "Intepretability", "VAE", "Genomic Perturbations", "Health" ] }, "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/d682c5136d996ce63e84ed25d12be6a57104abf8.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": "Interpretable Causal Representation Learning for Biological Data in the Pathway Space" }, "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": 3 }, "primary_area": null, "questions": { "value": "See Above" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well-structured with a clear methodology.\n\n2. It offers some insights into Aya's performance in multilingual contexts and addresses challenges faced by low-resource languages." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper evaluates the performance of the multilingual Aya model across tasks such as Aspect-Based Sentiment Analysis (ABSA), Hate Speech Detection (HS), Irony Detection (ID), and Question-Answering (QA) in Brazilian Portuguese. Through a few-shot learning approach, Aya demonstrates competitive results in QA, surpassing some Portuguese-specific models, though it underperforms in tasks involving nuanced or slang-heavy language like HS. The study highlights Aya's potential in low-resource contexts while indicating the need for further tuning for certain language-specific tasks to match or exceed specialized models​" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper only conducts an evaluation of the multilingual large language model on a specific language, Brazilian Portuguese, even though Aya supports 101 languages. The tasks are limited to Aspect-Based Sentiment Analysis (ABSA), Hate Speech Detection (HS), Irony Detection (ID), and Question-Answering (QA). Many other NLP tasks could be studied, such as reading comprehension, syntax parsing, named entity recognition, and event extraction. The evaluation scope and language focus are limited, reducing the paper's contribution.\n\n2. Describing the equations for precision and recall in such detail seems unnecessary and only increases the document length without adding value.\n\n3. This paper lacks inspiring conclusions from the experiments. It only presents main results and a confusion matrix, without providing in-depth analysis through fine-grained evaluation or insights into the working principles of the Aya model." }, "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 see the above." }, "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 centered around a very interesting and highly relevant topic.\n\n- Clearly scoped tasks and objectives." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an evaluation of Aya, a multilingual language model, on four tasks including Aspect-Based Sentiment Analysis, Hate Speech Detection, Irony Detection, and Question-Answering, highlighting its strengths and limitations, particularly when compared to transformer-based models for the Portuguese language." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are a few points that must be addressed:\n\nThroughout the text it feels like the authors use aggressive speech, hate speech, and offensive speech interchangeably. For example, in Figure 1 we see offensive vs non-offensive. The authors should pay attention to this aspect: clearly define the specific type of abusive phenomena they are focusing on, and use that terminology consistently throughout their work – please look into the nuances surrounding the overlapping abusive phenomena (the distinction between hate speech, abusive language, and offensive language). See for example the work of Poletto et al.:\n\n*Poletto F, Basile V, Sanguinetti M, Bosco C, Patti V. Resources and benchmark corpora for hate speech detection: a systematic review. Language Resources and Evaluation. 2021 Jun;55:477-523.*\n\nI would have liked for the authors to spend a little bit more space detailing the methodology. The authors should provide the criteria used for selecting the examples, as well as the exact prompts used for all the tasks (not just QA):\n\n- How did the authors ensure that the examples were representative and diverse? \n- What was the exact input for the Aya model? The authors provide the prompt for the QA, but not for the other tasks. Does that mean that for the other, only the examples were used? I am asking this because I am surprised by the fact that for hate speech *‘the generation was more efficient when using the labels as numbers, instead of the actual labels’* (cf. lines 309-311). For example, I just asked Aya if it is familiar with the hate speech definition provided by the OHCHR, and the answer was positive. How would the generation have changed if including this type of information? Did the authors provide any task-specific instructions to the model beyond the few-shot examples?\n\nA more in-depth error analysis would have been interesting to have. The authors could consider a subset of the misclassified data and construct aggregate statistics over modes of failure which would then inform us how prevalent each of the kinds of mistake are. This would be useful for future research, as it would become possible to prioritize on which kind of misclassification to work on next.\nIn regards to the ABSA example provided, I don’t agree that *‘hotel’* has a neutral sentiment – it seems to be conflict (i.e., both positive and negative) or, we could say that the entity hotel has a positive sentiment towards the attributes location and service, but negative towards the attribute that would incorporate size/room description.\n\nWas there any hyperparameter tuning performed for the transformer-based models? Interesting results for the Albertina models on the ID task. \n\nSuggestions:\n\n- abbreviations should be presented once, at their first occurrence, and always used afterwards (except for the abstract)\n- I believe the paragraph starting on line 089 is actually a continuation of the previous one and does not require splitting\n- line 124: an -> a\n- line 161: a -> an" }, "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": "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": "Other than Portuguese have the authors considered evaluating their model for any other low resource language? That would provide a more comprehensive idea of the performance variance of the Aya across languages." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper addresses an important problem that aims to mitigate technological inequities towards low resource language.\nThe methodology used in this paper is reasonable and easy to understand. In addition, the paper also makes a thorough inquiry of related research before carrying out their work.\nThe experiment section provides adequate amount of evaluation to properly assess the performance of the model for a number of language tasks in a low resource language like Portuguese." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study aims to assess the performance of Aya, a multilingual generative model trained on a wide range of low resource languages and a variety of downstream tasks like Aspect-Based Sentiment Analysis, Hate Speech Detection, Irony Detection, and Question-Answering. The objective is to evaluate Aya's effectiveness in these tasks but only on the pre-trained model without any finetuning. This would reveal its potential to improve the quality and accuracy of outputs in various natural language understanding tasks. Instead, this work employs a few-shot methodology to evaluate the model's effectiveness as this approach is particularly better suited in abscence of extensive labelled data in low resource languages. Results indicate that while Aya performs well in certain tasks like Question-Answering for languages like Portuguese, for other tasks like Hate Speech Detection the performances were significantly underwhelming. These results suggest that multilingual models like Aya can perform competitively in some contexts but may require further tuning to match the effectiveness of models specifically trained for Portuguese." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This work is low in novelty despite addressing an important problem. Training a generative model for a low resource language like Portuguese is certainly important work. While the authors employ few shot learning as a logical workaround for the issue of low training data, their evaluation reveals the relative limitation of this approach after a point. The authors could further look into some technical innovations in this regard to improve performance of the Aya model in portuguese for tasks like Hate Speech or Irony Detection." }, "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": "**Comments:**\n1. How do you ensure the examples of few-shot learning are representative and diverse? What is the measure for representative for example do you consider the classes or the related input text? \n2. L259-261, Do all the chosen questions begin with \"What\", “Where”, “Who” and “When” represents the dataset in a general way?\n3. L268-L269, \"we selectively remove instances from the dataset and include them alongside each test example during inference.\" -- Why do you remove instances from the dataset? The removed instances belong to which splits are not discussed.\n4. In Table 1, few-shot examples are 11 (SA), 10 (HS), 20 (ID), and 4(QA). Do you use the same examples for all inference data?\n5. \"In total, they mention nine different aspects, including four examples with negative polarity, four with positive polarity, and three that are neutral.\" -- Who mentioned? I believe there should be a citation.\n6. There are nine different aspects in the dataset, does every aspect represent only one sentiment? If not, how does the example represent the other classes that are not selected? \n7. Equations 1-6 are well known to the community, the information should be redundant.\n8. L421-422, why neutral is harder than positive and negative is not discussed properly.\n9. Comparing the results of two classes (Positive and Negative) with three classes (Positive, neutral, and negative) is not a good idea. Given that the model can easily differentiate two classes where prediction of multi-class is harder.\n10. Figure 2, the sum of the confusion matrix for columns is not 100. \n11. The authors stated that the model predicted exactly the same answer as the ground truth. However, it would be interesting to see some examples of those answers.\n12. SQUAD v1 is an old dataset and there is a possibility of adding this dataset to Aya's training data. It would be interesting to see the performance on some other QA datasets." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper uses state-of-the-art multilingual LLMs (Aya), that have been known for their capabilities for low-resource languages. Moreover, the study uses four different tasks (SA, HS, ID, and QA) ranging from classification to text generation/QA." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper assesses Aya’s performance in four tasks: Aspect-Based Sentiment Analysis, Hate Speech Detection, Irony Detection, and question answering in Portuguese language. The authors also compared Aya's performance with other language models. However, it is not understandable which one is from prior studies in Table 2. It is also unclear whether the Sabia-7B model is studied in this study or previous study." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main weakness of this paper is written poorly. The related work and the theoretical background sections are too long. The claim about the examples for few-shot learning could be partially correct but not properly correct. The details of prompting are missing in the paper. The performances are not well discussed in the paper.\nI believe the paper will be in good shape if the content is trimmed to a short paper rather than a long full paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024aya,\ntitle={Aya in Action: An Investigation of its Abilities in Aspect-Based Sentiment Analysis, Hate Speech Detection, Irony Detection, and Question-Answering},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3GMuudWmMV},\nnote={under review}\n}" }, "abstract": { "value": "While resource-rich languages such as English and Mandarin drive considerable advancements, low-resource languages face challenges due to the scarcity of substantial digital and annotated linguistic resources. Within this context, \nin 2024, Aya was introduced, a multilingual generative language model supporting 101 languages, over half of which are lower-resourced. This study aims to assess Aya's performance in tasks such as Aspect-Based Sentiment Analysis, Hate Speech Detection, Irony Detection, and Question-Answering, using a few-shot methodology in Brazilian Portuguese. The objective is to evaluate Aya's effectiveness in these tasks without fine-tuning the pre-trained model, thereby exploring its potential to improve the quality and accuracy of outputs in various natural language understanding tasks.\nResults indicate that while Aya performs well in certain tasks like Question-Answering, where it surpassed Portuguese-specific models with an Exact Match score of 58.79%, it struggles in others. For the Hate Speech Detection task, Aya's F1-score of 0.64 was significantly lower than the 0.94 achieved by the Sabiá-7B model. Additionally, the model's performance on the Aspect-Based Sentiment Analysis task improved considerably when neutral examples were excluded, but its handling of complex slang and context-dependent features in other tasks remained challenging. These results suggest that multilingual models like Aya can perform competitively in some contexts but may require further tuning to match the effectiveness of models specifically trained for Portuguese." }, "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": [ "Sentiment Analysis", "Hate Speech Detection", "Irony Detection", "Question-Answering", "Large Language Models", "Few-shot Learning", "Portuguese Language." ] }, "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/eb37281b432581d69ead4acd8e3f017817d486c5.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": "Aya in Action: An Investigation of its Abilities in Aspect-Based Sentiment Analysis, Hate Speech Detection, Irony Detection, and 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": 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": "- How do you ensure that the generated perturbations effectively introduce the desired bias without altering the correctness of the content? How well do the LLMs understand the instructions for generating biased content? Could there be unintended consequences or biases introduced by the LLMs themselves?\n- Would incorporating a human evaluation benchmark provide additional insights into the accuracy and fairness of LLM judgments?\n- Are there potential trade-offs between mitigating biases and maintaining the performance of LLM judges?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Originality: The authors expand upon existing work by identifying and categorizing 12 distinct types of biases.\n- Quality: The paper presents a thorough evaluation of the identified biases across multiple LLMs, using diverse datasets and specific metrics tailored for judging tasks. This rigorous experimental design ensures the reliability and validity of the findings.\n- Clarity: The examples in Table 1 provide concrete examples of how biases manifest in LLM judgments, making the abstract concepts more tangible and relatable.\n- Significance: The proposed CALM framework offers a valuable tool for stakeholders to assess and mitigate biases, leading to more fair and reliable LLM evaluation methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the potential biases inherent in using Large Language Models (LLMs) as judges in various evaluation tasks, such as scoring and pairwise comparison. The authors propose a novel framework called CALM, which systematically quantifies and analyzes each type of bias by using automated and principle-guided modification. The paper evaluates six popular LLMs using the CALM framework and finds that while some models demonstrate notable fairness in judgment, significant biases persist in certain specific tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I think this paper is more like a toolkit paper rather than a novel research paper, as they just integrate 12 types of existing biases in LLM-as-a-Judge. If we look at the appendix B, we can find that each of the 12 types can be referenced to another previous paper.\n- The paper primarily relies on automated metrics to assess bias, but human evaluation could provide a valuable additional perspective. Incorporating a human evaluation benchmark would strengthen the validation of the findings." }, "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": "The discussion includes prejudice against certain groups such as “homosexual,” “black,” “female,” and “HIV-positive.” HIV-positive.” I would be concerned that there would be some impact on particular groups." }, "flag_for_ethics_review": { "value": [ "Yes, Discrimination / bias / fairness concerns" ] }, "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 source of the basis for these assessments of Robustness Rate (RR) and Consistency Rate (CR)? Why are human correlations such as Pearson's coefficient not considered in the assessment. \n2. LLM does not take into account some of the popular LLM as Judge models, such as pandaLM, Prometheus, etc. LLM lacks a specific LLM as Judge model. Lack of specific LLM as Judge evaluation. \n3. Is the data randomly selected? For example, GMSK, how to choose 100 pieces of data from tens of thousands of pieces of data? How to prove that these 100 data are representative enough?\n4. Is it user friendly? Is the reproduction cost prohibitive?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. A comprehensive delineation and classification of twelve specific biases that can compromise the dependability and credibility of LLM-as-a-Judge.\n2. The proposal of the CALM framework for assessing biases within LLM-as-a-Judge systems, which enhances the rigor of the assessment process in a person-independent manner.\n3. An in-depth analysis of six widely-used LLMs through the lens of the CALM framework." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study identifies 12 significant potential biases and introduces a novel automated bias quantification framework called CALM. This framework systematically quantifies and analyzes each type of bias in LLM-as-a-Judge through automated, principle-guided modifications. Empirical findings indicate that there is still room for enhancing the reliability of LLM-as-a-Judge. The paper explores both the explicit and implicit impacts of these biases and offers recommendations for the dependable application of LLM-as-a-Judge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Lack of Transparency in Assessment Criteria: The source of the basis for the assessments of Robustness Rate (RR) and Consistency Rate (CR) is unclear. \nIncomplete Consideration of Popular Models: The evaluation does not include some well-known LLM as Judge models, such as pandaLM and Prometheus. This omission suggests a lack of thoroughness and may lead to biased or incomplete conclusions.\nQuestionable Data Selection Process: The method for selecting data is not well-defined. For instance, in the case of GMSK, the process of choosing 100 pieces of data from tens of thousands is not explained. This raises concerns about the representativeness and reliability of the selected data.\nUser-Friendliness and Reproduction Costs: There are concerns about the user-friendliness of the system and whether the costs associated with reproducing the results are prohibitive. This could limit accessibility and practical application for users." }, "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": "073: This is not limited only to \"humanities, social sciences, or general knowledge\" fields, a refined answer could be in any field or task.\n\n281: If i understand correctly, for CR you ask the LLM to generate two responses for the same prompt, and check if the answers are consistent. If so, why not use more than two responses? You can make the consistency measure more robust by comparing the variance of N>>2 generations. \n\n295: Which ACC do you use as a metric?\n\nMetrics paragraph: In my opinion, and even though the English is good and I understand each sentence, this paragraph is not clear enough. I would emphasize in the text which metric is used for each task, specifically at the start of the paragraph, and refer to the column in the table. In addition, the names, abbreviation are not consistent throughout the paper and specifically in the figures. Please see the weaknesses regarding the metrics." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I believe the paper is strong, well-written, and highly comprehensive. The topic is both timely and important, and the NLP/LLM community would greatly benefit from its publication. In my opinion, this paper should be accepted. The reason I initially rated it a 6 instead of an 8 is to encourage the authors to consider revising the metrics (as discussed in the weaknesses section)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors identify 12 distinct types of biases in LLM-as-a-Judge models, provide automated code for \"checklisting\" these biases in LLM judges, and conduct a comprehensive analysis of how these biases manifest in state-of-the-art LLM judges across various datasets. The key finding is that LLM judges are not robust. The implications of these findings are significant and should be effectively communicated to the community." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Metrics**: I have a few suggestions regarding the metrics used in this paper and how they are presented in the results. First, for the RR and CR metrics, I recommend making the CR metric more robust by sampling multiple generations when computing the CR for a given instance. Additionally, I propose adjusting the RR metric with the CR, as the authors note that LLMs are non-deterministic, and a low RR score might reflect this rather than a genuine lack of robustness. To adjust the score for each individual instance, I would subtract the individual CR from the individual RR. It is important to compute this adjustment at the individual level, as the CR varies between instances. The final score would be the dataset average of RR_i - CR_i. While this metric is more complex and falls within a -1 to 1 scale, it is far more reliable. If you choose not to present the difference between the two, I suggest at least presenting the average RR and average CR for each LLM in the results. The CR is not a constant value; it varies between models and across instances.\n\nRegarding the ACC metrics, I am unclear about which specific metric you are using in the results section. Should we compare the two metrics and examine the variations between them? Additionally, is this score presented as an absolute value? Could you clarify this aspect to ensure an accurate interpretation of the results? And why do you use \"hack\"? Isn't it essentially the CoT ACC? \n\nRegarding the Error Rate (ER) metrics, could you explain the rationale for using these metrics instead of the RR/CR in the paper? Additionally, could we apply the ER metrics to detect other forms of bias? I also find the ER_SE metric unreliable. From your description, it appears that Y_other represents the score assigned by other models to the evaluated model's response. However, I believe Y_other should represent the average score assigned by the explained model to the responses of other LLMs. This would better measure whether the LLM prefers its own responses. Otherwise, you're merely capturing the LLM’s general bias relative to the consensus. For example, one LLM might use scores in the range of 3-7, while another uses 1-6, yet they could still achieve a perfect Spearman's correlation. Moreover, why do you use absolute value? This can be misleading. For example, y_self, y_other = (5, 3) is the same as (1, 3). I believe you can think on a better metric for ER. \n\nIn general, I would recommend adjusting the metrics so that higher scores indicate more bias, unlike the current ones where higher scores represent robustness. Since you frequently use the term \"bias\" throughout the paper, this modification could make the results and the interpretation of the metrics more intuitive and easier to follow. \n\nIf the authors revise the metrics and provide this analysis, or at least clarify what I may be misunderstanding, I would be happy to increase the overall rating from 6 to 8.\n\n**Misinterpretations of the Results:** Specifically, the statement \"Bias is more pronounced in the alignment dataset compared to the fact-related dataset\" cannot be inferred from the results. The biases in these datasets differ, and you need to compare like for like - either the same bias across different datasets or the same dataset with different biases. While it's possible to compare the CR metrics between datasets (difference of differences), I believe this is insufficient on its own. First, I would like you to clarify (both here and in the paper) the rationale behind distinguishing between biases and datasets, as well as why certain biases may not be applicable to all datasets. I haven’t given this much thought, but it is critical that this distinction is explicitly explained in the paper, rather than leaving it up to the reader to infer." }, "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": "- How do we know that evaluations that rely on perturbations by LLMs can be trusted? How do we know that such perturbations do not introduce errors or biases other than those which are intended to be evaluated?\n- How could one interpret the results of evaluating a bias using the CALM framework in terms of its effect on real-life applications of the LLM-as-a-judge system? For example, if one LLM's robustness rate for diversity is 0.1 greater than another's, how does this the actual treatment of minorities by systems that utilize the LLM in a LLM-as-a-judge application?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper covers a comprehensive list of biases and conducts experiments on many state-of-the-art LLMs and across multiple relevant domains such as fact-based and alignment-related data.\n- They introduce a novel method for evaluating biases in LLM-as-a-judge systems that is well-principled and automatic. It is also flexible as the framework could even be extended to bias types that are not considered in the paper. \n- They systematically demonstrate that current LLMs are still susceptible to various biases. As far as I am aware, many of their evaluation results are completely novel, such as demonstrating how different types of fake-authorities interfere with LLM-judges to varying degrees." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces CALM, a framework for measuring bias in LLM-as-a-judge applications. The framework works by modifying the prompt or response that is to be judged by introducing a bias, and then measuring how this modification affects the judgement. They propose a classification of biases into 12 distinct types, each of which can be measured by their framework. The introduces typology covers a broad spectrum such as bias based on the length of answers, the use of an authoritative tone or faux citations. \n\nTo evaluate the magnitude of these biases when current LLMs are used as judges the paper introduces various metrics, most importantly the robustness of a judgement when a bias is added to an answer. Biases are evaluated on three types of datasets which are sampled from existing sources. They cover factual data for which responses should be evaluated according to factual accuracy, alignment related data for which judgements depend on user preferences, and refinement aware evaluation data which contains pairs of responses in which one is a refinement of the other.\n\nFor its main results, the paper evaluates the biases of multiple state-of-the-art LLMs when used in an LLM-as-judge system. The results demonstrate that current models are susceptible to various biases in their judgements. Some noteworthy findings include:\n- All models are significantly impacted by position bias.\n- Most models judge their own output more favorably than that of other models, even when sources are anonymized.\n- Different types of fake-citations influence an LLM's judgement to various degrees. Quote- and book formats have a higher chance of being convincing than URL citations.\n\nThese as well as other findings are discussed in the results section." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "# Main Concerns\nI have two concerns related to the soundness of the paper's methodology and experimental results. \n\nThe paper introduces a new method for evaluating biases but does not evaluate the trustworthiness of this method. The method is based on perturbing the responses that an LLM-as-a-judge is supposed to evaluate. In some cases this evaluation works via a separate LLM such as for the verbosity bias, fallacy-oversight bias, or authority bias where GPT-4-Turbo is used to rephrase a response. How do we know that using an LLM to modify responses does not introduce errors or other features that manipulate the judge's decision? While I can believe that GPT-4-Turbo is capable of applying the required modifications, this should be experimentally verified so that the results have scientific rigor. \n\nFurther, the paper provides scores of LLM-judge biases in the form of the robustness rates, but I can not tell what these scores mean for real-life applications of said LLM-judges. For example, LLM-as-a-judge is typically used for evaluation, such as a reward model during RLHF. If my LLM-as-a-judge system has a specific robustness score for a bias type such as diversity, how does this translate to the bias of a downstream system, such as an LLM that was trained using the judge? Without such results, it is unclear how to interpret the paper's numerical results.\n\nSumming up, I believe two types of experiments are necessary to improve the paper's soundness. \n- Demonstrate that perturbation using LLMs does not introduce unintended errors or biases that are different from what is intended.\n- Evaluate the effect of different bias scores for different LLM-as-a-judge systems on real-life applications of the systems.\n\nIf related experiments or arguments are added, I will improve my score.\n\n# Minor Comments (did not impact score)\n- It would be helpful to have example questions and responses from each dataset somewhere in the paper, to illustrate what the difference between e.g. the fact-related and alignment-related dataset is. They could be included in the appendix if there is no space in the main paper.\n- Figure includes some plots where a larger value of the y-axis means less bias (robustness rate) and some for which is a lower value is better (error rate). The plot would be easier to read if this was indicated somehow." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024justice,\ntitle={Justice or Prejudice? Quantifying Biases in {LLM}-as-a-Judge},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3GTtZFiajM},\nnote={under review}\n}" }, "abstract": { "value": "LLM-as-a-Judge has been widely utilized as an evaluation method in various benchmarks and served as supervised rewards in model training. However, despite their excellence in many domains, potential issues are under-explored, undermining their reliability and the scope of their utility. \nTherefore, we identify 12 key potential biases and propose a new automated bias quantification framework—CALM—which systematically quantifies and analyzes each type of bias in LLM-as-a-Judge by using automated and principle-guided modification. Our experiments cover multiple popular language models, and the results indicate that while advanced models have achieved commendable overall performance, significant biases persist in certain specific tasks. Empirical results suggest that there remains room for improvement in the reliability of LLM-as-a-Judge. Moreover, we also discuss the explicit and implicit influence of these biases and give some suggestions for the reliable application of LLM-as-a-Judge. Our work highlights the need for stakeholders to address these issues and remind users to exercise caution in LLM-as-a-Judge applications." }, "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", "LLM-as-a-Judge", "trustworthy LLM", "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/39c4fdd7f0cb600718984348a99acfb81ae33dfd.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/29f10f8da6359a118f02cccabb8e3cf0919a921d.zip" }, "title": { "value": "Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge" }, "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": "Please refer to the weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well-organized and easy to follow.\n- The authors conduct sufficient ablation studies to evaluate the proposed modules.\n- The experiments demonstrate the training efficiency of the proposed method and its capability to unify various visual conditions for generation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes CtrLoRA for better controllability of the conditional image generation. This framework trains a Base ControlNet for the general image-to-image generation and then uses the LoRA fine-tuning for specific user instructions. Experiments show the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The authors train a base ControlNet for the subsequent LoRA fine-tuning. However, why not directly fine-tune a pre-trained ControlNet or Uni-ControlNet?\n\n- Lack of comparison to: ControlNet++[1].\n\n- The paper does not explore whether this method can be generalized to other diffusion models such as SDXL and Pixart.\n\n[1] Li M, Yang T, Kuang H, et al. ControlNet++: Improving Conditional Controls with Efficient Consistency Feedback[C]//European Conference on Computer Vision. Springer, Cham, 2025: 129-147." }, "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 results in Figure11b demonstrate that the different conditions are effectively disentangled, with a direct summation module according to Figure 3c. Could you clarify why this module is effective, such as presenting the results of two elements both separately and after sum-up.\n\n2. A detail, why not presenting all 9 base-condition results comparison to UniControl in Table 2?" }, "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 focus on an important problem, extending ControlNet to a lightweight manner.\n2. Experimental results are impressive, especially the convergence experiment." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes CtrlLoRA, a two-stage parameter-efficient fine-tuning pipeline, to ease the original ControlNet's computation burden in terms of different conditions. The authors evaluate CtrlLoRA through extensive experiments by both the quality and the computation efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In line 70, the authors state that ControlNet with Canny edges requires 3 million images over 600 GPU hours for one condition. In contrast, line 244 indicates that Base ControlNet necessitates millions of images for 6000 GPU hours for 9 conditions. Although it is not fair enough, but it implies that the proposed method does not significantly reduce the computational burden.\n\n2. In line 239, the mechanism of training with 9 conditions is not clear enough. As different conditions have different levels of sparse information of input images, why they have equal training iterations? And continuous shifting between different conditions may make the training hard.\n\n3. the motivation why the new conditions are not trained as the Base ControlNet by a shifting mechanism is not clear enough.\n\n4. Most results are from \"Base CN + CtrlLoRA'', and results from \"Community Model + CtrlLoRA\" in Figure 11a are rare, not enough to convince that CtrlLoRA is effective when transferring to other community models.\n\n5. pretrained-VAE seems to be only an interesting trick.\n\n6. putting all the prompts in the appendix makes reading inconvenient." }, "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": "- Regarding the discussion of \"Adaptation to new conditions,\" while training a comparison method from scratch with a small amount of data may indeed result in slow convergence, what would be the results if we used a pre-trained conditional model (analogous to possessing a Base ControlNet) for fine-tuning?\n\n- I'm curious about the performance between a pre-trained controlnet model available in the community and a model trained using proposed \"Base + LoRA\" with same conditions." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- To address the high cost of separately training different models for conditional generation tasks, this paper proposes a training method that transitions from a base controlnet model to a lightly fine-tuned lora model. This approach ensures generation quality while achieving a faster convergence rate.\n\n- The paper shows many analyses of the proposed method and presents the results generated for a total of more than a dozen conditions.\n\n- The paper is well structured and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper draws on the idea of combining a base model with PEFT (Parameter-Efficient Fine-Tuning) for controllable generation. It trains a Base ControlNet obtained through several condition-specific training processes, and then fine-tunes it with a small amount of data for newly introduced conditions to obtain different condition-specific LoRAs. This approach improves the efficiency of training new condition generators at a lower cost." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper primarily aims to improve the training efficiency of all kinds of conditional models, hence it employs a series of LoRAs to train the newly introduced conditions based on the \"Base ControlNet\". However, there is relatively little comparison and discussion of existing methods that efficiently train ControlNet, such as T2I-Adapter, ControlLoRA, and SCEdit.\n\n- There currently exists a viable **controlnet-union** model, which can handle different conditions using a single model. This may be a higher-level representation of the training of the \"Base ControlNet\" model discussed in the paper. On the other hand, the use of LoRA for fine-tuning is relatively straightforward and has been implemented in previous community works, such as ControlLoRA. In comparison, the overall innovativeness of the paper is limited.\n\n- The paper does not discuss how many conditions to use or how to select conditions for training the \"Base ControlNet\" to achieve optimal knowledge transfer effects." }, "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": "1. Consider specifying 1-2 new image conditions and key metrics (e.g., adaptation speed, data efficiency, performance) for comparing UniControl [1] fine-tuning to CtrLoRA. This would provide a clear, focused comparison.\n2. Additional baselines are required for each base image condition. Comparisons should be made with a fully trained ControlNet, which has been trained exclusively under a single image condition, to establish a more comprehensive benchmark.\n3. Similarly, for the new condition, it is essential to compare the performance of CtrLora against ControlNet when ControlNet has been fully trained on a single modality. This will provide a clearer understanding of their relative efficiencies.\n4. It would be beneficial to explore how the number of image conditions used during the training of the base ControlNet affects its ability to learn new conditions. Insights into the scalability and adaptability of the base network could prove crucial for future applications.\n5. I have noted that CtrloRA can perform low-level image enhancement tasks, such as low-light image enhancement. Could the authors demonstrate how CtrloRA performs in comparison to other diffusion models for low-light image enhancement? This could highlight potential advantages or unique features of CtrloRA in practical applications.\n\n[1] UniControl: A Unified Diffusion Model for Controllable Visual Generation In the Wild." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The CtrloRA framework introduced in this paper allows users to quickly and efficiently fine-tune the ControlNet to new image conditions, with minimal resource consumption. The experimental results validate the effectiveness of this method. Additionally, the paper is well-structured and clearly written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a CtrloRA framework. This framework starts by training a basic ControlNet that handles various image conditions efficiently. With this trained network, one can quickly fine-tune it to adapt to new conditions using a task-specific LoRA—specifically, fine-tuning requires only 1,000 paired images and less than an hour on a single GPU. The experimental results confirm that this method greatly speeds up the training process for new image conditions. Based on these impressive findings, I recommend a weak acceptance. However, there are some unclear points and missing experiments in the paper (see the Question section), and my final decision will depend on the authors' responses to these issues." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are some unclear points and missing experiments in the paper (see the Question section), and my final decision will depend on the authors' responses to these issues." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024ctrlora,\ntitle={CtrLo{RA}: An Extensible and Efficient Framework for Controllable Image Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Gga05Jdmj},\nnote={under review}\n}" }, "abstract": { "value": "Recently, large-scale diffusion models have made impressive progress in text-to-image (T2I) generation. To further equip these T2I models with fine-grained spatial control, approaches like ControlNet introduce an extra network that learns to follow a condition image. However, for every single condition type, ControlNet requires independent training on millions of data pairs with hundreds of GPU hours, which is quite expensive and makes it challenging for ordinary users to explore and develop new types of conditions. To address this problem, we propose the CtrLoRA framework, which trains a Base ControlNet to learn the common knowledge of image-to-image generation from multiple base conditions, along with condition-specific LoRAs to capture distinct characteristics of each condition. Utilizing our pretrained Base ControlNet, users can easily adapt it to new conditions, requiring as few as 1,000 data pairs and less than one hour of single-GPU training to obtain satisfactory results in most scenarios. Moreover, our CtrLoRA reduces the learnable parameters by 90% compared to ControlNet, significantly lowering the threshold to distribute and deploy the model weights. Extensive experiments on various types of conditions demonstrate the efficiency and effectiveness of our method. All codes and model weights 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": [ "Controllable Image Generation", "Image-to-Image Generation", "ControlNet", "LoRA", "Resource-Efficient 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/3a5e48f0c23fbf52c7f2a14ab218f0738fdcaad5.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": "CtrLoRA: An Extensible and Efficient Framework for Controllable Image 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": 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": "Can you add more description of how self-contextualization works? This is the identical contextualization prompt just uses the LLM agent model (e.g. Gemini-1.5-flash) instead of the trained Phi-3-instruct model. \n\nTable 3 in appendix: Data and caption do not match. 33*15 = 495. Are the numbers the number of successful demonstrations collected? Some more information on the demonstration collection would be helpful. \n\nDoes figure 8 include both successful and failed tasks? -> Are the distributions over the same tasks? \n\nLine 182: “select the one that provides the most relevant context for the LLM agent to accurately predict the next action at as the target.” \n- This could be written more clearly. \n\nWill the model being released?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "LCoW shows a very clear benefit and improvement to LLM web agents. \n\nLCoW shows state of the art improvement on WebShop against strong baselines such as AgentQ and LASER. This is comparable to human expert level on WebShop tasks. \n\nThe experiments are pretty comprehensive showing improvement on different benchmarks with different agents and show continuous benefit for up to 3 iterations of LCoW training. \n\nThis method does not seem like it would add a ton of compute expenses and could be quite practical. \n\nThe paper is also well written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "LCoW aims to advance LLM based web agents by adding a contextualization step to the webpage observation in which an LLM reduces the html/raw observation by pruning irrelevant elements and adds contextual descriptive information. This significantly improves the performance of the downstream web agent and sets state of the art results. \n\nThe algorithm works by first collecting successful trajectories as ground truth. Then a contextualizer model (with a prompt instructions) is used to reduce and explain the UI elements. This now contextualized observation is give to a set of LLM agents that produce actions. The contextualized model gets a high reward if the agents give the same action as in the successful trajectory. The best contextualized observation is then collected. Finally the model is trained to produce the \"good\" contextual observations. This can be repeated for multiple iterations. \n\nThe main contributions are:\n- The novel approach to LLM-based contextualization and parsing, enabling state of the art performance on web agent datasets. \n- The algorithm for training the contextualization model\n- The prompt for contextualization\n- Many experiments\n\nTheir results include experiments on\nWebShop and WebArena across multiple LLM agents with strong baselines in WebShop (such as AgentQ and LASER). \nThey also include ablations/analysis on how the action matching reward improves with iterations, how LCoW affects the number of steps required for each task, and comparisons of the original collected trajectories against behavior cloning," }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are a few weaknesses of the proposed method, though some of this is more limitations than reasons to reject. \n\n1) It is unclear how this method translates across websites, domains, and to some extent tasks. Since this involves training the contextualization model, there is potential to overfit to the data available. It would be nice to have some experiments showing that LCoW trained on a few domains generalizes to many other domains. Perhaps LCoW when generalizing to a task on say LinkedIn after only being trained on a couple benchmarks. \n\n2) This is also true for generalizing over tasks as well. Perhaps LCoW fails when extending to tasks that require more contextual reasoning. \n\n3) The training details section notes that action matching based on parsing is infeasible for open-ended actions (such as filling in a text box) and uses GPT-4o to do matching. However, the bigger limitation is on open-ended tasks or tasks that have diverse ways/orders of completing them. How would LCoW when there are many actions that are reasonable? \n\n4) In real-world situations there may be many rollouts that include individual actions that are actually incorrect. LCoW would treat these as correct and may even train the model to drop the truly relevant areas of the page. \n\n5) The limitations section only notes novel UI elements as a limitation. It seems the limitation section should be expanded to cover some of the above concerns as well. \n\n6) This approach also relies on being able to collect successful trajectories, whereas other methods that employ search may be able to extend agent capabilities to new tasks.\n\n7) There are no experiments comparing to code-based html parsers for LLM agents. Though they are undoubtably not as good or there would already be models with performance comparable to LCoW. \n\n8) How long does the contextual observation generation take? In tasks that rely on parsing large amounts of text (e.g. Write a tweet based on this article), regenerating and contextualizing a whole article could be expensive, time consuming, and not necessary. (This should be addressed)\n\nThere are a few nice to have experiments that are not present: \nGeneralization across task type or difficulty\nGeneralization across websites" }, "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 reward obtained from multiple LLMs is only used to judge whether the current step correctly predicts the real action. Should the final expectation of task be used?\n2.How does LCoW handle web pages with novel UI elements or layouts that were not encountered during training?\n3.Have any measures been taken to prevent overfitting, particularly given the iterative training process that relies on self-generated data?\n4.Can the web page be partitioned and analyzed through the prompt function? And there is no comparison with the previous intelligent code analysis work." }, "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 comprehensive experiments on popular benchmarks, demonstrating that LCoW significantly improves the performance of LLM agents across various scales. The success rates surpassing human experts are particularly impressive.\n2.The paper shows that the contextualization module trained with LCoW can generalize well to different LLMs, including those not involved in the training process, which is a strong indicator of the method's robustness.\n3.The proposed iterative algorithm for training the contextualization module is a effective approach that allows for continuous improvement of the module based on real-world interactions and feedback." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces LCoW, a novel framework that addresses the challenge of enhancing decision-making capabilities of Large Language Models (LLMs) in the context of web automation tasks. The method distinguishes the comprehension of web content from the decision-making process by training a dedicated module that creates contextualized representations of intricate web pages." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.As mentioned in the paper, the contextualization module struggles with web pages containing UI elements not seen during training. This limitation could be a barrier to the framework's real-world applicability, especially given the vast diversity of web page designs.\n2.The contextualization module was trained on a relatively small dataset of fewer than 2,000 self-generated samples. Can the model's ability to generalize to a broader range of web pages and tasks?\n3.The paper does not extensively discuss the potential for overfitting, especially given the iterative training process that relies heavily on self-generated data. There is a risk that the model may perform well on similar tasks but fail to adapt to new, unseen scenarios.\n4.The contextualization module shown in Figure 7 is not intuitive enough." }, "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": "- How do you know the performance gain is enhanced by the LLM's decision-making?\n- Figure 1: How do you select the reported tasks from WorkArena?\n- How do you ensure that crucial elements are not removed during the contextualization?\n- It seems the contextualization module could only be trained with successful trajectories from LLMs. What about those tasks that even those LLMs could not fulfill?\n- The WorkArena contains up to 1,000 task instances, why do you evaluate only 115 tasks? How do you select them?\n- If the contextualization is shopping relevant, do you believe it's less convenient to write several human oracle rules than to train a contextualization module?\n- Figure 8: The mean of the number of steps doesn't seem to differ much. What is the exact number?\n- It seems the behavior cloning baseline is trained with fewer demonstrations than the contextualization module." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The observation contextualization idea is clever, and the training of the contextualization doesn't require human-labeled data.\n- The results reported on the shopping tasks are promising, proving that the idea should work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes to contextualize the observation of LLM-based online shopping agents to improve their performance. It trains a task-observation-reliant contextualization module to help locate the most important information on a page and provides explanations. The idea is clever and shows promising results on two shopping benchmarks. However, it doesn't include code or any playing episodes for the reviewers to verify the outcomes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I like the observation contextualization idea, but I've seen a highly ranked paper on the WebArena benchmark, a benchmark with a wider type of websites defined other than mere shopping here, using a similar but more general method that doesn't require task-related inputs. I believe the strong reliance on the web observation's format, as you mention in the limitation section, \"it often struggles to provide suitable contextualization for web pages containing UI elements unseen during the training,\" constrains this work's scope on shopping-related tasks only.\n- I would question your results as you didn't include code or episodes for reviewers to verify your conclusions, especially when you only play your agents on a partial WorkArena benchmark, whose results are easily controllable if you only select the tasks where your agents win.\n- I think you have sacrificed the agent's generalizability with the contextualization module specifically trained on the shopping tasks, as you put in the appendix." }, "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": "* Explain the performance of different choices of LLM-based contextualization modules\n* Discuss the practical efficacy of these said agents given the cost/tokens for using the reward models and the lack of simulation environment for new unseen web sites." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Pipelines for LLM-based web agents are complex and the proposed approach breaks down \"contextualizing\" the web pages separately from decision making ability of the agents. The approach of leveraging simulation data across the web environments to train (fine tune) a small model for contextualizing shows good results. The reward model is in essence a LLM-based judge system across multiple powerful LLMs. The qualitative results are interesting as they highlight how the proposed contextualization module works in removing irrelevant components of the web page." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes training a language model to contextualize complex web pages for improving the success rates of LLM-based web agents. To enable this the proposed method uses the web simulator environments to roll out multiple trajectories and uses multiple LLMs to score the different candidates. This strategy provides a significant improvement over baseline open and closed-sourced models across different benchmarks like WebShop and WebArena." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* With respect to generalization capabilities, the study can be strengthened by demonstrating performance across the different web environment bechmarks or types of web pages (e.g, instead of picking or holding out 500 examples randomly the type of Web tasks could be used for creating the train/test held out set). \n\n* Additionally, it is not clear if in the real world environment a simulation environment is available to bootstrap and roll out the candidate sequence of state and action(s). As listed in the limitations section the power/promise of these agents diminishes given that performance drops when dealing with unseen UI elements. \n\n* For different web benchmarks, different LLMs were used for training the contextualization module. This needs to be explained or justified." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Training language models to automatically contextualize web page observations to enhance the decision-making of LLM agents" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning to Contextualize Web Pages for Enhanced Decision Making by {LLM} Agents},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Gzz7ZQLiz},\nnote={under review}\n}" }, "abstract": { "value": "Recent advances in large language models (LLMs) have led to a growing interest in developing LLM-based agents for automating web tasks. However, these agents often struggle with even simple tasks on real-world websites due to their limited capability to understand and process complex web page structures. In this work, we introduce LCoW, a framework for Learning language models to Contextualize complex Web pages into a more comprehensible form, thereby enhancing decision making by LLM agents. LCoW decouples web page understanding from decision making by training a separate contextualization module to transform complex web pages into comprehensible format, which are then utilized by the decision-making agent. We demonstrate that our contextualization module effectively integrates with LLM agents of various scales to significantly enhance their decision-making capabilities in web automation tasks. Notably, LCoW improves the success rates of closed-source LLMs (e.g., Gemini-1.5-flash, GPT-4o, Claude-3.5-Sonnet) by an average of 20%, and demonstrates a 33.5% average improvement in success rates for open-source LMs (e.g., Llama-3.1-8B, Llama-3.1-70B) on the WorkArena benchmark. Moreover, the Gemini-1.5-flash agent with LCoW achieves state-of-the-art results on the WebShop benchmark, outperforming human experts." }, "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", "LLM agent", "Web automation" ] }, "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/917a517ab6936b1f63a1e8f4a62ed4d8566566a9.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": "Learning to Contextualize Web Pages for Enhanced Decision Making by LLM Agents" }, "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": "What happens to your method’s performance when it is run on noisy quantum hardware? \n\nHow resource-intensive is it to make your method fault-tolerant?\n\nWhy did you not try HAM and PEFF on the Fashion-MNIST dataset? Apologies if this was already covered, but I struggled to find it." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Strengths: The paper is exceptionally well-written. The exposition is clear, the protocol is well explained, and the background well-done. The idea to embed states in a Hamiltonian and then leverage VQE is interesting as well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Summary: In this paper, the authors put forward a new quantum machine learning (QML) classifier. Their new approach leverages the variational quantum eigensolver algorithm to find a parametrized quantum circuit that classifies inputs based on their expectation value. Underlying their new approach is their idea to embed input states within a Hamiltonian." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses: The article has three main weaknesses.\n - No theoretical performance guarantees: There are no theoretical results guaranteeing any a super-polynomial improvements over classical methods. \n - Weak results on simulated data: The new method is routinely outperformed by other methods including logistic regression on both text datasets. Sure, some of the competitors require more parameters, but none of the other models are honestly that big. These results would be far more compelling if there was a clear super-polynomial advantage with the new method.\n - No discussion of robustness to noise: This is a big one. There’s no discussion or evaluation of the methods robustness to noise. Unless you’re proposing a fault-tolerant algorithm, you need to discuss noise and compare your method’s performance in the presence of NISQ-era levels of noise to classical competitors." }, "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 to choose the {${P_j}$}$_{j=1}^p$ in Eq.12?\n2. is there any theoretical analysis that the generalization error is bounded with respect to the number of finite terms $P$?\n3. instead of using parametrized quantum circuit, whether is it enough to only optimize the the parameters in measurement (parametrized hamiltonian? like https://arxiv.org/pdf/2307.10560)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "it is well-written and clearly presents the concept and idea." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the Hamiltonian classifier, a quantum machine learning approach that enhances the data encoding by mapping inputs to Pauli strings, and provides related proof-of-principle experiments to demonstrate the advantages." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. the similar idea was proposed in Post-variational quantum neural networks (https://arxiv.org/pdf/2307.10560)\n2. missing some theoretical analysis of the performance of the proposed model." }, "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 the complexity analysis, section 3.6, you mention that your method incurs an additional $1/\\epsilon^2$ cost. Is this the same for other methods presented in Table 1?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper is clearly written and scientifically sound. The improvement in gate complexity is substantial, since no other methods achieve logarithmic scaling for these quantities.\nI believe the paper provides a good advance in the field of variational quantum circuits." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors provide a new method to implement variational quantum circuits (VQCs) that can be used for machine learning tasks using quantum hardware. This method achieves a logarithmic scaling in both qubits and quantum gate counts, while having worse sample complexity (as presented in Table 1). They then numerically simulate their quantum algorithm on text and image classification, and achieve results that are mostly better than other forms of VQCs and on-par with classical neural networks (such as MLPs and CNNs)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main critique of this paper is that the method achieves mostly the same performance than very simple classical models, such as logistic regression, MLPs and CNNs, while being based on having access to ideal quantum hardware which is not readily available (and will only be made available in the long-term). Therefore, the main value of the paper is an algorithmic advance for VQCs (and in how data is encoded into the quantum computer), but not for the global field of machine learning itself.\nSimulations that include noise could be useful, which hopefully would show that even on noisy quantum hardware the results are not significantly altered.\nThe benchmarks are somewhat sparse, and some of them include results that have 100% test accuracy for many methods, so we cannot know which is better. Considering more difficult datasets would be more interesting.\nThe maximum number of qubits considered is $n=10$ if I am not mistaken, which is significantly smaller than what can be simulated with even small computational resources ($n=20$ is feasible in the noiseless setting)." }, "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": "Here are some minor questions:\n(1) On Page 6, authors claimed that they can randomly define $p$ Pauli strings, and decompose the data matrix $H_{\\phi}(\\tilde{x})$ onto the sampled Pauli basis. Then, is there any creteria on selecting these Pauli strings? What is the scaling of the parameter $p$, and what is the relationship between $p$ and the power of Hamiltonain classifier model (such as generalization error upper bound or the effective dimension)? \n(2) In equation 10, $\\alpha_i$ has an exponentially small factor $2^{-n}$. Does this factor cause the measurement accuracy to become exponentially small, leading to a large amount of measurement overhead?\n(3) In Table 1, Cong et al. (2019) do not utilize the QCNN to solve a classical task; instead, they predict the quantum phase transition problem. Is it fair to compare Cong et al. (2019) to the authors' work on a classical task?\n(4) In Tables 2 and 3, it is observed that the proposed methods (HAM, PEFF, and SIM) do not outperform all the listed methods. Given these facts, what is the advantage of the proposed Hamiltonian classifier method?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "From a high-level perspective, this paper is well-written, clearly introducing the research background and presenting their results. Meanwhile, the authors have put significant effort into the numerical simulation section, where they compare various related quantum-classical methods across several datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper argues that previous NISQ quantum machine learning algorithms generally require a large number of computational resources, which may not be suitable for NISQ devices. To address the data-encoding problem, the authors encode classical data into a physical Hamiltonian, referred to as the Hamiltonian classifier. To demonstrate the power of the Hamiltonian classifier, they numerically test their model on several popular classical datasets, including text and image classification tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) One of the main concerns is that the \"Hamiltonian Classifier\" has been proposed and studied in several papers, such as [S. Jerbi et al., Shadows of Quantum Machine Learning, Nat. Comm., 2024] and [Y. Song et al., A Quantum Federated Learning Framework for Classical Clients, Sci. China Phys. Mech. Astron. (2024)]. However, this paper does not cite these highly relevant works. As a result, the authors' contributions are significantly diminished, especially the claim in Sec. 3, page 4, \"To the best of our knowledge...\"\n\n(2) As a high-level conference (ICLR) in the field of machine learning, we expect to see more surprising results in quantum machine learning. The authors still utilize the standard classical-quantum workflow, despite encoding the data into a physical Hamiltonian. While many papers adopt this framework and numerically benchmark their methods on some datasets, such research contributes very little to quantum computation theory and the quantum machine learning community, particularly at this stage. This is supported by recent findings: it has been shown that many classical-quantum QML methods (including the authors') are classically simulable when the model does not suffer from the barren plateaus phenomenon [A. Angrisani et al., Classically Estimating Observables of Noiseless Quantum Circuits, arXiv:2409.01706]. Furthermore, when the QML algorithm is limited to a 2D architecture, all constant-depth (or constant evolution time) quantum-classical approaches can be classically simulated [S. Bravyi et al., Classical Algorithms for Quantum Mean Values, Nat. Phys., 2021]. From this perspective, it appears that the Hamiltonian classifier method may not provide a clear quantum advantage." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "An efficient quantum machine learning method named \"Hamiltonian classifier\" that achieves logarithmic complexity in both qubits and gates by representing inputs as measurements rather than using traditional state preparation" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024an,\ntitle={An Efficient Quantum Classifier Based on Hamiltonian Representations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3HPOtZxs5s},\nnote={under review}\n}" }, "abstract": { "value": "Quantum computing shows great potential for expanding the range of efficiently solvable problems. This promise arises from the advantageous resource and runtime scaling of certain quantum algorithms over classical ones. Quantum machine learning (QML) seeks to extend these advantages to data-driven methods. Initial evidence suggests quantum-based models can outperform classical ones in terms of scaling, runtime and generalization capabilities. However, critics have pointed out that many works rely on extensive feature reduction or use toy datasets to draw their conclusions, raising concerns about their applicability to larger problems. Scaling up these results is challenging due to hardware limitations and the high costs generally associated with encoding dense vector representations on quantum devices. To address these challenges, we propose an efficient approach called Hamiltonian classifier inspired by ground-state energy optimization in quantum chemistry. This method circumvents the costs associated with data encoding by mapping inputs to a finite set of Pauli strings and computing predictions as their expectation values. In addition, we introduce two variants with different scaling in terms of parameters and sample complexity. We evaluate our approach on text and image classification tasks, comparing it to well-established classical and quantum models. Our results show the Hamiltonian classifier delivers performance comparable to or better than these methods. Notably, our method achieves logarithmic complexity in both qubits and quantum gates, making it well-suited for large-scale, real-world applications." }, "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": [ "quantum computing", "quantum machine learning", "variational quantum circuits", "quantum encoding" ] }, "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/449f128da0f62da574fd361fc9ca4025f2e35d32.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": "An Efficient Quantum Classifier Based on Hamiltonian Representations" }, "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": null, "code_of_ethics": null, "comment": null, "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": null, "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": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "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": "Adding additional evaluations with mode-free attack baselines as well as defense strategies will significantly improve the paper." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The idea of combining different types of attacks to enhance the performance of the primary attack is interesting.\n2. The performed evaluations mostly cover the key points made in the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem of conducting combined attacks by leveraging some secondary of different type to enhance the performance of the primary attack. For example, attackers can use property inference attacks first to better guess the target property distribution and then generate higher quality auxiliary datasets to further assist in the performance of the attribute inference attacks. The attack results indicate that the combined method outperforms the single attack, indicating the the strength of attacks in practice can be much more effective." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The performance evaluation is heavily focused on leveraging shadow model based attacks. I think a comprehensive evaluation should also include some model-free attacks (e.g., LiRA for membership inference attacks), and this is important because, shadow model based approaches do not always outperform the model-free attacks in all settings. And the current combined method strongly binds to the shadow model based attack and so, the inclusion of the model-free attacks is necessary. \n2. The evaluation of existing defense for the primary attacks should also be included. Some results such as showing that combined attacks can make the defense cost significantly higher (e.g., DP based defenses have to sacrifice more utility to empirically resist the attack effectiveness).\n3. The related work section also misses some relevant works [1], [2], which both consider combining training and test-time types of attacks, given that the authors believe Wen et al.'s work as relevant.\n\n[1] Feng & Tramer, \"Privacy Backdoors: Stealing Data with Corrupted Pretrained Models\", ICML 2024. \n\n[2] Tian et al., \"manipulating transfer learning for property inference\", CVPR 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": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "This paper violates plagiarism policies. According to iThenticate/Turnitin results, the authors directly copied the following content from a previous publication [1]: (1) Section 5.3, covering Attribute Inference and Membership Inference, and (2) Sections A.2 and A.3. The plagiarized material spans approximately one and a half pages, which is a substantial infringement of ethical guidelines.\n\n[1] Liu, Yugeng, Rui Wen, Xinlei He, Ahmed Salem, Zhikun Zhang, Michael Backes, Emiliano De Cristofaro, Mario Fritz, and Yang Zhang. \"{ML-Doctor}: Holistic risk assessment of inference attacks against machine learning models.\" In 31st USENIX Security Symposium (USENIX Security 22), pp. 4525-4542. 2022." }, "flag_for_ethics_review": { "value": [ "Yes, Research integrity issues (e.g., plagiarism, dual submission)" ] }, "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 my comments for more details." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The attack vector is interesting and easy to understand\n2. The attack setting is common in real-world applications\n3. The results of the individual/combo attack seem promising" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores \"attack combos\" where multiple ML attacks are combined, enhancing overall impact. Traditional studies examine attacks individually, but adversaries often employ multiple strategies at once. Focusing on four attacks—adversarial examples, attribute inference, membership inference, and property inference—the authors introduce a taxonomy for attack interactions across three stages: preparation, execution, and evaluation. They identify four effective combos, demonstrating their effectiveness across various ML models and datasets. A toolkit, ACE, is also developed to support research in this area." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper violates plagiarism policies. According to iThenticate/Turnitin results, the authors directly copied the following content from a previous publication [1]: (1) Section 5.3, covering Attribute Inference and Membership Inference, and (2) Sections A.2 and A.3. The plagiarized material spans approximately one and a half pages, which is a substantial infringement of ethical guidelines.\n\n2. The novelty of this paper appears limited. The authors primarily evaluate four standard attacks and their combinations on three basic datasets. Some key advanced attacks, such as model stealing and backdoor attacks, are equally critical in this domain and also need to be thoroughly evaluated and discussed.\n\n3. The lack of critical details makes it difficult to understand what the authors did or what motivated their choices. For instance, while the paper claims that the combo attack is more effective than individual attacks, this is not clearly explained in the methodology and evaluation sections.\n\n4. The paper does not clarify whether the proposed ACE tool can be used to evaluate the impact of inference attacks on commercial ML models, such as those provided through Machine Learning as a Service (MLaaS). \n\n5. Writing needs to be improved. Please proofread the whole paper carefully to correct typos and grammar errors.\n\n[1] Liu, Yugeng, Rui Wen, Xinlei He, Ahmed Salem, Zhikun Zhang, Michael Backes, Emiliano De Cristofaro, Mario Fritz, and Yang Zhang. \"{ML-Doctor}: Holistic risk assessment of inference attacks against machine learning models.\" In 31st USENIX Security Symposium (USENIX Security 22), pp. 4525-4542. 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": 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 above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- To the best of my knowledge, this is the first work to investigate attack combinations across different categories. This is important to develop more secure and reliable models.\n- The released toolkit is beneficial to further exploring the threat of other kinds of attack combinations.\n- The experimental results show that certain combinations of attacks perform better than existing attacks and are insightful for developing stronger attacks and more robust models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This is the first study that investigates the combination of different types of test-time adversarial attacks, including adversarial examples, attribute inference, membership inference, and property inference. The authors decompose the attack pipeline into three phases, i.e., preparation, execution, and evaluation. Through experiments, the authors identify four effective combinations of attacks, such as property inference assisting attribute inference in the preparation phase and adversarial examples assisting property inference in the execution phase. The authors also develop a modular and reusable toolkit that helps investigate effective attack combinations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The architecture of the attacked model is limited. To support the generality of the findings, it would be helpful to attack Transformer-based models, which are known to behave differently than CNNs against adversarial attacks.\n- The experiments are conducted on relatively small datasets. It would be helpful to investigate the behavior of attack combo on large-scale 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": 5 }, "contribution": { "value": 3 }, "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. In ADV2MemInf, the authors chose to constrain the amount of perturbation using the $l_2$ norm. What is the reason behind this design choice, given that both Square and PGD are applicable with different $l_p$ norms? A broader question is: What kind of criterion are the authors using when choosing specific attacks during the experimental evaluation? For instance, using more recent adversarial example generation techniques than PGD might give better results when combined with MemInf.\n2. Other prevalent inference-phase attacks are model extraction (or model stealing) and model inversion. Why have the authors chosen not to include these attacks within the ACE framework? For example, shadow models used in membership inference could potentially improve model extraction, or vice versa.\n3. In Table 2, PropInf2AttInf in CelebA does not improve the F1 score (only 1 pp on average) and the VGG19 model trained on the CIFAR10 dataset. What is the reason behind this while the remaining results show some improvement?\n4. In page 8, lines 419-420, the authors state that the overfitting does not have a significant impact on the membership inference attack. This conclusion is false; as demonstrated by other state-of-the-art works, overfitting, in fact, affects the membership inference attack (Shokri et al., 2017; Liu et al., 2022b)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. ACE is a novel framework that combines different inference-time attacks with the aim to enhance each other's performance. \n2. The rationale for each attack combination aimed at improving the primary attack is well-founded.\n3. The experimental setup is clear enough to replicate experiments if necessary." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors investigate intentional interactions among various inference-phase attacks against ML models and how an adversary can leverage the information obtained from one type of attack to increase the effectiveness of another attack. The proposed framework, ACE, combines various attacks in different phases of attack implementation (preparation, execution, and evaluation phases). Empirical results in three different datasets show that attack performance is increased when they are combined. I think ACE is a novel tool for sophisticated adversaries, but the paper lacks a systematic evaluation (see weaknesses and questions) and might need a revision to improve the quality/discussion." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper focuses on improving the performance of the attacks using various attack combinations. However, the paper lacks an evaluation of how common defense mechanisms (against adversarial examples, membership inference, etc.) affect the effectiveness of such enhanced attacks.\n2. Although the authors empirically show that the effectiveness of the attack increases when they leverage the information from another inference-phase attack, the paper lacks disussion on the efficiency. For example, the implementation of MemInf and PropInf requires the training of numerous shadow models. Although adversarial examples increase the effectiveness of MemInf, they do not reduce the attack's cost and further complicate the attack mechanism, potentially rendering it impractical as an attack strategy. \n3. Although the authors consider both white-box and black-box settings, the population of attack strategies in the paper is not enough to establish a benchmark tool as detailed as, e.g., ML-Doctor or TrojanZoo." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nliu2024ace,\ntitle={{ACE}: Attack Combo Enhancement Against Machine Learning Models},\nauthor={Yugeng Liu and Zheng Li and Hai Huang and Michael Backes and Emiliano De Cristofaro and Yang Zhang},\nyear={2024},\nurl={https://openreview.net/forum?id=3Hg5ufmfRu}\n}" }, "abstract": { "value": "Machine learning (ML) models are proving to be vulnerable to a variety of attacks that allow the adversary to learn sensitive information, cause mispredictions, and more. \nWhile these attacks have been extensively studied, current research predominantly focuses on analyzing each attack type individually.\nIn practice, however, adversaries may employ multiple attack strategies simultaneously rather than relying on a single approach.\nThis prompts a crucial yet underexplored question: when the adversary has multiple attacks at their disposal, are they able to mount or enhance the effect of one attack with another?\nIn this paper, we take the first step in studying the intentional interactions among different attacks, which we define as attack combos. \nSpecifically, we focus on four well-studied attacks during the model's inference phase: adversarial examples, attribute inference, membership inference, and property inference. \nTo facilitate the study of their interactions, we propose a taxonomy based on three stages of the attack pipeline: preparation, execution, and evaluation.\nUsing this taxonomy, we identify four effective attack combos, such as property inference assisting attribute inference at its preparation level and adversarial examples assisting property inference at its execution level. \nWe conduct extensive experiments on the attack combos using three ML model architectures and three benchmark image datasets.\nEmpirical results demonstrate the effectiveness of these four attack combos.\nWe implement and release a modular, reusable toolkit, ACE. \nArguably, our work serves as a call for researchers and practitioners to consider advanced adversarial settings involving multiple attack strategies, aiming to strengthen the security and robustness of 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": { "value": [ "~Yugeng_Liu1", "~Zheng_Li17", "~Hai_Huang4", "~Michael_Backes3", "~Emiliano_De_Cristofaro1", "~Yang_Zhang15" ] }, "authors": { "value": [ "Yugeng Liu", "Zheng Li", "Hai Huang", "Michael Backes", "Emiliano De Cristofaro", "Yang Zhang" ] }, "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": [ "machine learning security and privacy", "membership inference", "attribute inference", "property inference", "adversarial examples" ] }, "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": { "value": "liu|ace_attack_combo_enhancement_against_machine_learning_models" }, "pdf": { "value": "/pdf/fae7730f9756d2d4f27f28d8bd0bbd711f7f7f5b.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": "ACE: Attack Combo Enhancement Against Machine Learning Models" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_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": 1 }, "primary_area": null, "questions": { "value": "Q1)The “pass rate” is an important metric to evaluate the performance in section 4.3. Could you elaborate more on how it is calculated? What are the reasons to choose it? And how is related to training/fine-tuning phases?\n\nQ2)Could you contextualize the size and the quality of the generated dataset in terms of how significant it is in comparison with SOTA or related works? Are they general enough for RPA for just for applications around Apple Shortcuts and RoutineHub?\n\nQ3) The papers also mention about fine-tuning Worfflow Lalama and annotator in a very short paragraph in section. Could you elaborate more about setup, why such training parameters are chosen?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1)The observation and problems are interesting and relevant to the conference\n\nS2) The dataset might have potential uses\n\nS3)The experiments involve many systems" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper argues that state-of-the-art models like GPT-4o face challenges in effectively handling complex workflows. To address this, the paper introduces WorkflowLLM, a data-centric framework designed to enhance the workflow orchestration capabilities of LLMs. Central to this framework is WorkflowBench, a large-scale fine-tuning dataset comprising 106,763 workflows, 1,503 APIs, and 83 applications across 28 categories. WorkflowBench is constructed through a three-phase pipeline: data collection from sources like RoutineHub, query expansion using ChatGPT to create diverse and complex workflows, and workflow generation leveraging a trained annotator model for high-quality workflow synthesis. The dataset is enriched with Python-style code, comments, and hierarchical thought structures to improve LLM learning.\n\nBased on WorkflowBench, the paper fine-tunes Llama-3.1-8B, resulting in WorkflowLlama, which demonstrates superior performance compared to existing models, including GPT-4o, on workflow orchestration tasks. WorkflowLlama achieves notable success in generalization to unseen APIs and tasks, evaluated using metrics such as CodeBLEU and Pass Rate. Additionally, WorkflowBench exhibits robust zero-shot generalization on the T-Eval benchmark, achieving an F1 plan score of 77.5%." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1) The scientific or technical contributions are limited as the key contributions of paper are around data curation effort. And the significance of data size and quality is not clear (see Q2). \n \nW2) Section 3 mentions many steps to generate the needed data for its workflow using ChatGPT but the Quality control protocol is not very clear. Even the author provides some examples and algorithms in appendix, the details are very descriptive . Also section4.2 mentions using human evaluator to re-label the sampled data,It’d be better to provide more details towards the quality control protocol for this human-driven process as well.\n\nW3) Many technical details are not very clear (see questions)\n\nW4) Code is provided in supplemental material but there is no document how to reproduce the experiment results ( also see Q3)" }, "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": "Are the prompts used in the framework that drive the LLMs usage automatically generated? If yes, how to ensure the quality of the prompts? If not, how to scale up?\n\nFor each step from data generation to model evaluation, how to optimize the prompts?\n\nThe description on quality confirmation is a bit vague. The example given about the issues is not clear. How are such issues detected? How to prompt ChatGPT to refine A’ and P’, and how to ensure the quality of the refinement? How to perform the rule-based filtering, e.g., how to automatically detect if parameter constraints are violated or not?\n\nFor OOD settings, how are the other models evaluated? Do they also be given the same input like the one for Workflow Llamma, i.e., providing the list of required APIs? \n\nGiven synthetic data takes the large portion of the final benchmark (91k out of 11k) but the quality is unsure, it would be interesting to also see how Workflow Llama performs on the real-word dataset, i.e., without the synthetic data, compared to other models? \n\nThe proposed framework outperforms all the other models by an extremely large margin in the first four columns in Table 2 (such as 9.4% vs 0.5). This seems to be suspicious and needs further examination and explanation. \n\nIn Table 2, both CodeBLEU and Pass Rate are used, one examines the similarity of code and one examines if the code is executable. It would be interesting to combine both of them, such as setting a threshold on the CodeBLEU score when checking the Pass Rate to compare models, since Pass Rate itself doesn’t necessarily indicate the success of the workflow orchestration." }, "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-written and easy to follow. Examples are given to illustrate the important steps. \n\nThe extended dataset improves the diversity and complexity of the samples in the training data, allowing the models to adapt to the situations that are closer to real-world applications. \n\nThe experimental results show that the proposed framework outperformed the existing models by a large margin." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework that leverages LLMs to automate workflow orchestration. It extends the real world dataset by creating synthetic data through LLMs, to train the model. It conducts experiment study to compare the proposed framework with several proprietary and open source models, using CodeBLEU and Pass Rate as the metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper mainly relies on prompting LLMs to generate the results of the key steps, such as data generation, commenting actions, and plan generating. It seems that there is a large space of prompt engineering to improve the performance but left to be undone. \n\nThe proposed framework bypasses the API selection when handling the OOD setting by directly serving the required APIs as the input. This greatly simplifies the problem of automating workflow orchestration. \n\nThe paper relies heavily on LLMs for the entire process, including generating the dataset, training the model, and evaluating the models (only less than 82% accuracy in a small sample with 330 instances in total). This weakens the technical contribution and the reliability of the experiment 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": { "value": "Potential ethical considerations were taken into account in the Appendices." }, "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 am curious about the choice of ChatGPT to run both the query expansion phase of the dataset creation (section 3.2) and the experiments (sections 4.1 and 4.2). Altough I think it was a good idea, why rely on a proprietary model (the GPT-4o) with proprietary and non-disclosed prompt engineering behind it? Did the authors try to use directly the GPT-4o model? Did they try other models to run this phase? I assume that the \"ease-of-use\" of the ChatGPT is a fair point in favor of its use, but it should be evident in the paper." }, "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 well-presented and well-written overall. The authors identified a clear pain point in achieving an agentic LLM approach for process automation: lacking LLMs capable of correctly describing complex workflows. Although partially addressed by other works, as pointed out by the authors in the related work section (section 2), their approach is innovative in dealing with fairly more complex workflows than previous solutions.\n\nTheir approach to overcoming this limitation is sound and clever. They provide a fine-tuned LLM specialized to handle this kind of task.\nTo do so, they craft a well-defined dataset, and it is possible to highlight the process of dataset creation as one of the paper's main contributions, as crucial as the constructed dataset and the fine-tuned model.\nAfter the data collection, the query expansion phase is especially interesting because it uses another LLM (e.g., ChatGPT) to help overcome the lack of diversity and complexity of the gathered initial data with syntactic data. The same applies to the clever use of ChatGPT in the evaluation phase.\nCompared with other LLMs solving the same kind of problem, the presented results show the potential use of the author's solution to deal with the initial established problem, i.e., the automation of complex tasks.\nWorkflowLlama outperforms other LLMs when the workflow demands generalization to unseen instructions and unknown APIs.\n\nI also highlight the good approach to evaluating the effectiveness of the LLM-based evaluator (section 4.2), which strengthens the paper's arguments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a detailed explanation of the construction and evaluation of the WorkflowBench dataset, which contains many examples of workflow descriptions. The dataset is later used to fine-tune the open-source LLM Llama-3.1-8B.\nThe dataset creation follows a well-established and potentially reproducible approach. It starts with data gathering, expands to increase data diversity, and generates a final workflow dataset, enhancing real collected data with synthetic data.\nThe fine-tuning intends to overcome the limitations of existing LLMs in workflow orchestration and provide a consistent improvement in process automation by proposing an agentic approach.\nThe fine-tuned LLM called WorkflowLllama is detailed, and its capabilities are evaluated, showing a solid capacity to orchestrate complex workflows. Its performance was compared with commercial and open-source LLMs using the test set of Workflowbench." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The background on automation's relevance (lines 36-38) might be condensed, as automation's importance is widely recognized.\n- In section 3.1, the author can better explain the $\\mathcal{D}$. The other elements were presented before, but  $\\mathcal{D}$ is directly introduced on line 248 without prior explanation.\n- Lines 245-247 read (my emphasis):  \"This bottom-up [...], efectively **ensuring content realiability** and **minimizing the risk of hallucination**\". These are pretty strong assumptions without solid evidence, especially if generalized to every LLM. Do the authors know published papers that confirm them (since this paper is not about reducing hallucinations)? If not, the phrase can be removed or rephrased to be more humble.\n- The paper lacks a clearer section on threats to validity. The authors provided some in Appendix E, but they should be incorporated into the main text.\n- The paper also lacks a more evident running example/use case. Although the authors provide a small case study in section 4.7 and some examples in Appendix D, a consistent running example, incorporated throughout sections, could clarify the model's applications and emphasize practical use cases.\n\nMinor issues:\n- Line 52 mentions GPT-4, while the rest of the paper uses GPT-4o(-mini). Maybe it is only a typo, or the work mentioned indeed uses the GPT-4 model. Since they are different models, it can be good to double-check every reference to ensure it always talks about the correct model, maybe highlighting somewhere they are not the same.\n- The size of Figure 2 can be increased to improve readability.\n- Figures 3,5 and 6 could improve accessibility through higher contrast colors.\n- Although well-written, the paper is sometimes repetitive (e.g., sections 1 and 3). Proofreading with this issue in mind may help the authors achieve a better final text." }, "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 performance of generic LLMs in workflow orchestration through agent-like approaches would serve as an interesting baseline for comparison. Through this comparison, we could better understand whether the core challenge lies in the model's reasoning capabilities or in its ability to maintain consistent output formats. This point won't affect the score; it's just for discussion." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Accurate and valuable problem identification. The paper identifies the limitations of existing work in handling real-world workflows: difficulty in processing long-sequence workflows with complex logical relationships like control flow. This observation directly addresses practical application pain points, making the motivation very solid.\n2. Proposes a systematic data construction method. The paper designs a complete data synthesis process: from real data collection, workflow abstract representation, data synthesis to quality control, with detailed design considerations at each stage. The resulting large-scale dataset provides an important resource for this field.\n3. Comprehensive experimental validation. The effectiveness of the method is strongly supported through multi-dimensional evaluation using CodeBLEU and Pass Rate, cross-domain testing with T-Eval, along with detailed case analysis and ablation studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a pipeline for constructing workflow (RPA) data to enhance LLMs' workflow orchestration capabilities. The core motivation stems from the observation that real-world workflows require more complex logic and longer sequences than what current LLMs can directly generate. The primary contribution is the construction of a dataset containing over 100K samples. The main limitations lie in the lack of certain technical details (such as specific ChatGPT versions used for annotation) and the lack of deeper investigation into model capability improvements." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Important technical details are missing. The paper mentions using ChatGPT for data annotation and optimization multiple times but doesn't specify the model version. For work with data construction as its core contribution, this seriously affects reproducibility and rigor.\n2. Lacks methodological innovation and mechanism analysis. Although data construction is an important contribution, the paper lacks in-depth analysis of how this data enhances model capabilities. Specifically, it doesn't investigate whether the improvements come from enhanced logical reasoning abilities or simply better format matching. Without such analysis, it's difficult to determine if the model has truly learned to reason about complex workflows or has merely memorized patterns from the training data.\n3. Missing critical ablation experiments. The paper doesn't compare the performance difference between the Annotator Model and WorkflowLLM, despite both models using the same training method and data type. This results in question the necessity of the data synthesis strategy and weakens the core argument.\n4. The paper mentions that the Annotator Model can generate workflows with over 70 actions, and theoretically WorkflowLLM should be capable of the same. Given the general limitations of LLMs in long sequence generation, including such long-sequence workflows in the appendix would further demonstrate the work's contribution.\n5. Given that the core contribution is data construction, whether the dataset will be open-sourced directly affects the practical value of the work. The authors should consider open-sourcing the dataset to promote development in this field." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024workflowllm,\ntitle={Workflow{LLM}: Enhancing Workflow Orchestration Capability of Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Hy00Wvabi},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in large language models (LLMs) have driven a revolutionary paradigm shift in process automation from Robotic Process Automation to Agentic Process Automation by automating the workflow orchestration procedure based on LLMs. However, existing LLMs (even the advanced OpenAI GPT-4o) are confined to achieving satisfactory capability in workflow orchestration. To address this limitation, we present WorkflowLLM, a data-centric framework elaborately designed to enhance the capability of LLMs in workflow orchestration. It first constructs a large-scale fine-tuning dataset WorkflowBench with 106,763 samples, covering 1,503 APIs from 83 applications across 28 categories. Specifically, the construction process can be divided into three phases: (1) Data Collection: we collect real-world workflow data from Apple Shortcuts and RoutineHub, transcribing them into Python-style code. We further equip them with generated hierarchical thought via ChatGPT. (2) Query Expansion: we prompt ChatGPT to generate more task queries to enrich the diversity and complexity of workflows. (3) Workflow Generation: we leverage an annotator model trained on collected data to generate workflows for synthesized queries. Finally, we merge the synthetic samples that pass quality confirmation with the collected samples to obtain the WorkflowBench. Based on WorkflowBench, we fine-tune Llama-3.1-8B to obtain WorkflowLlama. Our experiments show that WorkflowLlama demonstrates a strong capacity to orchestrate complex workflows, while also achieving notable generalization performance on previously unseen APIs. Additionally, WorkflowBench exhibits robust zero-shot generalization capabilities on an out-of-distribution task planning dataset, T-Eval." }, "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", "Process Automation", "Workflow", "Tool 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/61b261072d2e135d1cbe1c150891ef1b3f631413.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/1d0fe4cb519a7fc7de0de5e56b9068fb29b1e1a4.zip" }, "title": { "value": "WorkflowLLM: Enhancing Workflow Orchestration Capability of 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": 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": "Clerical: In section 5.2 the $l_{1}$ option setting is mentioned as a baseline, but Table 1 compared the options to something called $l_{0}$. Do $l_{0}$ and $l_{1}$ refer to the same baseline? If so, using consistent notation will help make the results section more readable." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The need for efficiency in decision making in RL is clear, as single-step actions are slow and computationally expensive (even more so in slow simulators). Thus, the problem addressed by OptionZero is clear and its existence is well-motivated. Additionally, since much prior work in options appear to be in manually defined and demonstration-based settings, the generalisability of OptionZero is a strong selling point.\n\nWithin fixed computational constraints, the idea of decreasing the frequency of queries to the network is a strong idea for the current state of RL. Also important is the notion of learning subroutines which the options network will identify as useful in different scenarios and not have to re-learn temporal relationships.\n\nThe flexibility to play options or primitive actions results in tailored reactions to scenarios, as an agent may need the fine-grained approach taken by traditional RL. The main results in Table 1 indicate the validity of the method, as using options provides a performance benefit more often than not, with longer option limits sometimes outperforming shorter ones." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work proposes a revamped approach to the well-known options framework, which allows agents to take temporally extended actions as well as myopic ones. By combining a network which learns options with MCTS MuZero (which models transition dynamics) the authors propose a method to utilise options alongside single actions in self-play games." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It is unclear why options are outperformed by primitive actions in certain environments. The authors suggest that in environments with high combinatorial complexity, learning of the dynamics model may be difficult and thus options may simply produce more overhead than actual benefit. A more detailed analysis of these environments would be beneficial, for e.g. investigate whether there is a correlation between the stochastic branching factor of the environment and the performance of options.\n\nAdditionally, it seems that longer options may improve efficiency but not always increase performance when those options may be overextending in environments where more granular control is required. Have the authors considered implementing dynamic options lengths somehow? This may make the idea more viable, or at least a discussion on the complications of implementing that would be a good addition to the 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": 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. Why select these 26 Atari games instead of using the standard 57 Atari games?\n1. What is the hardware resource for conducting the experiments?\n1. How long does training a single Atari game with OptionZero take?\n1. How long does training a single Atari game with MuZero take?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well-written. The authors explain the use of options clearly with a toy example, demonstrating how options are used. The empirical results are also strong, achieving high mean normalized scores." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents OptionZero, a novel approach that incorporates an option network into MuZero and allows the agent to learn temporary extended actions. The authors conducted empirical experiments and find OptionZero outperforms MuZero in teams of mean human-normalized scores on 26 Atari games." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It's not clear the actual benefits options bring. In the intro, the paper claims options allow for \"searching deeper\", but the empirical analysis shows \"deeper search is likely but not necessary for improving performance\". While it's nice to have the option to do option, could the authors provide a more detailed analysis of options beyond a deeper search? \n\nThe paper could also benefit more from discussions of \n1) the trade-offs between increased complexity and performance gains\n2) how much tuning did the authors perform to make OptionZero work; were there failure cases/ideas during the development and how did the authors overcome 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": 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": "- How does the dynamics network handle complex action spaces, especially in games with highly varied option paths?\"\n\n- What are the specific computational costs of incorporating the option network, both in training and during MCTS simulations? Could the author discuss the associated overhead with the proposed method?\n\n- Can the model be applied to environments beyond games with less predictable state transitions, and how would option discovery be affected? I would suggest adding studies in robotic environments." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Novel idea for autonomous and adaptable option discovery. The proposed method's ability to autonomously discover and tailor options to diverse game dynamics removes the need for predefined actions, making it highly adaptable across different environments.\n\n- Convincing results for enhanced planning in RL. By integrating an option network, OptionZero reduces decision frequency, enabling computational efficiency, particularly in visually complex tasks like Atari games.\n\n- Strong Performance Gains: On Atari benchmarks, OptionZero achieves a 131.58% improvement over MuZero, a significant improvement compared to previous SOTA papers.\n\n- Interesting ideas to adjust option lengths, balancing performance and training efficiency, particularly useful in tasks needing variable action sequences." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce OptionZero, an advanced extension of the MuZero algorithm that autonomously identifies temporally extended actions, known as options, without the need for predefined options or expert-provided demonstrations. By incorporating an option network, OptionZero enhances planning efficiency by decreasing both decision-making frequency and the computational load required for complex environments. Evaluated on a suite of Atari games, OptionZero demonstrates notable improvements in human-normalized scores compared to MuZero, accompanied by a detailed analysis of how options are utilized across varying game states and scenarios" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Inconsistent Option Use Across Games: OptionZero's reliance on options appears to vary widely across Atari games. While longer options bring substantial gains in some games, they contribute less in others. This inconsistency suggests that the model’s option-based planning may struggle to generalize well across diverse, complex environments. The paper should discuss this limitation. \n\n- Challenges in Complex Action Spaces: In games with intricate action spaces, such as Bank Heist (Atari), OptionZero’s dynamic network encounters difficulty as option lengths increase, particularly with multi-step dependencies. This issue may restrict OptionZero’s application in environments where actions are highly combinatorial, relying instead on settings with more straightforward or predictable actions.\n\n- Reduced Prediction Accuracy for Longer Options: The model’s prediction accuracy tends to decrease as options become longer, affecting planning quality where extended strategies are essential. I would recommend adding an experiment to study this effect and discuss potential limitations of the proposed method. \n\n\n- Limited Application Beyond Games: Although the model shows promise in game environments, the paper does not investigate its potential beyond Atari-like settings. I would appreciate seeing results on other domains, maybe robotic such as Gymnasium-Robotics. Under the current evaluation, the method seems limited to game-based 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": 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": "Please refer to the weakness part." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. This paper is well-written with a clear structure, making the research content easily comprehensible.\n\n2. By introducing the OptionZero framework and leveraging self-play for automatic option design, this study paves the way for new approaches. The novelty is good.\n\n3. The experimental results robustly support the effectiveness of the algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the OptionZero framework, which incorporates options into MCTS and enables the automatic design of options through self-play, thereby avoiding cumbersome manual design. In Atari games, OptionZero achieves significant improvements compared to MuZero, achieving a 131.58% improvement in mean human-normalized score. It has shown promising directions for discovering and using options in planning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I am curious about whether the introduction of options has any impact on the theoretical optimality of MCTS.\n\n2. Is it possible to demonstrate the advancement of the algorithm more directly by solely utilizing the learned network for action selection, without relying on MCTS?\n\n3. What is the expected performance as the value of $l$ increases?\n\n4. What are the differences in the running wall-clock time required for OptionZero compared to MuZero?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper presents OptionZero, a method that integrates options into the MuZero algorithm, which autonomously discovers options through self-play games and utilizes options during planning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024optionzero,\ntitle={OptionZero: Planning with Learned Options},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3IFRygQKGL},\nnote={under review}\n}" }, "abstract": { "value": "Planning with options -- a sequence of primitive actions -- has been shown effective in reinforcement learning within complex environments. Previous studies have focused on planning with predefined options or learned options through expert demonstration data. Inspired by MuZero, which learns superhuman heuristics without any human knowledge, we propose a novel approach, named OptionZero. OptionZero incorporates an option network into MuZero, providing autonomous discovery of options through self-play games. Furthermore, we modify the dynamics network in MuZero to provide environment transitions when using options, allowing searching deeper under the same simulation constraints. Empirical experiments conducted in 26 Atari games demonstrate that OptionZero outperforms MuZero, achieving a 131.58% improvement in mean human-normalized score. Our behavior analysis shows that OptionZero not only learns options but also acquires strategic skills tailored to different game characteristics. Our findings show promising directions for discovering and using options in planning." }, "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": [ "Option", "Semi-MDP", "MuZero", "MCTS", "Planning", "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/2c03adaa2af6d28cea5615c047befbd4055e5098.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/13604cd8e5c1d54b2c3b5bc6b872907fd07b5235.zip" }, "title": { "value": "OptionZero: Planning with Learned Options" }, "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": "As I liked a lot the paper, here I also include a series of suggestions to further improve the paper:\n\n- At page 4, shouldn't the sup-norm be defined as $||f||_\\inf = sup_{x \\in \\mathcal{X}} |f(x)|$?\n- At page 5, I think it would have been great to have a small example with just a neural network with two outputs $y_0$ and $y_1$ and the constraint $y_0 \\ge y_1$. Then you could for example show that if $y_0 = 3$ and $y_1 = 4$ then $a(x)=[−1,1]$, $b(x) =0$ and \n$$\n\\mathcal{P}(f_\\theta)(x) = f_\\theta(x) - \\frac{a(x)}{\\||a(x)\\||^2} \\text{ReLU}(a(x)^\\top f_\\theta(x) - b(x)) = [3.5, 3.5].\n$$\n- At page 5, among the assumptions there is written that the constraints need to be feasible. Just to improve the readability of the paper and also make sure everything is well defined, it would help to add the meaning of the word feasible (i.e., \"that there exists at least one solution or point within the domain of interest that satisfies all the constraints simultaneously\")\n- At page 5 the authors give the assumptions for which the number of constraints needs to be lower or equal than $n_{out}$. I think it would be really helpful to add a simple example with a set of constraints that cannot be captured (e.g., $x \\ge 0, y \\ge 0, x+y \\ge 0$)\n- At page 8, in the experimental analysis, the constraints you show clearly define a non-convex space. However, for your layer to work you need to have a set of constraints that defines a convex space. Are you simply applying different projections on the ground of the value of $x$? If that is the case, I personally find this experiment a bit misleading as this only works because $x$ is a known input. I do not think your layer would work in a setting where you have a constraint of the type if $y_1 > y_2$ then $y_2+ y_3 < 1$, tight?\n- Finally, I think it would also be nice if you could extend on how this type of work is relevant for the SafeAI literature, as creating models that are complaint by design with a set of constraints obviously increases their safety." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "In general, I like this paper quite a lot and I think it has different strengths (listed below):\n\n- The paper handles a very important problem. \n- The paper is technically sound \n- The paper is written very well" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposes a new layer able to project the output of the neural network (which might be non-compliant with a set of hard constraints) back into a \"safe space\" where the constraints are guaranteed to be satisfied." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper has only one major flow: it is not well placed in the literature. \n\nIndeed, I do not think the authors are familiar with the Neuro-symbolic AI literature, where the problem of learning with hard constraints has already been studied. In particular, there is a research group that has worked a lot on creating layers that are make neural networks compliant by construction with a set of hard constraints [1,2,3]. [1] is the first work that proposed this kind of approach with constraints expressing hierarchies over the outputs. In the latest works they worked with hard constraints expressed in propositional logic [2] and as linear inequalities [3]. Obviously I believe [3] is particularly relevant to your paper and it would be nice to have a comparison between the two methods (at least in terms of discussion for the rebuttal phase and experimental only for the camera ready). Delving more on the logical side you have works like Semantic Probabilistic Layer that gives a probabilistic perspective to hard constraints expressed in propositional logic and can guarantee their satisfaction by construction [4]. Finally, you can find an entire line of work which maps the outputs of the neural network into logical predicates and allows reasoning on top of these predicates (see e.g., [5,6,7]) which then also guarantees the satisfaction of the constraint. \n\nThe final rate is below the acceptance threshold because of this. However, I am fully aware that it is often very hard to keep up with the extensive literature available in ML, so I will be very open to increasing my score. \n\nReferences:\n\n[1] Eleonora Giunchiglia and Thomas Lukasiewicz. Coherent hierarchical multi- label classification networks. In Proc. of NeurIPS, 2020.\n\n[2] Eleonora Giunchiglia, Alex Tatomir, Mihaela Catalina Stoian, and Thomas Lukasiewicz. CCN+: A neuro-symbolic framework for deep learning with requirements. International Journal of Approximate Reasoning, 171, 2024.\n\n[3] Mihaela C. Stoian, Salijona Dyrmishi, Maxime Cordy, Thomas Lukasiewicz, and Eleonora Giunchiglia. How Realistic Is Your Synthetic Data? Constraining Deep Generative Models for Tabular Data. In Proceedings of International Conference on Learning Representations, 2024.\n\n[4] Kareem Ahmed, Stefano Teso, Kai-Wei Chang, Guy Van den Broeck, and Antonio Vergari. Semantic probabilistic layers for neuro-symbolic learning. In Proceedings of Neural Information Processing Systems, 2022.\n\n[5] Robin Manhaeve, Sebastijan Dumancic, Angelika Kimmig, Thomas Demeester, and Luc De Raedt. DeepProbLog: Neural probabilistic logic programming. In Proceedings of Neural Information Processing Systems, 2018.\n\n[6] Connor Pryor, Charles Dickens, Eriq Augustine, Alon Albalak, William Yang Wang, and Lise Getoor. Neupsl: Neural probabilistic soft logic. In Proceedings of International Joint Conference on Artificial Intelligence, 2023.\n\n[7] Emile van Krieken, Thiviyan Thanapalasingam, Jakub M. Tomczak, Frank Van Harmelen, and Annette Ten Teije. A-neSI: A scalable approximate method for probabilistic neurosymbolic inference. In Proceedings of Neural Information Processing Systems, 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": 3 }, "primary_area": null, "questions": { "value": "- I am a little confused about part iii) of Proposition 4.2. Namely, the projection preserves the distance from the boundary of the feasible set when $\\bar{f}_\\theta (x)$ satisfies the constraint. Would you mind sharing a geometric intuition?\n- I am also confused about the $C_\\leq (f (x))$ notation in line 359. What does $C$ denote? This is different from the $C (x)$ in Eq.(4), right?\n- Regarding Figure 2, it looks like all models perform reasonably good in the region which the training data lie in, and the difference occurs outside of data coverage. I am confused why \"Soft\" seems much worse than others. If I understood it correctly, \"Soft\" penalizes when the model output violates the constraints. Since all training points are feasible, I intuitively expect \"Soft\" to behave similarly to \"NN\", but this is not the case. Could you please explain why such difference? Also, how would \"Soft + proj\" look like?\n- For the \"safe control policy\" experiment in Section 5.3, what do you think is the biggest advantage of the proposed method compared with non-learning methods such as model predictive control?\n- Line 500 mentions that the constraint in Eq.(12) can be conservative, leading to worse performance compared to \"Soft\" and \"DC3\". Is it possible to adjust the level of conservativeness by changing $\\alpha$?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I have not worked on constrained neural networks, and hence I am unfamiliar with a lot of the cited literature. That being said, judged based on the content of this submission, the results are promising and meaningful, and the presentation is mostly clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a type of hard-constrained neural network by introducing differentiable projection layers. Specifically, if the constraints are affine and the number of constraints are no greater than the output dimension, the projection can be found in closed form. For other convex constraints, the authors propose to apply the differentiable optimization framework to compute the projection iteratively. The authors use experiments including learning an optimization solver and controlling a safety-critical dynamical system to demonstrate the effectiveness of the proposed work." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- To use the closed-form projection algorithm Eq.(7), we need $n_{ineq} + n_{eq}$ to be no greater than the output dimension. Is this restrictive in practice? In the included experiments, which one of them uses closed-form projection?\n- For Eq.(11), should $u$ also be a function of $t$, i.e., $u (t)$?\n- In Table 4, why are rows 2 and 4 marked as red even though they are feasible?" }, "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": "- Could the authors train DC3 for longer, or with more inner iterations to satisfy the inequality constraints? If this is deemed infeasible, can the authors provide an explanation on the discrepancy with the results in the original paper?\n- Would it be possible to provide \"DC3 + Proj\" results?\n- Why is DC3 absent from Table 5?\n- In the toy example, training points are sampled from [-1.2, 1.2], but then the networks are evaluated on [-2, 2]. Aren't samples in that area OOD, in a sense? Couldn't that explain the performance of the baselines? I understand that guaranteed constraint satisfaction is an advantage of the proposed approach, but these points should be discussed. (e.g., by providing results on [-2, 2] training)\n- What is lost by the fact that HardNet-Aff does not rely on an orthogonal projection? Does this imply anything concerning the hardness of learning the function through gradient-based method? An interesting ablation would be to compare HardNet-Aff with HardNet-Cvx on a setup where both are supported.\n- It would be interesting to see some experiments on (even slightly) larger networks. Would some methods benefit more from the additional capacity than the others?\n\nIn general, I think the quality of the work would clearly increase if the authors were more honest on the limitations of the proposed approach (see weaknesses above)." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The idea to enforce hard constraints by construction through a projection layer is simple and neat. \nDifferently from previous work in the area, universal approximation theorems are provided.\nThe experiments show that, at least for affine constraints supported by HardNet-Aff, HardNet works quite well in practice (albeit at a small scale).\nFinally, I found the related work section to be well-written and fairly comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents HardNet, an approach to train neural networks that satisfy hard constraints by construction.\nThe core idea of the paper is to append a projection layer at the end of the network in order to bring the network output onto the feasible set.\nTwo different schemes are presented: one using a closed-form (non-orthogonal) projection for affine constraints, and one resorting to previous work presenting differentiable convex optimization solvers, in case of more general convex constraints.\nUniversal approximation theorems for the architectures are presented.\nExperimental results on a variety of benchmarks are presented, demonstrating that HardNet attains good performance while satisfying the constraints." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main weaknesses of the paper are threefold: HardNet-Cvx, the assumptions behind HardNet-Aff, and the experimental section.\n\n*HardNet-Cvx*: the idea to use differentiable optimizers to perform the projection does not appear to be completely novel. DC3 discusses it in related work, excluding it because of large computational cost (this large cost is definitely confirmed at inference in the experiments in Table 5). The rayen paper uses it as a baseline (named PP in their paper). I do not know if the authors are aware of this, but these points absolutely need to be acknowledged throughout the paper. Furthermore, the only example over HardNet-Cvx is used (Table 5) appears to nevertheless use affine constraints (albeit, as far as I understand, too many to be supported by HardNet-Aff). In this instance, its runtime is extremely large, questioning its practical applicability.\n\n*HardNet-Aff assumptions*: the assumptions required for HardNet-Aff seem very strong to me. It seems to be that a simple interval constraint per network output coordinate would already be unsupported, hence incurring the large cost associated to HardNet-Cvx. Could the authors comment on this?\n\n*Experiments*: my main concern over the experimental section is the surprisingly bad performance of DC3. In the original paper, all constraints appear to be satisfied in practice. Is there anything I am missing here? Was DC3 run for an insufficient number of iterations? I understand that for HardNet the constraints hold by construction, but DC3 appears to be fairly strong empirically, in the original paper. Important details such as training times for each scheme appear to be omitted (or at least, do not feature prominently). \"DC3 + Proj\" would also appear to be a missing, yet very interesting baseline. Further details are provided as 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": 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": "I'd like to explain more in detail my doubts.\nConsider the simple case of $1-d$ output and a single affine constraint. The projection layer reduces a simple rescaled ReLU, and the whole network has zero gradient where the constraint is not satisfied. \n\nThis effect is true in general. In fact, if we evaluate the Jacobian of the projection layer when the constraint is not satisfied ($J_{\\mathcal{P}} = I-a(x)a(x)^T$), we can see that the gradient of the network will be always orthogonal to the constraint vector $a(x)$. \n\nFor this reason, if $f_\\theta$ is initialized outside the feasible region, it should be impossible to \"re enter\" it by simply following the gradient. This means that the whole optimization would get \"stuck\" on the boundary of the feasible set, which might not be ideal. In stochastic gradient descent, this issue might be mitigated, however, i believe this is an important discussion to have in the paper.\n\nThe proposed projection (for the affine variant) works in two steps, reducing the dimensionality of the output space using the equality constraints, and performing a projection in the reduced space. Is this better than simply treating equalities as a pair of inequalities? This aspect should be investigated to justify the additional complexity of the method." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is very clear in its presentation. It is well structured and reads well. The contributions of the paper are clearly presented and well summarized, both in text and by images and tables. The empirical evaluation include plenty of relevant and diverse scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a simple framework for imposing constraints on input-output relations in neural networks.\nThe approach consists in appending a final projection layer to the network, ensuring that the constraints are satisfied by construction. Moreover, the authors show (formally) that this projection operation does not hinder the expressivity of the network, and empirically evaluate the approach on various scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have some doubts regarding the novelty of the paper and the technical discussion. \n\nThe idea of satisfying hard constraints using a final differentiable projection layer is mentioned in the works cited as reference, and many other works referencing the idea exist (i.e. https://arxiv.org/abs/2111.10785 and https://arxiv.org/abs/2307.10459). If the contribution is simply the extension of the idea to input dependent constraints, it should be more clearly stated.\n\nAs for the soundness of the approach, while the additional projection layer is differentiable, its derivative is not well behaved.\nClaims such as \"meeting the required constraints while allowing its output to be backpropagated through to train the model via gradient-based algorithms\" and \"This allows us to project the output $f_θ(x)$ onto the feasible set $C(x)$ and train the projected function via conventional gradient-based algorithms\" are not substantiated by a proper discussion on the gradient properties of the resulting network.\nIn fact, as presented, the gradient is always orthogonal to the constraint. \nThis observation is not novel, and from my understanding is the main motivation driving the development of alternatives to projection methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024hardconstrained,\ntitle={Hard-Constrained Neural Networks with Universal Approximation Theorem},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Imf21Jvwh},\nnote={under review}\n}" }, "abstract": { "value": "Incorporating prior knowledge of input-output relationships into machine learning models has gained significant attention, as it enhances generalization from limited data and ensures trustworthy predictions. However, most existing approaches use soft constraints by penalizing violations through regularization, which offers no guarantee of constraint satisfaction---a critical requirement in safety-critical applications. On the other hand, imposing hard constraints on neural networks may hinder their representational power, adversely affecting performance. To address this, we propose HardNet, a practical framework for constructing neural networks that inherently satisfy hard constraints without sacrificing model capacity. Specifically, we encode affine and convex hard constraints, dependent on both inputs and outputs, by appending a differentiable projection layer to the network’s output. This architecture allows unconstrained optimization of the network parameters using standard algorithms while ensuring constraint satisfaction by construction. Furthermore, we show that HardNet retains the universal approximation capabilities of neural networks. We demonstrate the versatility and effectiveness of HardNet across various applications: fitting functions under constraints, learning optimization solvers, optimizing control policies in safety-critical systems, and learning safe decision logic for aircraft 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": [ "constrained optimization", "universal approximation", "surrogate 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/d0ad082c6f735a75a6663071293d2c5603bb8de6.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": "Hard-Constrained Neural Networks with Universal Approximation Theorem" }, "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 will use this questions paragraph for questions/suggestions.\n\nRegarding the dependency on accurate camera poses:\n- How can the proposed method be integrated with SLAM or other online techniques?\n- How does the method work with imperfect poses? An experiment on this would be beneficial.\n\nRegarding the weak experiments: \n- Perform experiments on other datasets. The authors can find many in the cited papers.\n- Compare to NICER-SLAM.\n- Compare to the vast literature of photometric stereo with known poses; few examples [m1-m4].\n\nAll in all, I don't feel it is realistic for the authors to fix all my issues within the discussion period." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I don't find any obvious strengths in the paper. The method has many unjustified steps that completely ignore prior work. The experimental section is very weak, with the baselines outperforming the proposed method in many metrics, with the proposed method being quite slow even though it was sold as an online method, and with having results only on a single dataset." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method for estimating the 3D structure from an image sequence given known camera poses. The method incrementally adds images into the reconstruction, further optimizing the visible parts. Each image is encoded through a transformer-based encoder and feature pyramid. Then, monodepth prediction is applied to the first image in the sequence. Later, the 3D points are adjusted along their rays based on the new images." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major comments:\n- The 2D-3D matching approach appears highly dependent on precise camera pose estimation. Sampling along a ray and projecting back to the camera to confirm matches assumes a high degree of pose accuracy, as even minor angular deviations can significantly affect the rays, especially for distant points. Achieving such accuracy is difficult in incremental or real-time systems without extensive bundle adjustment rounds. Consequently, this method can only be applied once pose estimation (and potentially a full reconstruction) is complete. If so, the rationale for incremental processing and online capability in the proposed pipeline is unclear, as an offline method would first need to process the entire sequence. Moreover, image matching with known poses is well-explored, with numerous relevant publications that the authors seemingly overlook. Without being exhaustive, here are a few examples [m1-m4]. Actually, a large portion of the multi-view stereo literature addresses matching, making the proposed appear excessively complicated without clear justification (especially, since the main reason, i.e. being incremental, does not seem to make sense, as I write earlier).\n- The experiments are very weak. Showing results on a single dataset, while all other baselines work well on others as well, is clearly insufficient. Also NICER-SLAM [j] is missing that also only use RGB. \n- Table 1: Throughout the paper, the authors describe their method as \"online\", but it runs at ~1 FPS, which does not truly qualify as online. Additionally, its accuracy and precision are lower than methods that are nearly an order of magnitude faster.\n- Table 2: The same observations apply here as for Table 1. In many metrics, the proposed method is underperformed by significantly faster alternatives.\n- Conclusion: \"Experiments show that our approach achieves state-of-the-art performance.\" This statement is inaccurate. While some metrics may show strong performance, others reveal that it lags behind baseline methods.\n- The assumption of known camera poses should be stated upfront (e.g., in the abstract). The current wording suggests that the authors address both geometry and pose estimation by using the term \"reconstruction\" which is not true. \n\nMinor comments:\n- Experiments: Although the authors opt not to use the depth channel, it would still be informative to show comparative results with methods that do, such as [a,b,c,d], since the ScanNet dataset includes this data. This comparison would help readers understand how RGB-only performance currently compares to RGB-D methods.\n- L053: Missing related work on volumetric methods: [a,b,c,d].\n- L066: Missing related work: [e,f].\n- Paragraph at L141: Not all volumetric methods require a predefined grid [c].\n- Several methods for image ray to 3D point matching are entirely ignored: [g,h,i].\n- Fig.2: The image is very dark; consider using a different image or adjusting the visualization for better clarity.\n- L130: \"volume .\" -> \"volume.\"\n\n[a] Oleynikova, H., Taylor, Z., Fehr, M., Siegwart, R. and Nieto, J., 2017, September. Voxblox: Incremental 3d euclidean signed distance fields for on-board mav planning. In 2017 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) (pp. 1366-1373). IEEE.\n\n[b] Grinvald, Margarita, et al. \"Volumetric instance-aware semantic mapping and 3D object discovery.\" IEEE Robotics and Automation Letters 4.3 (2019): 3037-3044.\n\n[c] Zheng, J., Barath, D., Pollefeys, M. and Armeni, I., 2025. Map-adapt: real-time quality-adaptive semantic 3D maps. In European Conference on Computer Vision (pp. 220-237). Springer, Cham.\n\n[d] Miao, Y., Armeni, I., Pollefeys, M. and Barath, D., 2024. Volumetric semantically consistent 3d panoptic mapping. IROS 2024\n[e] Wang, S., Leroy, V., Cabon, Y., Chidlovskii, B. and Revaud, J., 2024. Dust3r: Geometric 3d vision made easy. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 20697-20709).\n\n[f] Leroy, V., Cabon, Y. and Revaud, J., 2024. Grounding Image Matching in 3D with MASt3R. ECCV 2024\n\n[g] Chen, B., Parra, A., Cao, J., Li, N. and Chin, T.J., 2020. End-to-end learnable geometric vision by backpropagating pnp optimization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 8100-8109).\n\n[h] Zhou, Q., Agostinho, S., Ošep, A. and Leal-Taixé, L., 2022, October. Is geometry enough for matching in visual localization?. In European Conference on Computer Vision (pp. 407-425). Cham: Springer Nature Switzerland.\n\n[i] Wang, S., Kannala, J. and Barath, D., 2024. DGC-GNN: Leveraging Geometry and Color Cues for Visual Descriptor-Free 2D-3D Matching. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (pp. 20881-20891).\n\n[j] Zhu, Z., Peng, S., Larsson, V., Cui, Z., Oswald, M.R., Geiger, A. and Pollefeys, M., 2024, March. Nicer-slam: Neural implicit scene encoding for rgb slam. In 2024 International Conference on 3D Vision (3DV) (pp. 42-52). IEEE.\n\n[m1] Goesele, M., Snavely, N., Curless, B., Hoppe, H. and Seitz, S.M., 2007, October. Multi-view stereo for community photo collections. In 2007 IEEE 11th International Conference on Computer Vision (pp. 1-8). IEEE.\n\n[m2] Žbontar, J. and LeCun, Y., 2016. Stereo matching by training a convolutional neural network to compare image patches. Journal of Machine Learning Research, 17(65), pp.1-32.\n\n[m3] Scharstein, D. and Szeliski, R., 2002. A taxonomy and evaluation of dense two-frame stereo correspondence algorithms. International journal of computer vision, 47, pp.7-42.\n\n[m4] Furukawa, Y. and Ponce, J., 2009. Accurate, dense, and robust multiview stereopsis. IEEE transactions on pattern analysis and machine intelligence, 32(8), pp.1362-1376." }, "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 part." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Using point clouds as a representation for the scene is more scalable and generalizable compared to voxel and implicit surface representations.\n\n2. Loosening epipolar constraints could potentially improve performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a real-time scene reconstruction approach from multi-view images, with a point cloud scene representation. When a new image is introduced, the global scene is optimized by adjusting the locations of existing points, adding new points, and removing redundancies. This process is achieved through a ray-based and learning-based 2D-3D matching technique. Although the method achieves high accuracy, it is more time-consuming and lacks distillation experiments to verify the effectiveness of the key designs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although this method improves reonstruction accuracy compared to baseline methods, it is $6\\times$ more time-consuming, with a processing time of 0.6 s/frame, which limits its feasibility for online reconstruction.\n\n2. In Line 243, the authors mention adjusting only visible points, which can lead to gaps at the edges between visible and non-visible regions. Although scene normals are supervised during training, there is no test-time guarantee to avoid this issue. \n\n3. Individually predicting an offset for each point in the scene adjustment introduces local noise, as seen in Figure 5.\n\n4. There is a lack of detailed ablation studies and analysis on the ray-based matching method (comparing with point-based) and the relaxation of epipolar geometry constraints.\n\n5. As an online reconstruction method, it is necessary to provide hardware testing information and comparisons of memory usage.\n\n6. The paper is somewhat hard to read and 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": 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": "- Line 244: How is visibility determined?" }, "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 sound: the authors address the limitations of monocular depth estimation, introducing multi-view matching to improve depth prediction and ensure consistent 3D reconstruction.\n\n- Extensive experiments benchmark the method against both offline and online approaches with different scene representations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a method for online 3D reconstruction from monocular RGB images. The proposed method maintains a global point cloud, as the 3D representation, by dynamically adjusting, adding, or removing 3D points as new frames arrive. The 3D point update is achieved through a ray-based 2D-3D matching technique, which projects 3D points along rays to another view to gather multi-view information to refine depth predictions along camera rays. The proposed method is evaluated against various prior methods on the ScanNet dataset," }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper would benefit from a high-level overview explaining (1) how the method is initialized on the first frame and (2) how the global representation is iteratively refined with each new frame before detailing individual steps.\n\n- Depth updates rely on a single new image at each step, which contradicts with most multi-view 3D reconstruction methods that integrate multiple views simultaneously. Using only one view at a time has potential drawbacks: 1) Reduced robustness in homogeneous regions compared to multi-view approaches; 2) Limited co-visibility, impacting point matching quality; 3) Suboptimal performance in extreme depth ranges, as the baseline (i.e. the distance between two cameras) is fixed. Could the authors clarify this design choice? I wonder if the less smooth reconstructions observed in the experiments relate to this limitation.\n\n- Since the method relies on stereo feature matching, the view-independent color jittering may negatively impact matching quality. \n\n- While the point cloud is lightweight, the final 3D reconstruction depends on an algorithm to convert the 3D point cloud to the underlying 3D surfaces. The authors use TSDF Fusion, which inherits its limitations in accuracy and resolution." }, "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": "/" }, "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": "- Geometric Metadata Selection: How were metadata elements chosen? Was there an ablation study or reference to prior work guiding the selection process?\n\n- View-Dependent Confidence: Shouldn’t confidence values for each point be view-dependent? For example, a 3D point visible from one viewpoint would have higher confidence in that view but lower confidence if occluded. The current approach to learning confidence seems unclear, particularly regarding occlusions. For instance, suppose two points are aligned along the line of sight from two cameras, where one point is visible in one camera but occluded in the other. The learned confidence may end up equal, averaging the depth between points and failing to handle occlusion naturally. Could you clarify how this method correctly handles occlusions?\n\n- Dataset Generalization: The evaluation is primarily based on the ScanNet dataset. Would PointRecon generalize effectively to outdoor or unstructured environments? What specific challenges might it face in these settings?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Global Consistency: The approach maintains a unified point cloud representation, enhancing consistency across views compared to independent depth predictions.\n\n2. Efficiency in Memory: By using a sparse point cloud approach, PointRecon avoids high memory demands, unlike volumetric methods.\n\n3. Flexibility and Resolution Independence: The point-based approach is free from fixed voxel size constraints, offering flexibility for detailed reconstructions.\n\n4. Competitive Performance: It demonstrates comparable or superior performance to current methods in both depth and mesh quality metrics on ScanNet." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces PointRecon, an online 3D point cloud reconstruction method that builds a 3D scene representation from monocular RGB video inputs. PointRecon employs a ray-based 2D-3D matching technique, allowing 3D points to adjust their locations along camera rays from newly observed images without predefined voxel resolutions. Once new points are added, the method integrates a point cloud merging technique based on learned confidence values, resulting in a globally consistent point cloud representation. PointRecon demonstrates comparable performance to state-of-the-art methods on the ScanNet dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Unclear Advantage Over Prior Methods: The benefits of PointRecon over SimpleRecon or VisFusion, which achieve similar reconstruction quality with lower latency, are not fully evident. Further clarity on the advantages or potential benefits of this method would be valuable. In terms of quality, the proposed PointRecon is simialr to SimpleRecon or VisFusion; In terms of speed, PointRecon is slower than the aforementioned prior work. Is there any aspect that the proposed method has stronger potential than prior work?\n\n2. Latency: The method's sampling approach introduces relatively high latency per frame, particularly during scene adjustment and depth prediction.\n\n3. Noise in Output: The absence of post-processing smoothing results in noisier meshes compared to other approaches with more advanced smoothing techniques.\n\n4. Complexity in Implementation: Ray-based matching and multi-level attention mechanisms increase computational complexity, which may affect scalability. For example, if this method is employed for a larger scene, the ray-based matching computation complexity will increase as the number of rays in the scene increases. Could author test this by running this method on a larger scene, e.g. multi-room environment instead of a single room scene?\n\n5. Limited Justification for Ray-Based Matching: Although an ablation study highlights the value of key components, the core concept of \"ray-based matching\" could benefit from further justification. More comparisons with alternative methods, such as point-based or traditional epipolar line matching, would strengthen the argument for this approach. The authors could run on the same dataset and test the alternatives by switching the matching module." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024pointrecon,\ntitle={PointRecon: Online 3D Point Cloud Reconstruction via Ray-based 2D-3D Matching},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3JfvvuPXsH},\nnote={under review}\n}" }, "abstract": { "value": "We propose a novel online point-based 3D reconstruction method from a posed monocular RGB video. Our model maintains a global point cloud scene representation but allows points to adjust their 3D locations along the camera rays they were initially observed. When a new RGB image is inputted, the model adjusts the location of the existing points, expands the point cloud with newly observed points, and removes redundant points. These flexible updates are achieved through our novel ray-based 2D-3D matching technique. Our point-based representation does not require a pre-defined voxel size and can adapt to any resolution. A unified global representation also ensures consistency from different views. Results on the ScanNet dataset show that we improve over previous online methods and match the state-of-the-art performance with other types of approaches." }, "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": [ "3D 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/99a9a2338fee735d6bd3b5c3a8c76ae55b717141.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": "PointRecon: Online 3D Point Cloud Reconstruction via Ray-based 2D-3D 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": 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": "**Improving clarity and motivations**\n\n- The distinction between low-level (coarse) and high-level (fine-grained) features is confusing, lacks clarity and contributes negatively to the overall appreciation of the paper. Could the authors precise what they refer to as high-level or fine-grained features and as coarse or low-level features? It is confusing to refer to \"semantics\" as coarse, as they are supposed to provide very precise and high-level descriptive information. Similarly, the spectrogram is often referred to as high-level, although it describes low-level features similar to the features extracted by the cochlea (line 86).\n\n- The description of the CLAP training (with aligned textual descriptions and audio) (line 147) and the use of prompts for incorporating music genres and phenotype information (line 117), are considered as coarse features, why? \n\n- The paper refers to as \"inverse pathway of auditory processing\" (line 17). If it is inverse; shouldn't it go from high-level to low-level? seems that the method goes from coarse-to-fine ... \"AudioMAE latent feature as the fine-grained acoustic embedding of audio\" (line 182) isn't it coarse information?\n\n- Typo line 466 ->\"it's better\"\n\n- In line 52, what are the DNN features referred to? could the authors provide more explanations?\n\n\n**Methodology/Results:**\n\n- It is unclear how the Semantic Decoder extracts semantically rich information (line 137); is it validated somehow? \n\n- One of the baseline modes is a Transformer that goes from voxel space to mel-spectrogram space. Could the authors provide more information about this implementation, as it does not seem straightforward?\n\n- Could the authors provide some motivations for using ridge regression to model brain signals (referred to as $x$ in the paper)? does this consider the autocorrelation of brain cortical activation and using the structure and spatial organisation of the auditory cortex?\n\n- Regarding the generation process and the use of the latent diffusion model, how much do the authors consider providing the ground truth as input to help the model (figure 3)? Isn't the conditioning should be enough?\n\n- In tables 1, 2, and 3, are the results averaged across subjects? if yes could you provide the std? \n\n- In Table 1, why is C2F-LDM low for PCC and PSNR?\n\n- The objective of section 3.3. and 3.4 are not very clear, for instance, lines 457 to 460. More importantly, the results do not seem to go in the direction of the conclusion of the paper, e.g. in Figure 5. b. why does the fine-grained decoding perform better for almost all experiments? The hypothesis made in section 3.3, e.g. \"This could be attributed to participants being more focused on the content of the stories during fMRI signal collection, potentially disregarding the speaking style of the speaker.\" (line 453) is used to motivate the next section 3.4 but is relatively weak and unclear. In section 3.4, The Brain2Music dataset performs better without prompts, although the prompts are supposed to incorporate important information such as the music genre. Why use prompts, then? seems somehow superficial. \n\n**Data**:\n\n- what about variation in performance across subjects? are the models trained per sujects? or models are used across subjects?\n\n- how are the voxels in Table 4 extracted? Is it from a template? subject-based parcellation?\n\n- Are the datasets aligned functionally (on top of anatomical registration) so that similar voxels and auditory regions across subjects can be compared?\n\n- For all datasets, the audio clips seem to be between 1.5 and 4 seconds. Is it enough to aim at reconstructing audio from a very short fMRI temporal window? What about using larger windows?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The technical contributions of the paper are strong. The methodology, with the threefold approach, is well thought out and carefully implemented. Many implementation details are provided, and the notations are clear and easy to follow. This is a great point.\n\n- Figure 2 is clear and provides sufficient details to understand the intricate methodology\n\n- The new method is compared against various baselines and three openly available datasets. Also, the code is made public, and additional results are provided. This is a great point for the openness of science. \n\n- Various ablation studies are done to confirm the contributions of specific modules (for instance the diffusion reconstruction vs using the MAE decoder)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to achieve audio reconstruction from an fMRI brain signal via a coarse-to-fine approach. The idea is to replicate the audio processing stream in the human auditory cortex. The method is threefold: first, it uses a CLIP-based approach to extract an audio representation (low-level description) from the fMRI data signal. From these initial features, a high-dimensional description of the auditory feature is obtained via a guided AudioMAE. Finally, these high-level features are used as a condition of a Latent Diffusion model for mel-spectrogram reconstruction and, thus, audio reconstruction. The method achieves high results on three publicly available datasets, demonstrating strong performance for both low-level and high-level audio metric reconstructions, showing improvement compared to the direct reconstruction of mel-spectrogram and other high-level approaches which omit low-level features." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- One weakness of the paper is the limited neuroscientific motivations for audio reconstruction as a way to understand auditory mechanisms (first paragraph of the introduction). It is not clear from the introduction or the conclusion how solving this task could help to understand better the auditory processes (line 28 \"This research contributes to the comprehension of the human auditory\nsystem\" ). Are these models generalisable between subjects? can a model trained on a single subject be used for another subject? can they be used to identify specific features related to language disorders? \n\n- Another major weakness in this study is the lack of clarity in the writing, which prevents us from understanding some of the motivations and results/discussion. The characterisation of \"high-level/low-level\" features and \"coarse/fine-grained\" features throughout the paper is very unclear. Some of these terms seem interchangeable in how they are employed; the paper would greatly benefit from a clear definition (see more related questions in the next section).\n\n- Some points of the methodology lack details and/or seem to diminish the results (see questions in the next section). For instance, the modelling of the brain signal with ridge regression, which overlooks the spatial autocorrelation of the brain signal and structure of the auditory cortices, is not motivated. Sections 3.3 and 3.4 lack clarity, it is not clear what they aim to achieve with respect to the general motivation of the paper, and the conclusion are not clear and even seem to diminish previous results (for instance Figure 5 and line 457 to 460) - see more questions below. \n\n- There are some important missing descriptions regarding the dataset and the training procedure. How is the data split between training and testing? is it subject-wise or dataset-wise? Are the same subjects used both for training and testing? functional alignment between subjects if direct comparison between 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": 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": 1 }, "primary_area": null, "questions": { "value": "1)\tFor each of the metrics, it would be helpful to understand what a good value is, and if these are getting at all close to that value. Currently, it is difficult to interpret whether the observed differences are at all significant. For instance, for the semantic representations one could use a baseline score that is two different samples in time from the same audio clip. And maybe a PCC and PSNR score could be referenced to samples with additive Gaussian noise of various SNR levels at the waveform? \n\n2)\tListening to the supplement audio files, I am not so sure that the proposed method is actually doing something reasonable. The samples “sound” more natural (which is expected from including a diffusion model in the pipeline), but they are often completely different sounds than the initialized audio, essentially hallucinations. PSNR levels of ~15-18 seem like they might be noise, relative to the original sound. \n\n3)\tIn the decoding section, one of the experiments is on male/female decoding which is called a “semantic” task. However, I believe a decent amount of male/female decoding can be achieved from simply looking at the overall power spectrum of the sound (male voices are generally lower than female voices). Thus, this doesn’t seem particularly “semantic”. \n\n4)\tThe metrics are not defined enough in the main text (Lines 311-319). At a minimum, the acronyms need to be defined. For instance, what is “PCC” and “PSNR”? Although some audio researchers may be familiar with these measurements, a more general ICLR attendee may not. \n\n5)\tHow many repetitions of each sound are present in each dataset and are these averaged for the analysis? More generally, how is fMRI measurement noise taken into account for the analysis? I.e., is there a sort of “noise ceiling” defined on the reconstruction? \n\nSomewhat minor suggestions for specific lines: \n* Line 075-077: This sentence doesn’t make sense to me. How does the “high-dimensionality” play a role here? The main challenge of fine-grained decoding is the lack of resolution in time and the inherently noisy signal of the fMRI response. \n* Lines 086-088: These references seem out of place, as the decomposition of sound into difference frequencies in the cochlea dates much further back than these papers. Some good references might be work by Shihab Shamma (maybe https://ieeexplore.ieee.org/abstract/document/119739, or https://pubmed.ncbi.nlm.nih.gov/16158645/) , but I would encourage the authors to perhaps cite classic textbooks or review articles on auditory processing \n* Lines 147-152: Which features are used in CLAP? The final output features or an intermediate stage? This should be mentioned in this section. \n* Line 231: “unpatchify” is not defined." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper tackles an interesting and challenging problem of trying to decode auditory fMRI activity. I like how it ties back into classic ideas of course-to-fine reconstruction. It also takes advantage of many recently proposed auditory models, similarity metrics, and auditory datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work designs a system to reconstruct audio signals from human fMRI data collected while listening to natural sounds. The system separately handles fine-grained information (i.e., the precise timing and frequencies of the sounds) and course-grained information (semantic information such as the class of sound) by using different neural network features for each branch of the architecture. The authors evaluate their method on three different fMRI datasets and multiple different evaluation methods capturing the fine-grained and semantic nature of the sounds." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The paper combines many different pre-trained systems. This makes the methods used difficult to understand, but also makes it hard to evaluate where things might be going wrong, and which pieces are critical. Due to the unknown inherit biases of each system, it is difficult to see how the work could actually be useful for understanding the human auditory system.\n \n* As currently written, the presentation of results is confusing. The text should point to each set of results when they are discussed, rather than stating (line 357-359) that qualitative results are displayed in Tables 1 and 2. Additionally, it would be helpful if the columns were in some way separated or labeled for the “fine grained” vs. “semantic” 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": 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": "Figure 5 is confusing. What are the ‘features’ that are input into the SVM? And given the results and discussion that there is ‘little semantic content in the semantic features’ (line 449) the initial claim that semantic prompts during decoding ‘[enhances] the quality of reconstructed audio’ seems perhaps exaggerated (lines 25-26). \n\nA lot of the text is pushed to the appendix, making the official 10 pages lacking in sufficient detail and discussion. \n\nA lot of the writeup feels pretty inside baseball to this reviewer (or perhaps this reviewer is just too far outside this particular game). e.g. CLAP is not explained (though there is a reference) until page 3." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Technically, the method introduced here is interesting and seems sophisticated. \n\nThe paper is well-written and makes use of highly relevant contemporary work—including studies done in machine learning and neuroscience—and uses this grasp of the literature to provide interpretations on different steps in their method. \n\nThe paper does a nice job of comparing many different methods in addition to their own." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new method for reconstructing audio signals from fMRI responses to sounds. Specifically, it proposes a course-to-fine decoding strategy in which fMRI responses from auditory cortex are decoded in a course-grained fashion into the semantic space of CLAP (which is not defined until page 3), and decoded in a fine-grained fashion into acoustic space.\n\nI would not recommend this paper for acceptance. While the high-level approach seems interesting, the purpose of integrating neural data seems lacking in motivation, and the use of pre-trained representations in both training and evaluation seem to present a major confound. \n\n\n\nFigure 5 is confusing. What are the ‘features’ that are input into the SVM? And given the results and discussion that there is ‘little semantic content in the semantic features’ (line 449) the initial claim that semantic prompts during decoding ‘[enhances] the quality of reconstructed audio’ seems perhaps exaggerated (lines 25-26). \n\nA lot of the text is pushed to the appendix, making the official 10 pages lacking in sufficient detail and discussion. \n\nA lot of the writeup feels pretty inside baseball to this reviewer (or perhaps this reviewer is just too far outside this particular game). e.g. CLAP is not explained (though there is a reference) until page 3." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of Clear Motivation: This reviewer got stuck at square one: what is the goal of this work? The authors don't tell us. If the goal were to understand brain function this work would proceed differently, for example separately analyzing primary versus non primary auditory cortex, asking which brain regions are best modeled by each component of the model, comparing decoded signals to human behavioral data, etc. Alternatively, perhaps it could be intended for BCI applications. But what would those applications be? If you have access to a person's auditory cortex then you would already have access to the sound they were hearing, so there would be no point trying to decode that information from the brain. A paragraph in the appendix (A.7) attempts to provide a motivation, but it doesn't make sense to me. For example it mentions that this work could \"aid individuals with voice disorders\". But if an individual had a voice disorder, one would want to generate the intended spoken information, not the heard information from auditory cortex.\n\nIt’s unclear how the findings here contribute to the ‘comprehension of the human auditory system’ as stated in the abstract (lines 27-29). While the coarse-to-fine method is inspired by the human auditory system, it is not quite a model of ‘each physiological structure of the auditory processing pathway’ (lines 96-97)\n\n2. Possible confound.\nThe results on many of the metrics suggest that the direct decoding methods better reconstruct audio than the C2F-LDM proposed here (Table 1, 2). The novel C2F-LDM method does improve on measures of FD, FAD, KL, and CLAP, but these results seem to present a confound. The reconstruction makes use of many model features from pre-trained models and are subsequently evaluated on their similarity to pre-trained model features. Thus would higher scores not be unsurprising?\n\n3. Is the fMRI data even relevant?\nThe paper should address the role of fMRI data and how it improves the methods described here. The majority of steps requires pre-trained models and predicting their ‘ground truth’ representations, but the optimal value of P=0.25 seems to suggest that using these ground truth representations without the neural data may improve reconstruction." }, "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": 4 }, "primary_area": null, "questions": { "value": "Questions:\n- Why was sex selected as the semantic class for the brain-to-speech case?\n- Just confirming that: semantic features of guidance = course-grained features in the CLAP space? (might be helpful to state this more explicitly)" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Overall, this paper represents a novel improvement for brain-to-audio decoding and its acknowledgement would benefit the advancement of the field. The approach introduced is novel, performance state-of-the-art, and the presentation clean." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a coarse-to-fine framework for audio reconstruction from fMRI brain recordings that outperforms the leading solely fine-grained approaches. This approach draws inspiration from the hierarchical processing found in the human auditory system. It first projects the audio and fMRI into a coarse-grained semantic embedding in the CLAP (Wu et al, 2023) space before paring that embedding again with the fMRI signal for fine-grained decoding, generating features in the space of AudioMAE (Huang et al, 2022). Lastly, an LDM is used to reconstruct the mel-spectrogram of the stimulus audio with the fine-grained embedding, before it is converted to the waveform using a pretrained HiFiGAN (Kong et al, 2020) vocoder.\n\nThe authors evaluate the performance of this framework over three different datasets encompassing three different classes of audio / reconstruction task (sound, music, and speech). The quality of the reconstructed audio is assessed by FD, FAD, KL, and CLAP score and the mel-spectrograms are evaluated using PCC and PSNR. Performance is benchmarked against direct decoding approaches and fine-grained only approaches. The authors also offer an investigation into the quality of the semantic information captured through these approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are a few places where choices are made without explanation and parts of the discussion on the semantic analysis are unclear. For example, in the discussion of the semantic analysis it is stated that there is less semantic richness in the brain-to-speech case because listeners might be more focused on content. However, this effect does not necessarily seem to hold for the fine-grained only decoding (which is not addressed). It seems more likely that coarse-grained decoding may just be ill-suited to capturing semantic quality from speech as the signal clearly exists given the performance of the solely fine-grained approach. This represents a potential limitation of this framework for brain-to-speech." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024reverse,\ntitle={Reverse the auditory processing pathway: Coarse-to-fine audio reconstruction from f{MRI}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3JoLo0mmHH},\nnote={under review}\n}" }, "abstract": { "value": "Drawing inspiration from the hierarchical processing of the human auditory system, which transforms sound from low-level acoustic features to high-level semantic understanding, we introduce a novel coarse-to-fine audio reconstruction method. Leveraging non-invasive functional Magnetic Resonance Imaging (fMRI) data, our approach mimics the inverse pathway of auditory processing. Initially, we utilize CLAP to decode fMRI data coarsely into a low-dimensional semantic space, followed by a fine-grained decoding into the high-dimensional AudioMAE latent space guided by semantic features. These fine-grained neural features serve as conditions for audio reconstruction through a Latent Diffusion Model (LDM). Validation on three public fMRI datasets—Brain2Sound, Brain2Music, and Brain2Speech—underscores the superiority of our coarse-to-fine decoding method over stand-alone fine-grained approaches, showcasing state-of-the-art performance in metrics like FD, FAD, and KL. Moreover, by employing semantic prompts during decoding, we enhance the quality of reconstructed audio when semantic features are suboptimal. The demonstrated versatility of our model across diverse stimuli highlights its potential as a universal brain-to-audio framework. This research contributes to the comprehension of the human auditory system, pushing boundaries in neural decoding and audio reconstruction methodologies." }, "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": [ "Brain-to-audio reconstruction", "Coarse-to-fine", "fMRI", "Auditory processing pathway" ] }, "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/296b37669aec07b55233dd1c1ade8dbb4a82ac73.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": { "value": "/attachment/a1016a7e03eab2baf41d0ed4b5a1cd1f419271ad.zip" }, "title": { "value": "Reverse the auditory processing pathway: Coarse-to-fine audio reconstruction from fMRI" }, "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": "In Line 017, what does \"targeting the training dataset\" mean?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "There are both theoretical and experimental demonstrations of the effectiveness of the algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an algorithm for dataset distillation by incorporating distributionally robust optimization into it. There is theoretical justification and empirical validation of the proposed algorithm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There seems to be a mismatch between the motivation and the experiments. The motivation emphasizes regions with low population density, which usually correspond to the worst-group performance in subpopulation shift [1]. However, the main experiments related to distribution shift are conducted on test sets with perturbations or the worst group induced by an additional clustering process. It would be better to conduct more experiments on subpopulation datasets included in [1]. \n2. The introduction has too many paragraphs, which makes the logic of the introduction tedious with poor readability. \n\nSome minor issues:\n\n- In Line 049, \"some technique\". \n- In Line 129 and 131, \"Algorithm\" \"Numerical Results\" their first letters do not need to be capitalized. \n- In the last paragraph of introduction, Section 3 is not mentioned.\n\n[1] Yang, Yuzhe, et al. \"Change is Hard: A Closer Look at Subpopulation Shift.\" *International Conference on Machine Learning*. PMLR, 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": 3 }, "primary_area": null, "questions": { "value": "Please refer to the detailed suggestions provided in the weaknesses section." }, "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 applying distributional robust optimization to dataset distillation, providing a reasonable approach to enhance generalization.\n\n2. Drawing on distributional robust optimization theory, this work establishes a theoretical foundation to support the proposed approach to dataset distillation.\n\n3. The paper is well-structured, with clear algorithm block and effective visualizations that enhance the presentation of the work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work proposes a robust dataset distillation approach that incorporates distributional robust optimization (DRO) to enhance generalization and performance across subgroups. This method combines clustering with risk-minimized loss to conduct dataset distillation. By prioritizing representativeness and coverage over training error guarantees, the approach enhances the models trained on synthetic datasets for real-world scenarios. Both theoretical analysis and empirical validation on multiple standard benchmarks are provided, demonstrating the effectiveness of the proposed approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The empirical experiments focus primarily on CIFAR-10, ImageNet-10. Extending the evaluation to larger, real-world datasets with a greater number of classes would better demonstrate the effectiveness of proposed approach in generalization and robustness under real-world conditions.\n\n2. Comparing the proposed approach with additional baseline method addressing generalization and robustness would provide a more comprehensive assessment of the proposed approach, including comparisons with baseline methods such as [1, 2, 3].\n\n3. It would be valuable to visualize the robust inference tasks using real-world data rather than illustrative visuals, which could provide more insight and observations in real-world scenarios.\n\nReference:\n\n[1] Domain generalization with adversarial feature learning. CVPR 2018.\n\n[2] Self-challenging Improves Cross-Domain Generalization. ECCV 2020.\n\n[3] HYPO: Hyperspherical Out-of-Distribution Generalization. ICLR 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": 3 }, "primary_area": null, "questions": { "value": "I have stated many of my suggestions and concerns in the weaknesses section. Below, I have further questions with some minor issues.\n\n* While theoretical analyses have been provided, I wonder whether the proposed method would affect the convergence rate and make it more difficult to find an optimal solution for the proposed objective. I would appreciate an empirical time complexity analysis regarding the proposed method.\n\n* In Eq.10, what does the *N* represent? I did not see any introduction of the *N*." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* This paper considers the group coverage and generalization ability of the synthetic dataset in Data Distillation(DD), which is interesting and novel.\n\n* The introduced algorithm is clear and the theoretical analysis seems solid.\n\n* The numerical results do show the effectiveness of the proposed algorithm. Meanwhile, the figures (Fig.3, 4) seem to indicate that the proposed method improves group coverage." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new data distillation approach emphasizing the subgroup coverage and generalization ability of the models trained on the synthetic dataset. This paper provides a theoretical analysis rooted in Distributionally Robust Optimization (DRO) and verifies the effectiveness of the proposed method with various experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* While the paper emphasizes group coverage and generalization in data distillation, the experiments are mainly conducted on IID datasets. More experimental results in scenarios with distribution shift between training and testing sets (such as long-tail classification, subpopulation shift, domain adaptation, domain generalization, etc) can further validate the improvement in group coverage and generalization ability,\n\n* According to Algorithm 1, the initialization of the synthetic dataset seems very important because it involves how the training data samples are clustered into subgroups. It may require further ablation study to verify the stability of the proposed algorithm.\n\n* The proposed method seems like a plug-in as it only modifies the objective for data distillation and could be combined with any data distillation methods. A more sophisticated comparison between the proposed method with other objectives regarding training gradients, feature distributions, and training trajectory could help readers better understand the improvement of the proposed 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": 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 weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- As far as I know, this paper firstly addresses a major limitation in standard dataset distillation (underrepresented or rare subgroups) by applying a double-layer distributionally robust optimization (DRO) framework. This approach ensures the synthetic dataset better represents the full diversity of the data, reducing performance drops across different data subgroups. \n\n- This paper provides not only empirical evidence of the proposed method's robustness against domain shifts such as noise, blur, and inversion, but also a theoretical analysis of the effectiveness of CVaR in the loss function to enhance robustness.\n\n- The proposed method is designed to be adaptable and can integrate with various existing distillation techniques, such as gradient or distribution matching. This modularity makes it compatible with a range of distillation methods, which are still actively developing and evolving." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a robust dataset distillation method that enhances generalization across underrepresented subgroups by using a double-layer distributionally robust optimization (DRO) approach. Traditional dataset distillation often compresses training data into synthetic sets by matching training characteristics but may struggle with subgroup generalization. To improve this, the authors cluster the training data around synthetic data points and apply a Conditional Value at Risk (CVaR) loss to minimize worst-case subgroup errors, making the synthetic dataset more representative of diverse data distributions. Experimental results show that this method significantly improves robustness under domain shifts, outperforming baseline methods on benchmarks like CIFAR-10 and ImageNet-10, particularly in challenging conditions like noise and blurring." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed DRO approach, particularly with clustering and CVaR, may introduce significant computational overhead. This added complexity could be amplified for large-scale datasets due to the clustering of training (real) data. However, the paper does not discuss the computational overhead of the proposed method.\n\n- Although the method shows promising results on benchmarks like CIFAR-10 and ImageNet-10, the experiments are limited to controlled domain shifts (e.g., noise, blur). Testing under more realistic settings, such as in transfer learning, would further validate its robustness and practical relevance. For example, one could (1) train neural networks on a synthetic dataset distilled from a coarse-grained dataset (e.g., ImageNet) and (2) fine-tune and evaluate them on a fine-grained dataset (e.g., Birds 200). This setup would better illustrate the method's effectiveness in addressing the challenges posed by rare subgroups." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024group,\ntitle={Group Distributionally Robust Dataset Distillation with Risk Minimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3JsU5QXNru},\nnote={under review}\n}" }, "abstract": { "value": "Dataset distillation (DD) has emerged as a widely adopted technique for crafting a synthetic dataset that captures the essential information of a training dataset, facilitating the training of accurate neural models. Its applications span various domains, including transfer learning, federated learning, and neural architecture search. The most popular methods for constructing the synthetic data rely on matching the convergence properties of training the model with the synthetic dataset and the training dataset. However, targeting the training dataset must be thought of as auxiliary in the same sense that the training set is an approximate substitute for the population distribution, and the latter is the data of interest. Yet despite its popularity, an aspect that remains unexplored is the relationship of DD to its generalization, particularly across uncommon subgroups. That is, how can we ensure that a model trained on the synthetic dataset performs well when faced with samples from regions with low population density? Here, the representativeness and coverage of the dataset become salient over the guaranteed training error at inference. Drawing inspiration from distributionally robust optimization, we introduce an algorithm that combines clustering with the minimization of a risk measure on the loss to conduct DD. We provide a theoretical rationale for our approach and demonstrate its effective generalization and robustness across subgroups through numerical experiments." }, "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": [ "dataset distillation", "distributional robustness", "generalization" ] }, "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/35a7cae7e063b234ddfba466ed26f98e53d5ac8a.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": { "value": "/attachment/3be5099d9326f830a4acdbe829f8de20f4fd4f94.zip" }, "title": { "value": "Group Distributionally Robust Dataset Distillation with Risk Minimization" }, "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": 1 }, "primary_area": null, "questions": { "value": "1. The 0-1 Knapsack Problem is known to be NP-hard primarily because its capacity can be as large as $2^n$. However, in this problem, the capacity is limited. Could this constraint make a brute-force search feasible?\n2. Quantization-based model compression techniques, such as W4A16, can reduce model size to 25%. If singular value decomposition (SVD) is combined with quantization, could this approach yield further compression? Given that different decomposition methods offer varying levels of precision, is it possible to identify an SVD approach with high tolerance to quantization?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. **Automated and Importance-Aware Truncation Selection**: AutoTrunc automates the selection of optimal truncation positions using a learning-based approach to model layer importance, focusing on layers critical to performance. This approach streamlines the compression process, maximizing compression efficiency while maintaining accuracy.\n2. **Theoretical Soundness**: Built on rigorous NP-hard problem analysis and efficient budget allocation strategies, AutoTrunc is a theoretically grounded method, ensuring reliable performance estimates and compression quality across applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents AutoTrunc, an automated framework for selecting optimal truncation positions in singular value decomposition (SVD) to compress large language models (LLMs) efficiently. Unlike previous methods that overlook layer importance, AutoTrunc uses a learning-based approach to model each layer’s contribution to overall performance, optimizing truncation to maximize compression while preserving model accuracy. By addressing the truncation selection problem as 0-1 Knapsack Problem with efficient algorithms and dynamically allocating memory based on layer sensitivity, AutoTrunc achieves superior compression results. Experimental evaluations show that AutoTrunc outperforms existing SVD-based methods, reducing perplexity by up to 38.63% on LLaMA-2-13B at a 50% compression ratio, enabling more efficient LLM deployment without retraining." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Limited Model Diversity**: The experiments focus solely on the Llama-2 model family, which uses multi-head attention. However, recent open-source LLMs, such as the Llama-3 and Qwen-2/2.5 families, have adopted Group-Query Attention, which might lead to different outcomes in weight compression. This lack of diversity in model structures limits the generalizability of the findings.\n2. **Limited Throughput Improvement**: AutoTrunc achieves only modest gains in inference throughput (approximately 1.1x from 0% to 60% compression), whereas other methods, such as SVD-LLM, achieve over 2x speedup at similar compression levels. This limited throughput improvement may reduce AutoTrunc’s impact in applications where throughput is a critical factor.\n3. **Lack of Comparison with Quantization Techniques**: The paper does not thoroughly compare AutoTrunc’s performance against other popular compression methods, such as quantization (AWQ, GGUF, GPTQ). Without these comparisons, it is challenging to assess AutoTrunc’s effectiveness, especially in contexts where quantization might offer a better trade-off between compression and 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": 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 above." }, "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 addresses a good research topic: SVD for LLM compression.SS\n \n2. The paper is well-organized." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an automatic way to search for the optimal truncation position when compressing large language models with SVD. The authors first empirically show that the truncation position is highly correlated to the final performance. Based on the observation, they modeled the layer importance based on the correlation and design a way to obtain the optimal configuration. Experimental results demonstrate the effectiveness of the designed searching strategy for the truncation position." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Poor Presentation**  Many aspects starting from Section 2.2 to Section 3.2 are not clarified clearly. Specifically,\n \n 1. The data collected in Figure 1 is also used for modeling the correlation, but why the method only needs to collect 6x40=240 data? why only need to measure compression ratio ranging from 10% to 50%?\n \n 2. Why the author collects the modeling data by only applying the uniform compression ratio and greedy search?\n \n 3. The computed upper-bound is also confusing. On the one hand, it is correlated to the manual configuration Fmin, meaning that setting a larger Fmin could increase the performance? On the other hand, it is correlated to the learned modeling parameter, indicating that changing the pre-defined modeling function from linear one to a more precise unlinear one could also impact the upper-bound. Therefore, it is hard to tell whether reaching this upper-bound is truly the optimal solution.\n \n 4. Since the both the empirical observation in Section 2.2 and the modeling in Section 3.1 are highly data-dependent. What if we change the data distribution for this empirical analysis and modeling?\n \n2. **Overclaim on Compression Speed:** The author claims that the search process of the recent work, ASVD, is slow, however, I found that the designed method still needs to measure the end-to-end perplexity under different compression ratios, which is similar to what has been done in ASVD. Additionally, the proposed method runs learning-based algorithm to model the correlation between truncation position and corresponding perplexity, which is also time-consuming. Given these two situation, it is hard to claim that the proposed automatic searching algorithm is more efficient than prior works.\n \n3. **Missing Experiments:**\n \n 1. Pruning-based compression methods are different from the SVD-based ones, and their compression ratios are not exactly equal to the ratio of parameter reduction in LLM. Therefore, it is not fair to compare these two types of methods under the same compression ratio.\n \n 2. Lack of experimental comparison on generation tasks.\n \n 3. Lack of comparison on quantization-based compression methods.\n \n 4. Lack of analysis on running the methods using data with different distribution." }, "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": "Following are a few questions that can benefit the paper \n- Is it possible to quantify both with experiments and asymptotically the extra effort required to derive layer-wise ranks? This can help demonstrate the gains of the method relative to the SVD-LLM.\n- What was the observation when the lower-bound was not applied and yet the AutoTrunc is allowed to automatically configure the layer-wise low-ranks for decomposition? It is understandable the performance might be taking a hit significantly, probably in many cases, worse than SVD-LLM, but it is worth showing that as a baseline to demonstrate the efficacy the lower-bound. This maybe especially useful to highlight the fact that ideal low-rank decomposition may not necessarily be good all the time and hence having a constrained (with the lower-bounds on each layer) adaptive/automatic truncation can be justified even better. \n- In the similar spirit, is it a better strategy to set an overall compression ratio as a single hyper-parameter as opposed to set a low-bound on each layer and let the algorithm decide the truncation of the singular values in each layer? This in itself might be a new direction can be time-consuming for this review, but certainly a potential future direction and a new method. \n- The proposed method is applied only on LLaMa-2-7/13/70B models, why not apply on other family of LLMs such as Mistral/Phi-3/X/Y/Z even if it is in the realm of 7B or lower parameter models? The questions here are, i) the learned `\\alpha(s)` or even the strategy of AutoTrunc can be transferred to other family of models, ii) If the answer is yes, what is common/different in all these models that is actually contributing to the improved performances or a layer-wise study can be better explained, iii) if the answer is no, then can this be transferred easily to new family of LLMs or do we have to repeat the whole method from scratch for a new LLM architecture?\n- The performance of the model drops >=50% compression ratio when compared to SVD-LLM, any hunches as to why?\n- Is it possible to compare and contrast different SOTA methods discussed in this paper, in terms of latency and memory efficiencies at different compression ratios?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well-written and easy to follow along.\n- The contributions in the paper are nicely positioned w.r.t the state-of-the-art in the literature.\n- The strongest point(s) of the paper are i) applying SVD-LLM on each layer of the LLM in an adaptive manner ii) learning the layer importance, applying the lower bounds on the compression ratios at each layer (of course this is different from the over-all compression of the entire LLM), LambdaRank for listwise ranking iii) empirically quantification on sub-layers the correlation between performance and the model quality." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the critical problem of compressing large language models using SVD based decomposition. The significant contribution of the paper is a theoretical backing with empirical evidence to automatically truncate the singular values/vectors of each layer instead of applying a uniform low-rank on each layer like in the SVD-LLM method case. The paper has a learnable strategy that learns to decompose a given layer in an LLM using layer importance modeling." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Follow the questions section" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024automatic,\ntitle={Automatic Truncation Position Selection in Singular Value Decomposition for Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3KEwJGYNzH},\nnote={under review}\n}" }, "abstract": { "value": "Model decomposition in large language models has drawn much attention due to its superiority and good interpretability, where activation-aware singular value decomposition (SVD) can achieve competitive performance by mitigating reconstruction errors brought by outliers in activation. However, the performance of the state-of-the-art SVD-based LLM compression method is limited to the selection of truncation positions. No work meticulously examines the details of this problem theoretically and empirically tests its correlation with model performance. To fill the research gap, we propose an efficient method that can automatically select truncation positions, namely AutoTrunc. In our work, we first analyze the correlation between truncation positions and the model performance. Then, the model layer importance is modeled based on the correlation, followed by mathematical proof to illustrate how to reach and obtain the optimal truncation position configuration for different layer types. Extensive experiments are carried out to verify our presumption and evaluate our proposed method. Our proposed AutoTrunc outperforms the state-of-the-art SVD-based LLM compression method, with perplexity scores dropping by 24.65% and 38.63% at the compression ratio of 50% in LLaMA-2-7B and LLaMA-2-13B, respectively. The code will be released upon acceptance." }, "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": [ "Model decomposition; Large Language Model; 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/8a50081b03350d69682ccd05f6638f5db4ced0fd.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": "Automatic Truncation Position Selection in Singular Value Decomposition for 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": 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 above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The study is highly relevant and timely for NAS-related fields, with the potential to generate a high impact on the community. \n\nThe effort invested in this work seems quite substantial.\n\nThe writing and the presentation are clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study produced a dataset of 7k unique models, Younger, from 174k publicly available models and 30 tasks. After processing and filtering, architectures are stored as acyclic graphs based on ONNX definitions. The study aims to automate architecture generation and refinement. It offers a range of statistical analyses to illustrate the diversity of Younger. Several experiments are also included to show the potential of Younger, especially as a benchmark for GNN." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The goal of this study is quite ambitious, allowing people to search for good architectures using Younger for a wide range of tasks. The experimental section shall match that ambition, e.g. demonstrating the applicability and benefits of Younger on a wide range of tasks. This would help illustrate the practical application of the dataset.\n* Provide a concrete example, or case study showing how Younger could be used for a specific task like CIFAR-10 classification and ImageNet classification.\n* Other possible cases can be performing time series prediction\n* Or creating generative models.\n* Or multimodal models for a specific task e.g. text-to-video or description or caption generation task.\n\n---\n\nAlso the paper claims Younger is advantageous in comparison with benchmarks like DARTS, NAS-Bench-201. Other than the difference between them, as shown in Table 1, how would they compare in a specific task, e.g. using different sets for the same task? The study should show Younger's advantages clearly, e.g. leading to better performance, reducing search costs etc.\n* Conduct a comparative experiment using Younger and a benchmark like NAS-Bench-201 on a common task, measuring metrics like search efficiency and final model performance.\n\n---\n \nClearly state whether AIGNNA is a novel term introduced in this paper. If not, provide the appropriate citation.\n\n---\n\nThe code and the dataset link cannot be found in the submission.\n* Provide direct links to the code repository and dataset, perhaps in a dedicated \"Resources\" section." }, "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": "Please see the weaknesses. Overall, this paper is interesting and the proposed YOUNGER dataset seems to be useful in several directions. Giving all these, I’d like to recommend the score 6 temporarily." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1) Constructing the dataset for neural architectures in the AI-generated manner is novel and interesting. \n2) The paper is nicely written and well organized. The details for the dataset and construction processes are clearly presented. \n3) The AI-generated architectures in different types are very helpful to several research directions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel dataset for AI-generated neural network architectures. Specifically, the construction process contains four core steps, i.e., retrieving NN models, converting the models to ONNX format, extracting DAGs from the ONNX models, and filtering out isomorphic DAGs to ensure the uniqueness of the architectures. Some experimental results support the statistical analysis and present some distributions of the whole dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) More details in terms of the use of the proposed YOUNGER dataset are expected. For example, it is possible to provide at least one case using the YOUNGER dataset, such as performing some NAS algorithms to search for architectures in this dataset? \n2) The experimental results are mainly focus on GNNs. However, it seems that there can be other types of architectures contained by YOUNGER. Could the authors provide additional experimental results in terms of other types of architectures beyond GNNs?\n3) More details for the distribution of the performance of architectures are needed." }, "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 the paper it is stated in line 284 that \"less than 1% of these models represent heterogeneous and effective architectures. This notably low proportion of heterogeneous architectures highlights the limitations of current neural network design methods, both manual and NAS-based, in fostering architectural innovation\". \n\nHow did the author decide whether an architecture is heterogeneous? and what does it mean by effective architecture? In addition, all architectures are from online sources should be mostly manually built right? It is really surprising that out of all models available online, less than 1% are really valid." }, "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 presents a good contribution to the community by offering a dataset that can potentially faciliate the research in the neural architecture search field.\n\n2. The dataset published can overcome challenges of previous related datasets in terms of operator scope and scale." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a dataset for neural network architectures extracted from public model sources. Then convert these models to intermediate representation operators for further research purpose. The dataset incorporates richer operator than existing neural architecture datasets, to facilitate more research in the field." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One complaint I have regarding this dataset comparing with existing NASbench dataset is that the proposed dataset doesn't seem to have the associated training performance on a standardized dataset. In comparison, the NASbench dataset is evaluated on a standard task (Cifar10) on different settings. Because of this, user can only perform unsupervised analysis like shown in figure 3-6." }, "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": "Section 4.3.2: different tasks have different evaluation metrics and ranges for different datasets [9], so how do you handle that?\nSection 4.3.3: Can you provide more information on exactly what this is, in terms of framing it through citations and potential tables/figures presenting the results? That would be a better use of page real estate than Figs 2-6 which are more appendix details.\n\nReferences:\n\n[1] Liu, Yuqiao, et al. \"Bridge the gap between architecture spaces via a cross-domain predictor.\" Advances in Neural Information Processing Systems 35 (2022): 13355-13366.\n\n[2] Mills, Keith G., et al. \"Gennape: Towards generalized neural architecture performance estimators.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 8. 2023.\n\n[3] Yang, Yichen, et al. \"Equality saturation for tensor graph superoptimization.\" Proceedings of Machine Learning and Systems 3 (2021): 255-268.\n\n[4] Zhang, Chenhao, et al. \"Towards better generalization for neural network-based sat solvers.\" Pacific-Asia Conference on Knowledge Discovery and Data Mining. Cham: Springer International Publishing, 2022.\n\n[5] Mills, Keith G., et al. \"Building Optimal Neural Architectures using Interpretable Knowledge.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.\n\n[6] Schrodi, Simon, et al. \"Construction of hierarchical neural architecture search spaces based on context-free grammars.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[7] Salameh, Mohammad, et al. \"AutoGO: automated computation graph optimization for neural network evolution.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[8] White, Colin, et al. \"How powerful are performance predictors in neural architecture search?.\" Advances in Neural Information Processing Systems 34 (2021): 28454-28469.\n\n[9] Mills, Keith G., et al. \"Aio-p: Expanding neural performance predictors beyond image classification.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 8. 2023." }, "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 dataset consists of many diverse architectures from different tasks.\n- Breaking the confines of predefined search spaces is a good step forward for NAS." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript proposes Younger, a dataset comprising of neural network ONNX graph representations collected from various online repositories. Each ONNX graph is annotated with some kind of task performance. The goal of Younger is to fuel Artificial Intelligence-Generated Neural Network Architecture (AIGNNA) generation, in order to break the confines of pre-established search spaces. The manuscript consists of tables and figures tabulating/illustrating features of the Younger dataset, as well as some limited experiments on AIGNNA Exploration." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses:\n- Primary weakness of this paper is that its key contribution is not novel at all. While it is true that predefined search spaces pose a large problem for NAS, there are existing works that break the confines of predefined search spaces, enabling generalizable prediction across different micro, cell-based NAS search spaces [1] and even between micro and macro search spaces [2], using ONNX as the generalizable representation [3, 4, 5] or otherwise. This paper does not seem to recognize such existing works.\n- AIGNNA in Section 4.3, the authors illustration is flawed for Fig. 7 left. You would not have a choice between 'Add' and 'ReLU' since ReLU is an activation applied to one single input, while 'Add' combines several inputs into one. A better choice would be to ask 'ReLU' or 'SiLU' or 'Add' vs. 'Concat'. Here still, there is existing work on generating new architectures outside of pre-existing search spaces [6, 7]. \n- Experimental results in Section 4.3.2 is not very convincing as you are focusing on testing different GNN designs (which govern the message passing rule, not necessarily graph feature design), while there are numerous neural predictors in the literature [8] which are better compared to. Also, ACC, F1, Prec. and Recall are generally not evaluation metrics for neural predictors; rather, rank correlation (Kendall's Tau or Spearman Rho) and regression error (L1) are more common [8]. \n- Overall presentation in the paper is very lacking. Table and Figure captions are too short and do not adequately convey enough information, while most figures (2 - 6) should be re-tweaked with larger fonts." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024younger,\ntitle={Younger: The First Dataset for Artificial Intelligence-Generated Neural Network Architecture},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3LFR5N2uv8},\nnote={under review}\n}" }, "abstract": { "value": "Designing and optimizing neural network architectures typically require extensive expertise, starting from handcrafted designs followed by manual or automated refinement, which significantly hinders rapid innovation. To address these challenges, Younger is introduced as a comprehensive dataset derived from over 174K real-world models across more than 30 tasks from various public model hubs. After extensive processing and filtering, Younger includes 7,629 unique architectures, each represented as a directed acyclic graph with detailed operator-level information based on ONNX operator definitions, enabling compatibility across different deep learning frameworks. The dataset is designed to support the emerging research area of Artificial Intelligence-Generated Neural Network Architecture (AIGNNA), which aims to automate their generation and refinement. Comprehensive statistical analysis, including architecture component analyses, highlights the diversity and complexity of architectures in Younger, revealing the potential for future research in this domain. Initial experiments, including operator and dataflow predictions, demonstrate the dataset's utility for architecture exploration and evaluation, and highlight its potential as a benchmark for graph neural networks. Furthermore, an online platform ensures continuous maintenance and expansion of the dataset, supporting global researchers in their endeavors. The dataset and source code are publicly available to encourage further research and lower entry barriers in this challenging domain." }, "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": [ "Artificial Intelligence-Generated Neural Network Architecture", "Neural Architecture Design", "Graph Neural Network", "Benchmark", "Dataset" ] }, "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/e5114cbff2460c3340f5a44e414d247dd7599255.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": null, "title": { "value": "Younger: The First Dataset for Artificial Intelligence-Generated Neural Network Architecture" }, "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": "1. Section 4.1 lists joint training and post-training for only 1 epoch. Usually multiple epochs are required to train a model. While post-training can be understood as vision encoder and much of language decoder may already be well-trained from prior stages, 1 epoch for joint training seems pretty small. Any reason why that worked? Is there a study on how more epochs affected the outcome? Is it possible that there isn't much data diversity between training and test set, and hence, 1 epoch is enough?\n\n2. What are the training/inference latency gains by using a smaller size model like GOT compared to Qwen-VL-MAX or others?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper presents a unified end-to-end model for a gamut of OCR documents, including sheet music, geometry and number-centric charts. It replaces the cascaded OCRs specialized in different document types.\n\nThe way the three stages of training are applied to unify a diverse set of OCR tasks (scene, document, chart, music sheets, etc.) within a single OCR is interesting. The task-oriented fine-tuning is limited to post-processing the language decoder. Freezing the vision encoder avoids increasing the computational demands and ensures foundational visual understanding is stable across the tasks. \n\nThe results are compared against the SOTA methods on a variety of metrics including F1-scores, edit distances, BLEU and METEOR values, and seem to outperform majority of the methods. For box-guided and color-guided OCR, specific comparison to Fox Lie et al. seems to outperform against all the metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The manuscript proposes a unified end-to-end 2.0 model for OCR, called GOT (General OCR Theory) using LVLMs (Large Vision Language Models). The architecture contains 80M parameters in the encoder, and 500M parameters in the decoder tackling long-contexts. Region-based recognition, dynamic resolutions and multi-page OCR are few other properties of GOT. It supports English and Chinese and can produce structured formats like markdown, tikz, smiles and kern. \n\nGOT has a 3-stage training process: pre-training the vision encoder, joint-training of encoder and decoder, and finally the post-training of the language decoder. The performance is compared against SOTA methods on various scores like edit distance, F1, BLEU and METEOR, and seems to out-perform against majority of the SOTA methods. The results on markdown, sheet music, geometry and number-centric charts are also presented." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weakness of the paper lies in its novelty. The 3-stage training process is well known in the literature. For example, many existing frameworks in OCR, vision-language and LVLMs decouple encoder pre-training from the rest of the pipeline. The vision encoders are usually pre-trained on a wide variety of data to create a foundational understanding of text and scene. The joint training of vision and language pieces is again known in models in UniT, BLIP and LVLMs. Lastly, the fine-tuning of the language decoder piece is again seen in T5 etc. Perhaps, the prime novelty is the application of these methods to an OCR problem, smarts about synthetic data generation and OCR-specific fine-tuning.\n\nThe other weakness of the paper is in its presentation. The paper is overall hard to follow, as it continues to mix, architecture, training, data and task-specific details all together, and does not lay out in separate sub-sections. E.g. the section 3.2.1 starts with the architecture, dives into input sizes, parameter sizes, goes through data peculiarities (natural scenes, cropped slices) and training process all in one paragraph. A lot of architecture diagrams can be added to aid the reading.\n\nLastly, in experiments, ablation studies are missing to underscore the importance of each of the stages, data types. Latency studies, comparisons with SOTA methods, and failure cases are missing." }, "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": 3 }, "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": 4 }, "primary_area": null, "questions": { "value": "1. How does the performance of the proposed method fare on openly and widely used page-level datasets (such as CASIA-HWDB, HCCDoc, CTW1500, ReCTS, IAM, CROHME16/19, etc.)? Why was the effectiveness of the proposed method not tested on these commonly used datasets in the community?\n\n2. Are the test datasets used in sections 4.2.2, 4.2.3, and 4.2.5 open-source?\n\n3. In the references, proprietary acronyms should be capitalized, for example, CASIA, IAM, HDWB, BLIP, etc.\n\nAdditional comment: I do not agree that low training and inference costs must be a characteristic of OCR 2.0. As a new technology framework or paradigm for OCR in the era of AGI, it should also possess scalability capabilities." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper introduces a unified OCR-2.0 model, emphasizing an end-to-end architecture whichl is designed to handle various OCR tasks efficiently.\n\n- GOT demonstrates versatility by recognizing a wide range of artificial optical characters, including sheet music, geometric shapes, and charts. The model can adapt to different input styles and output formats, enhancing readability for formulas and tables.\n\n- This paper is well written and well organized.\n\n- The idea of OCR 2.0 is interesting and novel." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an so-called OCR-2.0 model named GOT, designed for advanced Optical Character Recognition tasks. It proposes a new OCR model, emphasizing end-to-end architecture, low training costs, and versatility in recognizing a wide range of artificial optical characters. The model, with 580M parameters, incorporates a high-compression encoder and a long-context decoder for handling various OCR tasks.\n\nGOT is evaluated on multiple OCR tasks, demonstrating superior performance in plain document OCR, scene text OCR, formatted document OCR, fine-grained OCR, and more general OCR tasks like sheet music, geometric shapes, and chart OCR." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) The term \"general OCR theory\" is not appropriate as the paper does not present any rigorous theory. It is suggested to consider alternative terms such as General OCR Technology/Framework/Pipeline/Methodology.\n\n(2) Dataset construction is a significant contribution of this work. The authors utilized data engineering methods to create a substantial amount of non-public training data. If these datasets are not made publicly accessible, it will make it challenging for other researchers to perform fair comparisons under the same settings as this paper.\n\n(3). In section 4.2.2, the authors collected 400 natural scene text images for testing. Why did they not use publicly available datasets in this domain (such as CTW1500, ReCTS, etc.) to evaluate the performance of GOT on natural scene text? I am wondering if the proposed method on these public datasets can achieve state-of-the-art (SOTA) 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": 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": 2 }, "primary_area": null, "questions": { "value": "Show in the part of 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 article presents a unified approach to OCR recognition tasks, making it one of the most comprehensive OCR models to date with sufficient tasks.\n2. The proposed GOT method employs a three-stage pre-training and fine-tuning process to achieve the experimental results outlined in the paper.\n3. The GOT method addresses various OCR recognition problems across multiple scenarios (natural scenes, documents, etc.), as well as different levels of granularity, such as document-level and region-level recognition.\n4. Multiple datasets are constructed to conduct these diverse settings of these OCR recognition tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the GOT model (General OCR Theory), to improve upon traditional OCR systems (OCR-1.0). With 580 million parameters, GOT processes various artificial optical signals and supports multiple output formats. It features interactive region-level recognition and incorporates dynamic resolution and multipage OCR, showing superior performance in experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing of this article needs further improvement, as several key details are missing. For example, when discussing the method, it is unclear how to distinguish between different tasks. Does it involve using a question as input to the decoder, similar to existing MLLMs?\n2. In the experiment, the paper does not conduct the comparisons on the benchmarks of OCRBench, InfoVQA, and DocVQA. Is this because the proposed method does not support QA? (You did not clarify how you distinguish between different tasks?)\n3. This paper mainly focuses on recognition issues related to OCR tasks and does not address detection problems. One possible reason could be that both the current encoder-decoder and decoder-only architectures struggle with coordinate regression prediction, which may have prevented you from tackling detection tasks.\n4. Additionally, there is a lack of comparison with methods like Kosmos?" }, "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": "- Some more details would be necesary on how metrics are computed given the full recognized text and ground-truth . \n- Also some more details on how the OCR task on charts is defined" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Compared to other unified end-to-end frameworks for multi-task OCR based on multimodal LLMs, the proposed approach is efficient and the model is relatively smail. \n- The proposal of a new training strategy adding complexity increasingly to the model, either from the point of view of the model and the data used for training. \n- The generation of a large collection of data to train the model can be useful for advancing research in generic OCR (if the data is made public after publication)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper describes a single unified framework to perform end-to-end OCR in different kinds of images (documents, scene images, handwritten text, charts, music sheets, math formula). The framework relies on collecting a large amount of data for every type of image, partially from public data sources, partially automatically rendered. Then, a curriculum strategy is employed to train the model based on standard encoder and decoder architectures. In a first stage, only a limited number of OCR tasks with limited variability are used to train the encoder using a simple decoder and, progressively more data, tasks and the final decoder architecture are included in subsequent training stages. Experimental results compare the proposed approach with other generic models based on multimodal LLMs" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper lacks contextualization and comparison with previous SoA OCR methods not based on LLMs, specialized on each of the individual OCR tasks. Related work lacks a much better discusion and reference to all existing specific methods for text recognition in different tasks (scene text, documents, handwritten text, ...). In the experimental results I also miss comparison with specific OCR methods in each task, even in some tasks comparison with existing commerical OCR tools. \n- Following the previous comment, I think that the papser should also use common standard benchmarks and datasets in some specific OCR tasks. In the past years there has been a huge effort in the text recognition community to create standard benchmarks for evaluation, that are ignored in the paper. Using these common benchmarks (for all the tasks where this is possible) would help to get a better understanding of the contribution of the proposed approach in comparison with existing OCR techniques. \n- As far as I understand, most of the images used to train and evaluate the proposed approach are very clean images, collected from clean pdf documents or automatically rendered, without the kind of noise, distortion, low resolution problems, ... that can be encountered when dealing with real images. \n- I miss some analysis of the contribution of each of the training stages in the final performance of the model." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024general,\ntitle={General {OCR} Theory: Towards {OCR}-2.0 via a Unified End-to-end Model},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3LOcwfB4JX},\nnote={under review}\n}" }, "abstract": { "value": "Traditional OCR systems (OCR-1.0) are increasingly unable to meet people's usage due to the growing demand for intelligent processing of man-made optical characters. In this paper, we collectively refer to all artificial optical signals (e.g., plain texts, math/molecular formulas, tables, charts, sheet music, and even geometric shapes) as \"characters\" and propose the General OCR Theory along with an excellent model, namely GOT, to promote the arrival of OCR-2.0. The GOT, with 580M parameters, is a unified, elegant, and end-to-end model, consisting of a high-compression encoder and a long-contexts decoder. As an OCR-2.0 model, GOT can handle all the above \"characters\" under various OCR tasks. On the input side, the model supports commonly used scene- and document-style images in slice and whole-page styles. On the output side, GOT can generate plain or formatted results (markdown/tikz/smiles/kern) via an easy prompt. Besides, the model enjoys interactive OCR features, i.e., region-level recognition guided by coordinates or colors. Furthermore, we also adapt dynamic resolution and multi-page OCR technologies to GOT for better practicality. In experiments, we provide sufficient results to prove the superiority of our model." }, "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": [ "OCR", "LVLM", "Multimodal" ] }, "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/0db133a9ee6a504fb4672927d1f04a368b4c5580.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/a3be147f4e6febd6dceb8256d6b2dc7f703d89c5.zip" }, "title": { "value": "General OCR Theory: Towards OCR-2.0 via a Unified End-to-end Model" }, "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": "The same as in the Weaknesses part." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- A simple yet effective method to elicit the internal representations of black-box models\n- A series of detailed experiments demonstrating QueRE's effectiveness across various benchmarks and settings, comparing it favorably against more complex, resource-intensive methods.\n- Strong practical application value.\n- Mathematical foundation" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an effective method called QueRE, designed to infer the internal representations of black-box language models through self-queries, particularly in follow-up question scenarios. The authors represent the black-box model by using the probability of the \"yes\" token as vector values, and then train a linear model to achieve the following objectives: 1) accurately predict model performance at the example level, and 2) assess various states of the language model, such as detecting if the model has been influenced by harmful prompts, determining if it has been replaced by another model, and identifying the architecture and size of the models. The authors also demonstrate that this approach generalizes beyond top-K token reliance, making it applicable to sample-based methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am impressed with this paper, both by its strong results and its practical applications. Could you elaborate on the intent behind your design choices? Additionally, did you explore other methods that ultimately proved less effective or failed to yield similar 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": 3 }, "primary_area": null, "questions": { "value": "- While your method is superior to the baselines, can you comment on how much more expensive/cheaper it is?" }, "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 ideas are conveyed in a way that is accessible and logically coherent, making the methodology and results easy to follow.\n2. This work is particularly significant given the increasing reliance on black-box LLMs by researchers and developers who lack full access to model internals. By providing a practical and scalable way to elicit representations from these models, the paper addresses a growing need.\n3. The method of querying black-box models with elicitation question to construct representations is creative and the method outperformed strong baselines. The method was also applied to three different tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method to extract black-box representations from large language models (LLMs) by querying them with elicitation questions. The approach leverages the probabilities of model responses to these questions, creating low-dimensional representations that can be used to predict model performance on specific tasks, detect adversarial manipulation, and distinguish between different model architectures and sizes. The authors demonstrate that these black-box representations are effective, often outperforming other techniques that even rely on internal model states." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper labels the extracted low-dimensional vectors as “representations,” but this may be overstated. These vectors are simply derived from yes/no responses to elicitation questions, which only provide limited insights into the model’s deeper knowledge or reasoning structures.\n2. My interpretation is that the classifier is primarily learning to detect shifts in the model’s calibration—the confidence in its yes/no responses—rather than meaningful behavioral changes. This is limiting since if a model provider adds the system prompt (“Be helpful and cautious.”), it could alter the model's calibration and trigger your classifier as detecting an adversarial/harmful LLM (task 3). I think any added system prompt for that matter would trigger a false positive. Furthermore, if system prompts were appended to all models by the model providers, I'm not sure you could still reliably classify between models (task 2)?\n\nI’m not sure how useful the proposed method is for actual tasks beyond predicting model performance (point 1), and I'm not sure if the method is actually robust to the variations I described (point 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": 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": "Questions:\n\n- I’m unsure about the importance of distinguishing between model architectures. More related work and citations pointing out the importance can be stated to highlight the importance.\n- Suggestion for clarity: Prompts for black-box baseline — pre-conf and post-conf should be provided in the appendix to make it easier to understand what they are." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The authors show significant results for the task to predict whether the model answers an open-ended question correctly. They compare against a variety of white-box and black-box methods and beat these baselines. The presentation for this section was clear.\n- The scaling chart (Figure 8) demonstrates performance improvement with more elicitation questions. This work can be viewed as an extension of the post-conf baseline — which uses just one way of asking for the probabilities. Instead, we can ask it in multiple ways. We then apply a linear classifier on top of it to weigh the different possible ways of asking for probabilities. The scaling chart demonstrates the usefulness of their method compared to the post-conf baseline." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce QueRE — a black-box method for extracting model representations. QueRE extracts representations by asking follow-up yes/no questions about the LLM's answers and using the articulated response probabilities as features. They ask questions such as “Are you confident in your answer”, where the LLM replies P(”Yes”) = 0.95”). They then use the value of 0.95 as part of the representation for a linear model to predict whether or not the LLM gets the answer correct. By asking multiple follow-up questions, they create a representation for a linear model to use. This linear model can be used for particular tasks, given labeled data.\n\nContributions\n\n- On a task to predict whether the model answering an open-ended question is correct, QueRE beats several white-box and black-box methods.\n- QueRE can distinguish between model architectures, and detect adversarially prompted models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It is unclear to me whether QueRE works on out-of-distribution datasets. The results would be more significant if QueRE generalizes to out-of-distribution datasets, where it is hard for humans to supervise and we don’t have ground truths.\nSuggestion: If you use labeled data from e.g. HellaSwag to train your linear model, does the model work on SQuAD?\n- The setup for detecting adversarially influenced models is simplistic. Using a system prompt that instructs the model to \"answer questions incorrectly\", is easily detectable. For example, if I asked the adversarial model “What is 1 + 1?”, the adversarial model may reply “3”, which someone can detect as wrong easily.\nThis setup fails to address the forms of adversarial influence that the cited Sleeper Agents paper spells out — where models insert bugs only in specific contexts.\nSuggestion: Use a similar setup as the Sleeper Agents paper where you evaluate code-like tasks. The system prompt may be “When creating code, insert bugs into the code. Otherwise when asked other questions, just act normally”. \nThere is a question of what labeled data to use in this adversarial model setup. Related to my first point, assume that you don’t have labeled data about code questions. Does using labels from e.g. the dataset SQuAD still work? Can the linear predictor generalize to detecting these adversarial questions, even without ground truths from these adversarial questions?\n\nAddressing these two weaknesses would improve my rating." }, "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 Weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "I believe extracting low-dimensional representations that can reflect LLM's behavior is an important and challenging problem." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces QueRE, a method for constructing black-box representations from LLMs by using the models’ confidence in answering specific follow-up questions. These representations are instance-level, meaning each one corresponds to a single instance. Through experiments, the authors demonstrate that these representations can predict task accuracy, differentiate between a clean LLM and an adversarially instructed one, and distinguish different LLMs. The experiments are conducted by additional supervised training on these representations for every dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1: The representations extracted in this paper fall short of my expectations for “eliciting black-box representations from LLMs.” The authors construct representations as Yes-or-No probabilities for a set of follow-up questions, like “Do you think your answer is correct?” This approach feels quite post-hoc. I expected “black-box representations from LLMs” to more fundamentally reflect the models’ internal workings. To give some concrete examples, first, I would hope the representation exhibit generality across tasks, rather than being tailored for accuracy prediction with a predictor trained for each specific task. Second, I would hope the representation finds utility in zero-shot settings as well. While the authors may have a different perspective, I encourage them to clarify their definition of “black-box representation” early in the paper.\n\nW2: 2. Following up on the previous comment, if the authors could demonstrate that a single predictor can be trained across tasks to predict accuracy and generalize to new, unseen tasks without any known ground truth, it would help demonstrate the generality of the extracted representations.\n\nW3: Tables 2, 7, and 8 show that random sequences can outperform QueRE for some models and tasks, suggesting that the specific follow-up questions chosen may not be critical. Random sequences are considered among the weakest baselines, yet they outperform QueRE, leading me to believe that simply elicitating more information from the LLM is the main reason giving the empirical performance. Better constructed follow-up questions can perform better, but a “random kitchen sink” approach already works to some level. If this is the case, it should be acknowledged in the paper, as it, in my view, lessens the contribution of this work. Could the authors conduct a more thorough analysis of why random sequences sometimes outperform QueRE, and discuss the implications this has for their method?\n\nW4: Building on the idea of random sequences, I suggest the authors compare QueRE with more diverse follow-up question strategies. For example: (1) rephrasing the same question in different ways and eliciting confidence, or (2) using questions from the same task with known ground truth (since all experiments assume access to a small amount of labeled data from downstream tasks), or (3) using questions from other tasks with known ground truth.\n\nW5: Please discuss empirically the importance between the number of eliciting questions and the elicitating question strategies (e.g., between one designed in this paper, random sequences, others mentioned in W4).\n\nW6: I’m curious why QueRE outperforms full logits in Figures 2 and 3. Shouldn’t full logits contain strictly more information than QueRE? Could you provide a more detailed analysis of why QueRE outperforms full logits in these cases?\n\nW7: If this difference is due to the fact that full logits are only extracted for the initial response while QueRE includes additional follow-up questions, could the authors concatenate full logits from all follow-up questions and include this as a comparison?\n\nW8: Additionally, I suspect the use of linear classifiers might contribute to these results, since QueRE is low-dimension and full logits are high dimension. Could the authors compare QueRE’s performance with more complex classifiers? Linear classifiers may favor low-dimensional representations.\n\nW9: QueRE appears to extract and combine uncertainty or confidence signals. Could the authors compare QueRE to a baseline that directly concatenate different existing LLM uncertainty estimates? For example, [1].\n\nW10: If I understand correctly, the experiments in Sections 4.2 and 4.3 are ablation studies, rather than comparisons to external methods. All baselines—pre-conf, post-conf, and Answer Probs—are components of QueRE, as noted in lines 297-307. This setup makes it impossible for QueRE to perform worse than these baselines. I suggest that the authors highlight the ablation nature of these experiments in Sections 4.2 and 4.3. Additionally, it would be beneficial to compare QueRE with existing methods that can be applied.\n\nW11: The experimental setup in Section 4.3 may be seen as limited and less realistic. Adversarial LLMs explicitly designed to answer incorrectly are easy to distinguish and less reflective of real-world scenarios. I suggest the authors consider existing methods that contaminate or manipulate LLMs in more subtle ways.\n\n[1] \"Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs.\"" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a technique to extract representations from black-box language models through querying about its behavior; these representations are useful for predicting performance or if a model has been adversarially influenced." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024eliciting,\ntitle={Eliciting Black-Box Representations from {LLM}s through Self-Queries},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3LifGYAD0W},\nnote={under review}\n}" }, "abstract": { "value": "As large language models (LLMs) are increasingly relied on in AI systems, predicting when they make mistakes is crucial. While a great deal of work in the field uses internal representations to interpret model behavior, these representations are inaccessible when given solely black-box access through an API. In this paper, we extract representations of LLMs in a black-box manner by asking simple elicitation questions and using the probabilities of different responses \\emph{as} the representation itself. These representations can, in turn, be used to produce reliable predictors of model behavior. We demonstrate that training a linear model on these low-dimensional representations produces reliable and generalizable predictors of model performance at the instance level (e.g., if a particular generation correctly answers a question). Remarkably, these can often outperform white-box linear predictors that operate over a model’s hidden state or the full distribution over its vocabulary. In addition, we demonstrate that these extracted representations can be used to evaluate more nuanced aspects of a language model's state. For instance, they can be used to distinguish between GPT-3.5 and a version of GPT-3.5 affected by an adversarial system prompt that makes its answers often incorrect. Furthermore, these representations can reliably distinguish between different model architectures and sizes, enabling the detection of misrepresented models provided through an API (e.g., identifying if GPT-3.5 is supplied instead of GPT-4)." }, "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": [ "LLMs", "representations" ] }, "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/abc81c3a8bc98e07dac2f02620859bdf9f4fd8c8.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/c83ab5f1bd2d340131b8ffb61213da9364b37fd0.zip" }, "title": { "value": "Eliciting Black-Box Representations from LLMs through Self-Queries" }, "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": { "value": "No ethical 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": 2 }, "primary_area": null, "questions": { "value": "1) Can the authors list the advantages of using your benchmark instead of existing, human labelled or synthetic datasets for reasoning? In other words, is your benchmark a proxy of some high-order reasoning capabilities? What are the advantages of your approach over standard tests (e.g., compositionality tests that one can generalise to prevent memorisation and data leakage?).\n\n2) Why didn’t the authors show that your results correlate (or do not) with those on popular benchmarks in reasoning and/or simple translation?\n\n3) Say that one of those questions you use to create your benchmark is present in the training set of an LLM. How does that affect the performance of your translation? How do you ensure that a model is not interpolating the answer they potentially have to make the translation easier?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The idea of using translation to test a model’s capability is interesting, especially considering that such a dataset can be scaled automatically in size and breadth. The experiments with their dataset are comprehensive and cover many sota models and standard translation metrics. Despite rushed, the article is easy to read and the ideas expressed are clear enough." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a new task to test the reasoning capabilities of LLMs based on translation. The idea is to take standardised tests, define the rules of a new language, and ask an LLM to translate an example expressed in the new language and aided by a sufficient set of rules. By construction, the translations are not part of the training data and require simple symbol manipulation (e.g., logical operations). The authors test their benchmark on different LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have one big concern for this article.\nThe lack of details on what the benchmark is testing caused me some issues in understanding what capabilities it is testing. While the authors describe the task and provide a few examples, its small size (140 samples) and the lack of a detailed analysis of what kind of capabilities are required by a model to solve it (the authors mention multiple time logics and, I reckon, compositionality) make it very hard to draw comparisons with existing datasets. Furthermore, the authors do not run concurrent experiments on similar tasks (e.g., a baseline), to show correlation with existing benchmarks. In other words, is this benchmark telling us something about other popular LLMs’ benchmarks? If that’s the case, one can argue that your dataset is a proxy for some high-order reasoning and use your dataset (for example, because one can synthetically create new instances easily and does not suffer from memorisation) instead of one created by human experts.\n\nThe article seems rushed (see the paragraph on related work or the methodology, paragraph “Closed Frontier Models”). \nFurther, in lines 269-270, it’s your duty to run the experiments in time and before the deadline; we cannot discount the fact that GPT-4o-mini has a longer response rate, especially considering that your benchmark is very small. It’s better to present organic experiments on all the models or exclude them from the evaluation rather than preliminary results that may not be statistically significant or affected by sample bias.\n\nFigure 1 is confusing and difficult to interpret after reading the first section. Even after reading the methodology, I still do not fully understand what it represents.\n\nRelated works are rushed, with a few references missing (line 97) or inconsistent formatting (see 107-108 vs. 122-123). Some sentences are not grammatically consistent (lines 112-13), and others express vague concepts (line 125)." }, "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. Is it possible to design a CON->CON task? \nAs you mentioned in Sec. 5.3, the knowledge of English pattern will affect the overall performance. \nSince you are trying to use new information to evaluate, CON->CON is a natural thought to eliminate the influence of the English pattern.\n\n2. Can you add the analysis of the impact of the tokenization? You also mentioned this limitation in the paper, but I assume this is an important factor to prove that CLEAR is truly evaluating the reasoning ability. I suggest creating a token-based translation that every token is not an actual word. If this setup aligns with the existing performance, readers can assume that CLEAR is truly evaluating the reasoning ability.\n\n3. Can you add more reasoning paradigms, such as deductive reasoning, in the benchmark? I think you can provide unseens rules as premises and see whether LLMs can reason over these premises. This is an important reasoning paradigm when you want to evaluate the reasoning ability, especially when LLMs cannot do well in it." }, "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 studies an overlooked problem in current reasoning tasks that they, to some extent, rely on the internal knowledge of the LLMs, rather than completely on the reasoning ability.\n2. CLEAR evaluates the reasoning ability in a comprehensive way, supported by staged tasks that incrementally increase in complexity. It provides a nuanced understanding of the model's capability.\n3. The evaluation is comprehensive, covering a diverse range of models, and it points out the substantial room for improvement in reasoning capabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new reasoning benchmark CLEAR, aimed to assess the ICL and reasoning ability of LLMs.\nThey evaluate these abilities through introuducing new information, specifically, a new language.\nGiven the new information, the LLM should derive logical rules and update the meaning of tokens based on new information.\nThe results show that this is a challenging tasks for recent LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The size of the dataset is limited, which is also reflected on the analysis of prompt complexity. This limitation is noted in the paper but may affect the robustness and generalizability of the findings.\n2. There is a potential problem of the entanglement of tokenization and reasoning in this particular tasks. I think further analysis is required rather than just mentioning it in limitations.\n3. CLEAR mainly focuses on inductive reasoning, but general reasoning often involves deductive and abductive reasoning. Improving the diversity of the task forms could significantly increase the 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": 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. Are there plans to expand the dataset to include a broader range of Conlangs with varying linguistic complexities? \n2. Do you have more insights on failure analysis to show if the error of existing models comes from not being able to understand the grammar, or maybe bias towards the language they're pre-trained on? For example, provide step-by-step accuracy for each question. This could also help to extend the scale of the benchmark. As mentioned in weakness 2, do you have reasoning paths generated from each LLM? What's the error rate for each step? Also, what is the distribution of different error types, e.g. mismatching between words (vocabulary), failure to identify plural (grammar), etc? \n3. The identification of the translation direction's strong impact is a good direction to look into. Could the authors provide more error analysis to illustrate specific difficulties models face in this direction? Are there common patterns in errors that suggest improvements in task framing or prompting?\n4. In section 4.3, you briefly mentioned the structure of the prompt including the system prompt and few-shot examples. From what I understand the examples are like the ones shown in Figure 2. It'd be better if you could include the system prompt you used in the paper as well to show how the model is guided for this specific task." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. CLEAR is an innovative benchmark that addresses the limitations of current benchmarks by testing logical reasoning without relying on prior exposure to specific linguistic patterns. The idea of using Conlangs ensures that the task requires genuine reasoning rather than memorization or retrieval, which helps to better evaluate such ability of LLMs.\n2. The paper is thorough in constructing a well-defined benchmark, with clear methodologies for dataset creation and evaluation. The variety of evaluation metrics and the stratification of translation tasks by difficulty level add rigor to the framework, enhancing the robustness of the results. Visuals such as example translation tasks (Figure 1) and the ranking results (Table 1) are useful in illustrating the key aspects of CLEAR and help readers understand the task structure." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces CLEAR (Conlang Logic Evaluation And Reasoning), a new benchmark aimed at evaluating the reasoning capabilities of large language models (LLMs) using translation tasks between constructed languages (Conlangs) and English. Unlike natural language tasks, Conlangs present unique challenges, as they are intentionally designed to be outside the training data, reducing the risk of memorization and promoting logical reasoning. The benchmark includes translation tasks from English to Conlang and vice versa, with increasing complexity in translation rules. Popular LLMs, both closed-source and open-source, are evaluated on CLEAR, and multiple evaluation metrics, such as BLEU, ROUGE, METEOR, and character-based metrics, are applied to assess model performance on logical reasoning within language translation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. With repeated exposure to Conlang structures, LLMs may develop specialized reasoning paths tailored to the benchmark, rather than generalizable reasoning skills. The limited dataset could falsely encourage models to adapt to specific patterns in the Conlangs, thereby reducing the benchmark’s efficacy in evaluating general reasoning capabilities.\n2. The benchmark focuses on output accuracy without providing insights into the reasoning paths which is crucial for evaluating the reasoning ability. As the grammar and structure of Conlang are relatively simple, including an analysis of intermediate reasoning steps could reveal where models tend to struggle, offering diagnostic insights beyond final translation accuracy. There are related previous works like chain-of-thought, and least-to-most that prompt the model to show explicit reasoning paths. In your evaluation, do you ask the model to generate reasoning paths like in Figure 2, or it only outputs the answer? Some analysis on such paths could improve the quality of the work. \n3. As mentioned by the authors, the size and task of the benchmark is quite limited. This would affect the robustness of the benchmark and could make it harder to generalize the findings." }, "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 weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tThis paper provides a compelling approach to evaluating LLMs on unfamiliar language tasks, offering new insights into model capabilities beyond conventional datasets.\n2.\tThe experiment design is sufficiently thorough to support the paper’s objectives, ensuring a robust evaluation of the models' in-cotext learning capabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces CLEAR, a benchmark designed specifically to assess the translation and reasoning capabilities of LLMs in novel tasks. This benchmark evaluates LLMs through few-shot translation tasks using constructed languages (conlangs)—artificial languages crafted to be unfamiliar and absent from model training data. By engaging models in translation tasks that combine logical reasoning with pattern recognition, CLEAR aims to evaluate the models’ abilities to infer grammatical rules and apply logical operations without relying on prior knowledge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe \"learning a new language\" task feels more akin to in-context learning, emphasizing the models’ imitation and induction abilities rather than the “reasoning abilities” claimed in this paper. While \"reasoning\" is a term that can describe various tasks and skills, its use here in a “translation-like” task seems unsuitable.\n2.\tThe paper devotes excessive space to fundamental concepts. For instance, the section on few-shot prompting in related work and the descriptions of common metrics like BLEU could be more concise. Additionally, sections on logical reasoning in related work, including *Logical Reasoning of Large Language Models* and *Logical Reasoning*, are somewhat repetitive and lengthy.\n3.\tThe paper is somewhat difficult to follow. The task remains unclear even after reading the introduction, with clarity only starting to emerge in Figure 2. Several typos are present, such as “To‘learn’” in Line 72. Additionally, the text in Figure 1 is too small to read comfortably. Page 8 would benefit from reorganization if it only includes two tables.\n4.\tThe analysis is somewhat superficial, focusing mainly on translation direction without offering deeper insights to inform future work. Also, could the unexpected relationship between model performance and question complexity be due to an ineffective measure of complexity?\n\nOverall, I feel the current version seems rushed and does not meet the average ICLR standard." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present CLEAR, a new benchmark to evaluate the reasoning capabilities of LLMs through the translation of novel constructed languages." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024clear,\ntitle={{CLEAR}: Understanding the Reasoning Capabilities of Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3LnTTHDWER},\nnote={under review}\n}" }, "abstract": { "value": "Despite significant progress, accurately assessing the reasoning capabilities of Large Language Models (LLMs) remains both a challenging and divisive subject.\nMany existing benchmarks either suffer leakage, or reflect patterns in the training data, leading to ambiguous results.\nWe present CLEAR (Conlang Logic Evaluation And Reasoning), a novel benchmark designed to test the reasoning and problem solving capabilities of LLMs in new environments.\nCLEAR uses Conlangs (Constructed Languages) for few-shot translation tasks,\nwhich require some linguistic knowledge to solve, but primarily the ability to make new patterns from tokens in unfamiliar contexts using logical operations.\nThese conlangs represent a unique challenge, as while translation examples are plentiful, these conlangs each have a unique combination of rules, are self contained, and are absent in the training corpus.\nWe present an evaluation of current frontier models over multiple metrics as a baseline for future research. \nWe will be releasing \\dataset as a public benchmark to drive progress towards AI systems more capable of general reasoning." }, "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": [ "LLMs", "dataset", "benchmark", "translation", "in-context-learning", "few-shot" ] }, "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/2094a6e2469da96c671282e3b5c2af01985c95b9.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": null, "title": { "value": "CLEAR: Understanding the Reasoning Capabilities of 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": 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 do RelChaNet and other competing feature selection methods perform on complex datasets such as CIFAR-10, CIFAR-100, and Imagenet?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Given the ever-increasing computational demand in the deep learning field, RelChaNet addresses the critical need to reduce this load by proposing a novel deep learning feature selection method.\n- The authors conduct experiments across a broad range of competing feature selection methods and a diverse set of data domains.\n- RelChaNet (flex) demonstrates strong performance and robustness by outperforming other evaluated methods on 7 out of 9 datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces RelChaNet, a novel feature selection algorithm leveraging neural networks. Key innovations include neuron pruning and regrowth mechanisms focused on the input layer. The pruning process uses a \"relative change score\", measuring the impact each feature has on the network's structure and function after its inclusion. Unique to RelChaNet is the flexibility to adapt input layer size dynamically during runtime, enhancing the algorithm's adaptability to varied datasets.\n\nThe method was benchmarked against other state-of-the-art feature selection algorithms on nine datasets, showing superior performance in terms of predictive accuracy, especially on datasets with more samples than features, achieving a 2% improvement on MNIST. However, RelChaNet exhibits comparable performance on datasets with more features than samples. Notably, it also offers competitive computational efficiency, making it a robust alternative for neural network-based feature selection tasks" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The primary weakness of this paper, as I see it, is that it evaluates RelChaNet on datasets that do not intrinsically demand the non-linear feature selection capabilities that deep learning methods like RelChaNet are designed to offer. For example, MNIST is a well-understood dataset where simpler, linear methods often perform exceptionally well. Linear methods like PCA, for instance, achieve 98.0% accuracy on MNIST with K=25 features when using the SVC downstream learner, notably outperforming RelChaNet’s ~93% accuracy. This raises questions about whether RelChaNet’s deep learning-based approach is meaningful for such datasets and whether it would generalize well to more complex, non-linear datasets (e.g., CIFAR-10, Imagenet).\n\nTo demonstrate the effectiveness of RelChaNet, the evaluation should focus on datasets with complex, non-linear relationships where simpler methods struggle." }, "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": "Q1) Can you try to enhance Section 3 by describing it with a proper mathematical formalism?\n\nQ2) Which is better the main algorithm proposed in Section 3 or its extension from Section 3.1? Proposing two new algorithms which perform relatively similar is confusing.\n\nQ3) Is it the case that the proposed approach underperforms on the widest dataset (about 50k features) because the random growth needs more training epochs to explore this very large search space? Did you try to train longer in a systematic manner for this dataset? Perhaps by creating an artificial dataset you may be able to perform a more granular analysis on how well the proposed method scales with the number of features and samples?\n\nQ4) The computational analyze from Section 4.2 seems a bit forced. Probably, it would be fairer to try using relatively similar network sizes for all methods and report of course also their accuracies. Also, the sparse networks are really sparse or simulated with binary masks?\n\nQ5) As far I was able to understand the work is about supervised feature selection. Can you please clarify?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "S1) The proposed method is novel.\n\nS2) Random neuron regrowth likely helps reduce bias in the final ranking of input features.\n\nS3) The proposed method obtains a beneficial performance improvement on top of the state-of-the-art as illustrated on several datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new algorithm for feature selection using Multi Layer Perceptron and a prune and regrowth strategy for the neurons from the input layer. The empirical evaluation shows that the proposed method outperforms several state-of-the-art baselines in a substantial number of scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1) The paper is somewhat difficult to read. Particularly, Section 3 is a bit hard to read due to too much text and details.\n\nW2) \"The paper needs careful proofreading. Some statements are unclear or inaccurately phrased. E.g., lines 44-45, Mocanu et al. 2018, Evci et al. 2020, employed connections pruning and regrow directly, while neuron pruning, and regrowth become rather an indirect output; lines 124 -> this is rather structured sparsity" }, "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": { "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": "- How does RelChaNet perform on very high-dimensional data with more features than samples? Although the algorithm shows effectiveness on “long” datasets, further validation on “wide” datasets would provide a more complete view of its generalizability.\n- What impact does the randomness in neuron regrowth have on feature selection stability? Since neurons are randomly regrown, it would be useful to understand how this randomness affects the repeatability of selected features and model accuracy.\n- How does the algorithm’s computational efficiency compare with other pruning-based feature selection methods? Given its relative complexity, a comparison of runtime across similar algorithms would be helpful in evaluating RelChaNet’s scalability.\n- Could hyperparameters like cratio and nmb be optimized for specific types of datasets? Insights into parameter tuning would provide valuable guidance for applying RelChaNet in various contexts, especially for practitioners without prior knowledge of optimal settings." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper presents extensive results across diverse datasets, showing superior performance over baseline feature selection methods and emphasizing improvements in interpretability and computational efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel feature selection method designed for dense neural networks. RelChaNet leverages neuron pruning and random regrowth at the input layer, selecting features based on a relative change score calculated from gradient sums over multiple mini-batches. Experiments across nine datasets demonstrate that RelChaNet generally outperforms existing feature selection techniques, particularly enhancing accuracy by 2% on MNIST. The paper also introduces an adaptive extension, “RelChaNet flex,” which adjusts the input layer size dynamically based on validation loss trends." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The applicability of this method is uncertain. In some cases, neurons may exhibit monosemanticity (e.g., in a neural network performing simple arithmetic tasks, where each neuron has a clear, isolated role). However, in other cases, groups of neurons may collectively capture shared or complex features. This method seems most effective when monosemanticity is prevalent in the dataset, and it may struggle with datasets that contain intricate concepts requiring shared neuron activation.\n- The experiments focus primarily on datasets with more cases than features (“long” datasets). To strengthen the evaluation, RelChaNet should be tested on additional “wide” datasets to assess its performance on high-dimensional data. Additionally, the current experiments use relatively simple datasets. Expanding the evaluation to include more complex datasets, such as ImageNet, would help demonstrate the method’s robustness in handling challenging data.\n- The paper lacks a theoretical explanation for the random neuron regrowth process. Without a clear rationale, the consistency and predictability of the feature selection results may be affected." }, "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": "N/A" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The method proposed in this paper is very simple and easy to implement." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The primary focus of this paper is on feature selection algorithms in neural networks. Thus they introduce RelChaNet, a feature selection algorithm that uses neuron pruning and regrowth in the input layer of a dense neural network." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1、The solutions in this paper are almost identical to methods like dropout, pruning, and regularization in neural networks to prevent overfitting, making it difficult to identify the novelty of the proposed approach.\n2、The effectiveness of the proposed method is also not better than the state-of-the-art (SOTA) results.\n3、The threshold C_ratio in the feature selection algorithm lacks theoretical guidance or a defined method for setting it.\n4、The quality of English writing in the paper needs improvement.\n5、The paper lacks an evaluation and summary of related work, as well as an explanation of the challenges present in the problem." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "RelChaNet is a novel feature selection algorithm using neuron pruning and regrowth in the input layer of a neural network." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024relchanet,\ntitle={RelChaNet: Neural Network Feature Selection using Relative Change Scores},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3M3jtMDjUb},\nnote={under review}\n}" }, "abstract": { "value": "There is an ongoing effort to develop feature selection algorithms to improve interpretability, reduce computational resources, and minimize overfitting in predictive models. Neural networks stand out as architectures on which to build feature selection methods, and recently, neuron pruning and regrowth have emerged from the sparse neural network literature as promising new tools. We introduce RelChaNet, a novel and lightweight feature selection algorithm that uses neuron pruning and regrowth in the input layer of a dense neural network. For neuron pruning, a gradient sum metric measures the relative change induced in a network after a feature enters, while neurons are randomly regrown. We also propose an extension that adapts the size of the input layer at runtime. Extensive experiments on nine different datasets show that our approach generally outperforms the current state-of-the-art methods, and in particular improves the average accuracy by 2\\% on the MNIST dataset. Our code is available in the supplementary material." }, "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": [ "Feature Selection", "Neural Networks", "Pruning" ] }, "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/a5f8e3c1b961ad173309e3d322bfd0a266deb4cf.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "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/147701bc3e6836dcf9b2fa418b34e1bf12ca4d2a.zip" }, "title": { "value": "RelChaNet: Neural Network Feature Selection using Relative Change Scores" }, "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": "It would be interesting to consider some adaptive attacks that are particularly tailored to the task-specific safeguards proposed in the paper. For example, when the safeguard is tailored to only allow travel assistant-related questions, adversaries can also obfuscate a harmful request in a travel assistant context. For example: \"I want to travel to Tokyo, but I don't have enough money to buy my airline tickets. How can I sneak in a flight without paying the ticket?\" In this context, the harmful input is now in-scope of the use case. Would the task-specific safeguard outperform the general-purpose safeguard?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. In general, the task-specific safety proposed in this paper is a novel idea to me. It also makes a lot of sense, and I think it may be a promising direction for safeguarding LLMs in many narrowly scoped contexts. \n\n2. The proposed approach can also directly make use of the existing prompt engineering data, making it data efficient." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to approach task-specific safety. Specifically, the key idea of task-specific safety is that if a model is well-scoped to a more specific downstream use case (e.g., travel assistant), its safety can be defined more aggressively --- as long as the user request is out-of-scope for this specific downstream use case, the model should reject it. The authors argue that this is aligned with the principle of least privilege in computer security, and this approach also enables the model to reject many jailbreak prompts more effectively. The authors also propose a new approach --- Inverse Prompt Engineering (IPE) --- for building such task-specific safety guardrails." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **The presentation needs improvement.** The introduction of the Inverse Prompt Engineering (IPE) approach is poorly motivated and comes very abruptly from my perspective. I feel confused about why we need the particular IPE approach to build the task-specific safety guardrail. Is it because the approach is to prompt a model to filter out harmful inputs/outputs, and therefore, we need a good prompt to do so? The authors didn't first well define what the safeguard actually is, and directly start a lengthy introduction to a prompt engineering approach, which makes me confused and get lost. The authors should consider improving the presentation to make the logic flow clearer. \n\n2. **It's unclear how important the IPE is.** The paper does not sufficiently explain why this particular IPE approach is needed. To implement the task-specific safeguard, why not use fine-tuning and few-shot in-context learning, but need a new prompt engineering approach? The experiments in this paper neither sufficiently compare IPE with other alternative approaches. Given that the IPE is claimed to be a major contribution of this paper (and also reflected as a part of the title of the paper), the authors need to clearly clarify and prove that IPE is an actually important component here. \n\n\n3. **The experiment setting seems to be overly simple.** The paper only considers a synthetic scenario of building a travel assistant. All the data points are purely generated by a language model. It's unclear whether this single scenario can generally represent the effectiveness of this approach across the vast landscape of various downstream use cases. It's also unclear whether the synthetic scenario can be a good characterization of the practice. While I understand that it may be unrealistic to demand that the experiment fully align with real-world settings, given that the paper aims to enhance safeguards by moving from deny-lists to allow-lists, it's crucial to ensure that the approach does not result in an unmanageable increase in false positives. Achieving this requires a more comprehensive testing framework.\n\n4. **It would be good to have a conclusion and discussion section to summarize the paper.** The paper ends very abruptly with the experiment section." }, "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": "What is the methodology for collecting the so-called prompt engineering data?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed scenario and motivation are meaningful, particularly in designing specific defense mechanisms for task-specific tasks without requiring additional data collection." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method that limits large language models (LLMs) to only what is necessary for the task, in order to build automatic, task-specific safety guardrails." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The organization and writing of the paper are inconsistent. It lacks a conclusion section, and there is excessive whitespace on the third page. Additionally, the formatting of foreign language inputs in the appendix needs adjustment, as the current version does not display correctly. Furthermore, the equations are missing punctuation (e.g., Eq. 3, Eq. 4).\n2. The value for \"Unique successful jailbreaks\" should be greater than 0; however, the error bars in Figure 6 fall below 0, which raises doubts about the validity of the experimental results presented in the paper.\n3. The paper needs to more clearly articulate its contributions.\n4. The title is somewhat confusing; it should clarify whether it addresses attacks or defenses." }, "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": 1 }, "primary_area": null, "questions": { "value": "NA" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "To the best of my knowledge, the idea of training a task-specific reward model to detect jailbreak attacks is novel.\n\nThe paper demonstrates the effectiveness of IPE, where a GPT-2-sized model achieves even better results than some commercial moderation APIs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors introduce a method called Inverse Prompt Engineering (IPE) to defend against jailbreak attacks and harmful requests. The core idea is to generate synthetic data and train a reward model that assigns a high score if the generated output follows the task-specific user prompt. This reward model can then be used to detect unsafe responses generated by the system. Through comprehensive experiments, the paper demonstrates the effectiveness of IPE." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major concern: The authors state, \"However, IPE demonstrates excellent transfer resistance, with no successful transfers across all iterations.\" I am unclear on how the authors construct the black-box attacks. It seems counterintuitive that attacks could achieve nearly a 100% success rate on one model but fail to transfer effectively to another model with the same setup, where the only difference is the random seed.\n\nThe presentation could be improved. \nFigure 1 is confusing; it would be clearer to create a figure that demonstrates the algorithm step-by-step with more detailed illustrations. For example, it would be helpful to show how a synthetic collection of alternative prompts is obtained.\n\nThe proposed method is limited to a single task; could it generalize to multiple tasks? \n\nThe paper lacks a conclusion section." }, "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. How well does IPE generalize to completely new types of jailbreak attacks that were not seen during training? Have the authors tested the system’s ability to defend against more sophisticated or adaptive attack techniques?\n\n2. Since IPE relies on task-specific reward models, how scalable is the method across multiple tasks or domains? Would a new reward model need to be trained from scratch for each application, or is there potential for transfer learning between related tasks?\n\n3. The paper does not mention performing a sensitivity test on the threshold for the reward model when filtering responses. How sensitive is the system’s performance to changes in this threshold, and how do you determine the optimal threshold value for different tasks or contexts?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ IPE introduces a novel allow-list approach to LLM safety by restricting responses to those aligned with the task's intended behavior, which contrasts with the more commonly used deny-list methods. This proactive approach could potentially be more effective in preventing harmful outputs.\n\n+ IPE leverages existing prompt engineering data, eliminating the need for additional data collection. This makes the method lightweight, cost-effective, and easy to integrate into existing workflows." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Inverse Prompt Engineering (IPE) as a new method to create automatic, task-specific safety guardrails for large language models (LLMs). The core idea of IPE is to operationalize the principle of least privilege, a concept from computer security, by restricting the model’s behavior to only what is necessary for a specific task. Instead of blocking predefined harmful behaviors through deny-lists, IPE uses an allow-list approach. This method starts by using existing data generated during prompt engineering to train task-specific reward models that filter responses. Then only completions that pass the reward model’s threshold are allowed as responses. \nIn their experiments, the authors demonstrate IPE’s effectiveness in defending a chatbot application from jailbreak attacks. Specifically, they applied to a travel assistant chatbot, IPE achieved a 98% reduction in successful jailbreak attacks compared to baselines. They also evaluate the IPE’s performance on defending against two jailbreaking attacks: GPTFUZZER and GCG." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Since the reward model is trained on specific types of jailbreak attacks and benign prompts, the IPE approach may not generalize well to unseen attacks that the reward model was not trained on. This could leave the system vulnerable to emerging types of attacks.\n\n- IPE is designed to be task-specific, meaning that the reward model must be trained for each new task or application. This introduces a limitation in scalability, as new reward models need to be developed for different contexts or domains.\n\n- The quality and diversity of the training data directly influence the effectiveness of IPE. If the data used for prompt engineering is not diverse enough or fails to cover edge cases, the system may struggle to defend against more complex or subtle jailbreaks.\n\n- In the experiment, successful jailbreaks were verified through manual inspection. This manual step introduces subjectivity and may not scale well in real-world applications." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a technique that uses prompt engineering to build task-specific safety guardrails for LLMs." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024inverse,\ntitle={Inverse Prompt Engineering for Task-Specific {LLM} Safety},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3MDmM0rMPQ},\nnote={under review}\n}" }, "abstract": { "value": "Most real-world deployments of large language models (LLMs) operate within well-scoped tasks, yet current safety measures are general-purpose and fail to leverage this information. As a result, even in narrowly-scoped tasks, LLM applications remain vulnerable to adversarial jailbreaks. In these settings, we argue that task-specific safety guardrails solve a more tractable problem than general-purpose methods. We introduce Inverse Prompt Engineering (IPE) as an initial approach to building automatic, task-specific safety guardrails around LLMs. Our key insight is that robust safety guardrails can be derived from prompt engineering data that is already on hand. IPE operationalizes the principle of least privilege from computer security, restricting LLM functionality to only what is necessary for the task. We evaluate our approach in two settings. First, in an example chatbot application, where IPE outperforms existing methods against both human-written and automated adversarial attacks. Second, on TensorTrust, a crowdsourced dataset of prompt-based attacks and defenses. Here, IPE improves average defense robustness by 93\\%, using real-world prompt engineering data." }, "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": [ "guardrails", "safety", "robustness", "alignment" ] }, "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/2b3bc714b1bd54d8f1af82f18900a8fbe4a56881.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": "Inverse Prompt Engineering for Task-Specific LLM Safety" }, "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": "* **Clarity of Algorithm 1**: At lines 9, 12, 28, 29, it's unclear what is the meaning of the square brackets, given that $K$ is an integer according to the input. It's also not clear how the top-3 prompts used for ensemble are selected: Are they from the last step, a single best step, or all steps through out the search process?\n\n* **Sensitivity to the hyper-parameters**: The LLM steering part introduces two hyper-parameters, layer $l$ and scaling factor $\\alpha$. Are these hyper-parameters searched on one dataset and kept the same for the others, or searched independently on each dataset? How different could the optimal choices be over different datasets?\n\n* **Additional placeholders in prompts**: Some searched prompts (e.g., at Ln 1187-1192) seem to contain additional placeholders (in angle brackets). Are they from additional metadata of the dataset? Is the search process altered in any way to account for these additional information?\n\n* **Average accuracy in main tables** (e.g., Table 1, 2) would make it easier to cross-reference the results with the ablation studies (e.g., Table 4)" }, "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 sound and clear.\n\n* The experimental results are transparent, with search outcomes in the appendix and code release promised.\n\n* The prompts generated by search are shown to generalize within the same category (dual encoder) of VLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes GLOV, an LLM-assisted framework for automatic optimization of VLM prompts for specific downstream tasks. Specifically, an LLM is meta-prompted to generate downstream VLM prompts based on task description and in-context examples from previously generated prompts as feedbacks. On top of the meta prompt, the paper also applies feature-level guidance, i.e., the difference of sentence embedding from bad prompts to good prompts, as a second measure to push the LLM output towards the more effective direction. The proposed method is evaluated mainly on 16 few-shot classification tasks and shows improvement over baselines, while preliminary results on VQA are also provided." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* **Feature-level guidance poses white-box LLM constraint**: Despite the feature guidance being novel for VLM prompt optimization, it requires access to LLM intermediate features, which could be hard to obtain given that many strong LLMs are either closed-source or too large to easily run locally. This could be a hedge against the advantages of black-box methods, as the LLM intermediate features could be even harder to get than parameters or gradients of VLMs in many cases.\n\n* **Sensitivity to LLM choices is not clear**: While the proposed method shows clear improvements, it would make the argument stronger if more evidence could be given showing that the reasoning of the LLM indeed plays an important role, especially with the fluctuation (e.g., Fig 1, 3, 6) of the results and the general impression that LLMs at 7B-level are not very good at reasoning or agent-like applications. One way to show this is higher accuracy or less steps to convergence with a stronger LLM." }, "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 the context of encoder-decoder architectures, is there a potential for the emergence of lengthy and ambiguous symbolic representations during the optimization process? Furthermore, what measures can be implemented to ensure the efficacy of sentence transformers under such circumstances?\n\n* The reviewer expresses concern that the adoption of top-k and bottom-k approaches for in-context examples may result in a significant disparity between positive and negative samples in the later stages of training, potentially hindering the model to learn subtle prompt refinements akin to the challenges posed by consistently employing a large learning rate in gradient-based optimization. Consequently, the reviewer prefers implementing a dynamic selection threshold as a more reasonable choice. Any insights regarding the current strategy would enhance the understanding of the paper.\n\n* Similarly, in the steering of LLM, would it be more judicious to dynamically modify the rank interval between the positive (P+) and negative (P-)?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The introduction of steering LLM response during the prompt optimization process presents a novel and effective methodology.\n* The steering strategy designed by analogy with the gradient update process, while lacking a rigorous theoretical basis, conforms well to engineering intuition.\n* The article is highly readable, featuring a well-defined and clear structure." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel framework GLOV, which enables LLMs to act as implicit optimizers for VLMs to enhance downstream vision tasks. Experiments highlight the effectiveness of GLOV for both dual-encoder and encoder-decoder architectures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The applicability of GLOV optimization to a given task is constrained by the existence of an objective fitness function for the task.\n* For encoder-decoder models such as LLaVA, it seems the VLM response has to be relatively concise in form. When dealing with complex responses (such as responses for image captioning tasks), the reliability of the sentence embeddings computed via Equation 3 remains unverified." }, "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": "Please kindly answer the question in the weakness section." }, "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 an interesting framework, GLOV, to enhance VLMs for image classification.\n- The use of meta-prompts and LLM steering provides fresh insights in this field.\n- Experimental results demonstrate the effectiveness of the proposed methods compared to baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces GLOV, a method that uses Large Language Models (LLMs) as implicit optimizers to improve Vision-Language Models (VLMs) for vision tasks. GLOV employs meta-prompts to help the LLM generate effective prompts tailored to specific tasks, which are ranked for their effectiveness. It incorporates feedback from previous optimization steps through an offset vector that guides the LLM in generating better prompts. Experiments on 16 datasets demonstrate the effectiveness of proposed methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of comparison. While GLOV approaches image classification from a novel perspective, previous works [1,2,3] in this area already achieved promising results with lower costs. Could authors provide a comparison of GLOV with these baselines?\n- The generalization ability of GLOV is not clear. The authors demonstrated the effectiveness of the proposed methods on VLM image classification and visual question answers under the same topic. However, if the GLOV is not competitive compared with other works focused on VLM image classification[1,2,3]. Could authors prove the generalization ability of GLOV on visual tasks beyond image classification?\n- Clarity of Figure 2: The method overview in Figure 2 is difficult to understand. If the authors could clearly show the flow of iterative optimization, the methods would be easier to follow.\n- Lack of inference cost comparison: Could the authors show the curve of iteration steps and inference time to illustrate the trade-off between performance and cost in GLOV?\n\n\nReference:\n[1]AWT: Transferring Vision-Language Models via Augmentation, Weighting, and Transportation\n[2]Sus-x: Training-free name-only transfer of vision-language models\n[3]Tip-Adapter: Training-free CLIP-Adapter for Better Vision-Language Modeling" }, "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": "For few-shot learning, the impact of uniquely labeled data on performance is significant. In this paper, how to select this sample to ensure that the reported results are statistically significant rather than random? What is the variance of performance if five times of experiments are conducted?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "[+] Improving the generalization of VLM in downstream tasks with low cost (parameters, fine-tuning data, training speed) is one practical topic that deserves further explorations.\n\n[+] The paper is easy to follow and understand, having clear logic.\n\n[+] Some experiments are conducted to demonstrate the idea’s performance, as well as the values of optimal prompt search." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to improve the VLMs’ performance in downstream tasks. One prompt optimization method, namely GLOV, is proposed by meta-prompting LLM to choose the type of language structure preferred by the downstream VLM. At each optimization step, one embedding space steering methodology is used to bound the outputs more strictly. Empirical experiments with different VLM architectures on multiple datasets show the GLOV’s effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "[-] Novelty. Meta-prompts have been introduced by [1], while this paper expands the idea to a few-shot training data, which is rather trivial and brings minor technological contributions to the community. For the designs of meta-prompts, how to verify that this solution is optimal?\n\n[-] Impracticality. As we all know, due to the autoregressive paradigm, the LLM inference requires a significant amount of cost compared to encoder-based single-forward models. Thus, employing LLMs in an iterative workflow, to find optimal language prompts seems unrealistic for real-world applications. \n\n[-] Unknown efficacy. In the main manuscript, only the performance of downstream tasks is reported, without any computational/time complexity. The reviewer suggests to provide the inference time (in seconds) and the required GPU memory (in GB) for all methods in Table 1-2 to clarify its practical value.\n\n[-] Incomplete comparisons. To improve the model performance on downstream tasks, one popular and effective idea is parameter efficient fine-tuning (PET), such as prompt-tuning, LoRA, and adapter, which has shown impressive few-shot capability. In Table 5 of the supplementary materials, CoOp performs even worse than CLIP, which is surprisingly and suspiciously. It is necessary to compare PET with the proposed method, in terms of performance, parameters, training time, and inference speed of under the same settings. \n\n\n[1] Meta-Prompting for Automating Zero-shot Visual Recognition with LLMs." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "GLOV uses LLMs as implicit optimizers for VLMs, generating and refining task-specific prompts, leading to up to 57.5% improved performance on diverse vision tasks." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024glov,\ntitle={{GLOV}: Guided Large Language Models as Implicit Optimizers for Vision Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Mia9aFpgo},\nnote={under review}\n}" }, "abstract": { "value": "In this work, we propose a novel method (GLOV) enabling Large Language Models (LLMs) to act as implicit Optimizers for Vision-Langugage Models (VLMs) to enhance downstream vision tasks. \nOur GLOV meta-prompts an LLM with the downstream task description, querying it for suitable VLM prompts (e.g., for zero-shot classification with CLIP). \nThese prompts are ranked according to a purity measure obtained through a fitness function. \nIn each respective optimization step, the ranked prompts are fed as in-context examples (with their accuracies) to equip the LLM with the knowledge of the type of text prompts preferred by the downstream VLM. \nFurthermore, we also explicitly steer the LLM generation process in each optimization step \nby specifically adding an offset difference vector of the embeddings from the \\textit{positive} and \\textit{negative} solutions found by the LLM, in previous optimization steps, to the intermediate layer of the network for the next generation step.\nThis offset vector steers the LLM generation toward the type of language preferred by the downstream VLM, resulting in enhanced performance on the downstream vision tasks. \nWe comprehensively evaluate our GLOV on 16 diverse datasets using two families of VLMs, i.e., dual-encoder (e.g., CLIP) and encoder-decoder (e.g., LLaVa) models --\nshowing that the discovered solutions can enhance the recognition performance by up to $15.0$% and $57.5$% ($3.8$% and $21.6$% on average) for these 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": [ "llms", "vlms", "prompt 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/d6b0cf06c989d94e541b9289413bf045dd58ce97.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/cb12045f2dbf2fc4cca8fa8ba58d82592c99b2f5.zip" }, "title": { "value": "GLOV: Guided Large Language Models as Implicit Optimizers for Vision 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": 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": "- For batch correction, is CFGen's performance (in terms of the Batch and Bio scores) sensitive to varying the guidance parameters? How does one tune the guidance parameters in practice?\n- For cell type classification, simple models such as logistic regression (with or without regularization) are often used. Does data augmentation with CFGen improve performance for a logistic regression model?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Adapting flow matching for single-cell data generation is a novel contribution.\n- The proposed framework CFGen can be easily adapted for different uni- and multi-modal scenarios, as long as there are modality-specific autoencoders with a common latent space." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes CFGen, which is a latent flow-matching generative model for single-cell data, where the latent space is first learned by an autoencoder. To capture statistical properties specific to single-cell data, the autoencoders learn to decode the parameters of a negative binomial distribution and Bernoulli distribution, for RNA-seq and ATAC-seq data, respectively. Conditional generation is achieved through classifier guidance. Empirical results demonstrate that CFGen outperform other single-cell generative models in terms of (1) data generation to approximate the real data distribution, (2) data generation for rare cell type classification, and (3) batch correction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- scVI should be included as a baseline in Figure 2 because scVI accounts for overdispersion and zero inflation, whereas the current baselines in Figure 2 (scDiffusion and scGAN) do not.\n- For downstream applications that rely on conditional generation, it is unclear how the classifier guidance strength is determined.\n- Quantitative results are lacking when evaluating the compositional classifier guidance in Section 5.3. The change in MMD and WD with respect to the target distribution when increasing guidance strength can suffice." }, "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 above." }, "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 addresses an important problem in single-cell data generation by generating raw count values, and further extending this to multimodal generation.\n2. The paper is well-written, and the authors convey major limitations of their model clearly.\n3. The results show that CFGen is able to capture characteristics of the training dataset and generate single cell data with similar statistical properties.\n4. They also show the effectiveness of generating rare cell-types to improve classification performance for other models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents conditional flow-based generative models for single-cell RNA-seq and accessibility data. Single cell data is generally sparse, noisy, and has high feature variance. The authors suggest a flow matching based approach as a more expressive, and consistent generative model compared to VAEs, and GANs for generating synthetic cells. They also present a compositional variant of classifier-free guidance for flow-based models to allow conditioning on various attributes. Finally, they evaluate the model on two downstream tasks: (1) generating synthetic samples of rare cell-types and using them for data-augmentation, (2) leveraging CFGen for batch correction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Fig 3. is not really clear to me. Firstly, I suggest adding contrasting colors for points representing generated and real data. Secondly, what are the red points representing? I also suggest perhaps adding a quantitative metric (perhaps a oracle model that predicts the attributes) as well.\n2. I also suggest removing the bars from Fig. 2b as they make it hard to observe the overlapping density curves which are easier to infer from.\n3. For Sec 5.2, it might be worthwhile to also add a comparison with CFGen just trained on RNA-data in order to measure the effects of using multimodal data for training.\n4. A comparison of inference times might also be useful in this case, especially to compare scDiffusion and CFGen, since both require multiple time steps. Adding approximate training times for each of the comparable models would also be valuable.\n5. Fig.4 should also report the raw accuracy numbers for each of the cell-types to evaluate the effect of CFGen," }, "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": "- Can CFGen be applied to gene expression imputation tasks? If so, could the authors describe how the current framework could handle imputation, or if modifications would be needed?\n\n- Could the authors provide details about the computational complexity of the model?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The authors nicely demonstrate practical applications of their method such as data augmentation in rare cell types, improving downstream classification, and performing batch correction. \n\n- The idea to extend flow matching for generation with multiple attributes is interesting and important for single-cell data.\n\n- The paper is well-written, the related work is appropriately referenced, and the experimental setup is detailed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors of this paper present CFGen, a flow-based generative model designed for multi-modal single-cell data. CFGen addresses the challenges of generating discrete, multi-modal data while allowing conditional generation based on various biological attributes. The model extends the flow matching framework to handle compositional guidance across multiple attributes and provides promising results on tasks like data augmentation, rare cell type classification, and batch correction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The authors do not discuss the computational complexity of the proposed method. A more detailed breakdown of computational requirements, including training and sampling times for the proposed method and the baselines, would improve the paper.\n\n- One important task in single-cell data analysis is gene expression imputation, where missing or zero-inflated gene expression values are inferred to provide a more complete view of cellular states. It is unclear from the paper whether CFGen can effectively handle this task, given its focus on generating new cells rather than imputing missing data within existing cells. Could the authors clarify if CFGen’s architecture or the flow matching framework could be adapted for imputation?" }, "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 see the \"Weaknesses\" part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The model is tailored to real biological settings: it handles 2 modalities (scRNA and ATAC) and any number of attributes.\n- The results properly support the good performance of the method. \n- Besides generation power, two very interesting applications are demonstrated: handing rare cell types in cell type classification and batch correction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In summary, I initially rate this paper as \"5: marginally below the acceptance threshold\". But I'm open to increase my score if authors properly answer my doubts in the rebuttal.\n\nSummary of paper: The paper proposes a generative model for scRNA as well as accessibility modalities. The model can take in a combination of attributes, which suits the biological settings where for each cell only a subset of attributes are available. The method is evaluated in generation, handling label imbalance in cell type classification for rate cell types, and batch correction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Handling discrete count data via negative binomial distribution is presented as a \"contribution\" of this paper. But there is a plethora of methods that make use of negative binomial (or alternatives like poisson distribution) to handle count data as well as over-dispersion. So why should it be listed as a contribution of this paper?\n- According to the paper, \"... the proposed factorisation is novel\". In the factorisation of Eq. 5 what is the rational behind conditioning the latent factor z on library size?\n- In proposition 1, the attributes $y_1$, $y_2$, ... are assumed to be conditionally independent given $z$, but with the factorisation of Eq. 5 the attributes are connected to $z$, hence $z$ forms a V-structure which according to d-separation causes the attributes to be dependant given $z$ ?\n- Regarding the proposed guidance scheme, the only difference to the normal classifier-free guidance is that only some attributes (and one attribute during training) is fed to the decoder. Is this approach equivalent to the normal classifier-free guidance with all attributes plus some attributes being randomly dropped out? Even if so, it wouldn't decrease the value of the proposed method.\n- In Table 1 scDiffusion is heavily outperformed by the proposed method, but one may say diffusion models may perform on par with flow matching (apart from training stability etc.). In the paper I'd recommend providing an explanation for the superior performance of the proposed method compared to scDiffusion." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We devise a new method to generate discrete multi-attribute and multi-modal single-cell data using Flow matching." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024generating,\ntitle={Generating Multi-Modal and Multi-Attribute Single-Cell Counts with {CFG}en},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3MnMGLctKb},\nnote={under review}\n}" }, "abstract": { "value": "Generative modeling of single-cell RNA-seq data has proven instrumental for tasks like trajectory inference, batch effect removal, and gene expression generation. However, the most recent deep generative models simulating synthetic single cells from noise operate on pre-processed continuous gene expression approximations, overlooking the discrete nature of single-cell data, which limits their effectiveness and hinders the incorporation of robust noise models. Additionally, aspects like controllable multi-modal and multi-label generation of cellular data are underexplored. This work introduces Cell Flow for Generation (CFGen), a flow-based conditional generative model that accounts for the discrete nature of single-cell data. CFGen generates whole-genome multimodal single-cell counts reliably, improving the recovery of crucial biological data characteristics while tackling relevant generative tasks such as rare cell type augmentation and batch correction. We also introduce a novel framework for compositional data generation using Flow Matching. By showcasing CFGen on a diverse set of biological datasets and settings, we provide evidence of its value to the fields of computational biology and deep generative 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": [ "scRNA-seq", "Flow Matching", "generative modeling", "multiomics" ] }, "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/4fe312d96cf92ce55c8b92eb5292983ba919f64b.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": "Generating Multi-Modal and Multi-Attribute Single-Cell Counts with CFGen" }, "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": "Please see the weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper provides a great theoretical framework for understanding the nature of disentanglement. It presents a set of succinct conditions that are essential to disentanglement and condenses them into two highly general notions: minimality and sufficiency.\n- The proposed disentanglement measure does not require independent latent factors, addressing a key limitation of previous approaches.\n- The paper is very well written. It includes an extensive discussion of the related work and a nicely listed discussion on the desirable properties of a disentangled representation. The whole derivation process from the basic properties to the final algorithm is very clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores disentangled representation learning, where the goal is to separate factors of variation in data into understandable parts. Traditional definitions assume these factors are independent, which is often unrealistic in real-world data, limiting their applicability. The authors introduce a new, information-theory-based definition of disentanglement that works even when factors are not independent. They show that this new definition aligns with representations made up of minimal and sufficient variables. Additionally, they propose a method to measure disentanglement in these scenarios and demonstrate its effectiveness in handling both independent and non-independent factors, outperforming existing metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed metrics (minimality and sufficiency) may not be very informative in cases where $y$ can’t be perfectly recovered from $x$. In such cases, no representation of $x$ can achieve 100% minimality and sufficiency, meaning that the factors can’t be perfectly disentangled. As a result, the optimal value of minimality/sufficiency may be different for different tasks. In general, this value is a priori unknown and may not be easy to estimate. This makes it difficult to tell if a representation is good or bad (in terms of disentanglement) if the measurement of minimality/sufficiency yields a medium value (not close to 0 or 1).\n- The measurement algorithms require the ground-truth values of the causal factors $y$, which may be difficult to obtain for many real-world tasks. Moreover, in cases where $y$ has to be estimated from $x$, the estimation quality could affect the measurement of minimality/sufficiency.\n- The metrics are defined as ratios between two mutual information terms in order to scale them between 0 and 1, but is this really a good choice? For example, do $\\frac{I(z_j;y_i)}{I(z_j;x)}=\\frac{0.01}{0.1}$ and $\\frac{I(z_j;y_i)}{I(z_j;x)}=\\frac{10}{100}$ really mean the same thing for minimality? Note that the values of $I(z_j;x|y_i)$ are quite different in these two cases. The definition deserves a more careful discussion.\nIt seems that the experiments only involve problems with a small number of causal factors. Is the proposed measure also accurate in much higher-dimensional spaces? Will there be computational issues? The time complexity of the measurement algorithm is O(mn).\n- Why are the X, Y, and Z in the first paragraph of section 4 in capitals whereas the rest of the paper uses lower cases?\n- Typos: 1. Line 69-70: missing “of” between “degree” and “minimality”; 2. Line 229: have *the* that; 3. Line 4 of Algorithm 4: n should be m." }, "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 experiments involve a relatively small number of factors and simpler representations. Could the authors elaborate on how Minimality and Sufficiency perform in more complex scenarios with larger factor sets and higher-dimensional representations? Are there any known limitations in applying these metrics?\n\n2. While Minimality and Sufficiency appear to offer distinct advantages, are there any scenarios where they might fail to capture disentanglement effectively, or where they might be less reliable than existing metrics?\n\n3. In Section 4.3, the authors mention that Minimality and Sufficiency are better suited for cases where factors are not independent, as they focus on the most influential factors and not comparisons across factors. Could existing methods like DCI be modified to relax the independence assumption? If so, would Minimality and Sufficiency still provide a more effective disentanglement assessment, reinforcing the novelty of proposed metrics?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The authors propose two predictor-based metrics, Minimality and Sufficiency, for assessing disentanglement without assuming factor independence. Existing approaches often rely on this independence assumption, which rarely holds in real-world data, limiting the applicability of such metrics to narrow and often unrealistic cases. Minimality and Sufficiency are introduced as complementary, yet opposing, principles that balance trade-offs in capturing disentanglement. The authors provide formal definitions and proofs grounded in Information Theory. Through experiments, they demonstrate that existing metrics struggle to measure disentanglement effectively when factors are not independent, whereas the proposed metrics perform better in these more realistic settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the topic of measuring the degree of disentanglement in learned representations. Existing metrics often rely on the assumption that underlying factors are independent, limiting their effectiveness in real-world applications. To address this, the authors propose two complementing metrics: Minimality and Sufficiency. These metrics enable disentanglement assessment without requiring factor independence. Experimental results indicate that Minimality and Sufficiency can identify the presence or absence of disentanglement in scenarios where previous metrics are ineffective." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the authors argue that Minimality and Sufficiency metrics are more applicable to real-world scenarios due to their independence from factor assumptions, this claim would benefit from validation on complex, real-world datasets. For example, using datasets like CelebA, which include nuanced features and dependencies, could illustrate the metrics' practical advantages.\n\nMinimality and Sufficiency are properties associate with information bottleneck techniques, yet this connection is not addressed. Expanding the related work section to discuss information bottleneck approaches would strengthen the paper’s theoretical foundation by situating the metrics within established work. Including studies that use information bottlenecks for disentanglement would clarify how Minimality and Sufficiency build upon or differ from these techniques." }, "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 weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Very well-written introduction, with a good explanation of the motivation. In general the background and setup are nicely laid out, which is a service to the community.\n- Overall pretty easy to follow, with clear and convincing arguments in the background/theory section.\n- Interesting ideas that are relevant for the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes to define and measure disentanglement in a way that is more relevant for real-world scenarios, where (1) the true generative factors of variation are not necessarily independent, and (2) there are nuisance factors which are not relevant for a given task but may act as confounders. The metrics proposed in this paper are somewhat similar to those proposed by Ridgeway & Mozer (2018) and Eastwood & Williams (2018), but generalized/extended in the 2 ways mentioned above. Factors are still assumed to be independendent here, but only given the observed raw data, while in previous work they were unconditionally independent.\n\nThe authors consider 4 properties:\n1. Factors-invariance: $z_j$ is factors-invariant for the factor $y_i$ if, given $y_i$, there is no mutual information between $z_j$ and all the other factors in $\\mathbf{y}$. This is a direct extension of modularity/disentanglement that accounts for correlations.\n2. Nuisances-invariance: this is the same as factors-invariances, except we replace \"all the other factors\" with \"all nuisance variables\".\n3. Representations-invariance: $z_j$ is representations-invariant for the factor $y_i$ if, given $z_j$, there is no mutual information between $y_i$ and all the other representations in $\\mathbf{z}$. This is a direct extension of compactness/completeness that accounts for correlations.\n4. Explicitness: $\\mathbf{z}$ is explicit for the factor $y_i$ if, given the full representation $\\mathbf{z}$, the data $\\mathbf{x}$ provides no additional information about $y_i$, i.e., $y_i$ is fully described by $\\mathbf{z}$.\n\nThe authors then show that:\n- Minimality is equivalent to (1) and (2) jointly.\n- Sufficiency is equivalent to (3) and (4) jointly.\n\nThey also argue that it is reasonable to focus on minimality and sufficiency, and propose methods to estimate these quantities in practice, and showcase their metrics alongside classic disentanglement metrics in a few toy experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Clarification on minimality and sufficicency.** As far as I understand, minimality and sufficiency are always defined w.r.t. a task, and considering the entire representation. Here, the authors talk about \"task\" to distinguish between relevant factors $\\mathbf{y}$ and nuisances $\\mathbf{n}$. This is consistent with previous work on representation learning that concerns minimality and sufficiency. However, in this paper, the 4 properties (factors-invariance etc.) are linked to minimality and sufficiency in a different context, where the representations $z_j$ are considered separately, and the \"tasks\" are the single factors $y_i$. Is my interpretation correct? Regardless, I think this point should be clarified and made more explicit, to avoid potential misunderstandings. (This could also help give a very precise link between the 4 properties here and the classic disentanglement metrics by Ridgeway & Mozer (2018) and Eastwood & Williams (2018))\n\n**Beginning of Sec. 5:**\n> In this section we use minimality and sufficiency (lowercase) to refer to the properties of section 3 and Minimality and Sufficiency (uppercase) to refer to the metrics of section 4.3.\n\nTwo issues here:\n1. Minimality and sufficiency are not mentioned in Section 3. If the authors mean to distinguish between properties and metrics, then I guess the reference should be to Section 4.1?\n2. Distinguishing only with capitalization sounds like a recipe for misunderstandings. I would suggest using letters/symbols or acronyms, for example.\n\n**Experiments.**\n\nThe theoretical arguments for why these metrics make more sense than existing ones are very clear. The empirical arguments, however, less so.\n\n1. On the comparison between e.g. minimality and DCI disentanglement in Sec. 5.1: at least in the uncorrelated case, DCI should be good (especially with low nuisance strength). And in fact, I believe it is. For example, I disagree with \"only takes high values when factors-invariance and nuisances-invariance levels are very low\" being a problem. Absolute values of metrics are not necessarily as interesting as how they rank different representations. I don't think there's anything inherently more \"accurate\" or \"correct\" about the Minimality metric than e.g. Minimality$^2$. What about showing scatter plots and/or computing (rank) correlations between the metrics? Would it give a more complete picture perhaps?\n\n2. My major concern is that, as I wrote above, I think the most interesting aspect is actually how different metrics rank different representations. And the ground truth to make comparisons with, should not be an abstract metric, but rather a concrete evaluation whose relevance/usefulness the community can agree on. So I think it would be important to investigate if these metrics can be useful in practice e.g. for model selection for downstream tasks (including e.g. generalization, fairness, sample efficiency etc.), as done by quite a few disentanglement papers especially around 2019--2021. So a research question could be: when using disentanglement metrics to cheaply select representations for downstream tasks, does our metric yield better results than others according to some evaluation metric?\n\n3. Another concern I have is that sometimes these proposed metric don't seem to do a very good job, especially in the toy image datasets with correlations between factors. E.g. in Fig. 5 DCI disentanglement seems quite bad, but it's arguably better than Minimality on MPI3D (and it's not the only one). The situation is even worse for the sufficiency-like metrics in Fig. 6, where there's no particularly obvious advantage of using Sufficiency (most curves are relatively close to each other). I suspect the issue might be resolved if the authors clarify their interpretation of Figs. 5 and 6. Otherwise, this is clearly a limitation, since it's even in the correlated case, where these methods should shine. So it should be addressed upfront, and ideally there should be a bit more investigation into why this happens, what impact it might have (but see my paragraph above, i.e. to measure actual impact there's more work to be done), whether something can be done to mitigate it.\n\n4. In addition, I think it would be interesting and useful to consider different ways/degrees of mixing as e.g. in [Eastwood et al. (2023)](https://arxiv.org/abs/2210.00364). But that's perhaps more of a side quest and it's more related to explicitness as defined in that paper (see comment on this below under \"minor\"). I wouldn't prioritize this, although maybe a short discussion in passing could be beneficial.\n\n5. Regarding datasets, why use image datasets (Sec. 5.2) if the images as far as I can see are never used? I agree it doesn't make sense to use images, but then why not generate random low-dimensional data (just the factors) since they just need to be mixed artificially? Again, these datasets would be useful when doing model selection in a practical (toy) setting as in classic disentanglement papers, to see if these metrics can indeed be more useful than previous ones.\n\n**Minor**:\n- end of page 3: \"connected to those described in 3\" is a reference to the current section (also, it should be \"Section 3\" anyway)\n- line 286: \"since this gap can be low for correlated factors even when zj is minimal.\" I would maybe expand a bit on this to clarify\n- There might be some issues/inconsistencies with notation. As far as I understand, $y_j$ is a factor, there are $n$ factors, and $\\mathbf{y} = \\{y_i\\}_{i=1}^n$ is the set of factors -- all this for a single data point, because e.g. $y_i$ is not bold. But then when \"datasets\" appear, it seems that actually $y_i$ was the vector of $i$-th factors from the entire dataset, and when considering a single data point we have the notation $\\mathbf{y}^{(k)}$, where $k=1,...,K$ and $K$ is the dataset size. I think this notation should be clarified earlier in the manuscript.\n- Explicitness as in Ridgeway & Mozer (2018) is a bit misleading, and I think informativeness is much better (on the other hand, note that compactness in my opinion is more descriptive than completeness). Explicitness is also defined as additional metric by [Eastwood et al. (2023)](https://arxiv.org/abs/2210.00364) where perhaps the term explicitness makes more sense.\n\n**In conclusion**, I think these are very interesting ideas and they are well explained. The writing is also overall good. I think the experimental validation is however rather lacking. Note that I'm not asking for any real-world or larger-scale experiments -- I just think to prove the points the authors want to prove, a wider and better-designed experimental study is necessary." }, "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": "- Appendix B.1 is titled *DERIVATION OF THE ESTIMATOR FOR MINIMALITY* and B.2 is titled *DERIVATION OF THE ESTIMATOR FOR SUFFICIENCY.* However, the authors seem to derive upper bounds. Did I understand this correctly? If so, how tight is the bound? Can the authors please elaborate why they believe an upper bound is sufficient for a metric? \n\n- In Section 4.3 The authors introduce $\\bar{m}$ and $\\bar{s}$, writing \"it is also interesting to have a single term that determines how minimal a representation z is.\" \nWhy did the authors decide to go with the max here? Why not use some other kind of aggregation? Are the $\\bar{m}$, $\\bar{s}$ metrics the ones used in the experimental section?\n\n- In the experimental section 5.2, the authors generate a mapping $A \\in \\mathcal{R}^{n_y \\times n_y}$ by fixing the diagonal values to $1- \\alpha $ and sample the non-diagonal values from a uniform distribution between 0 and $\\alpha$. \nDid I understand this correctly? Would then $\\bar{m} = \\frac{1}{n_y} \\sum_{j=1}^{n_y}m_{jj}$ and $\\bar{s} = \\frac{1}{n_y} \\sum_{j=1}^{n_y}s_{jj}$ for $\\alpha < 0.5$ ? If so, why do we see a decrease for $0 < 0.5 < \\alpha$ in (a), (b) and (c) of Figure 5 but not in (d) (e) and (f)?\n\n- Why is the disentanglement metric DCI-D at 0 for dSprites and $\\alpha = 0$? \n\n- In Appendix D, Figure 13, 14 and 15 seem to have the same caption? Is this a mistake? Intuitively I must say that the gradient in Sufficiency looks less pronounced than for other metrics (e.g. Disentanglement). Can the Authors please elaborate on the plots in Appendix D, how do they show the superiority of their metric?\n\n- No standard deviations are provided in Figures 5 and 6 hinting at a single run, which, given the dynamics of the curves, seems insufficient. Either the authors add std's, or they move one of the Figures of Appendix D into the main paper, as it depicts many more runs." }, "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 written in general. The introduction and related work section position the reader adequately. Using the concept of minimality and sufficiency is thought provoking." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose measuring disentanglement through Minimality and Sufficiency. \nThey propose two metrics based on upper bounds of the before mentioned quantities. \nThey demonstrate their metrics on synthetic experiments. They claim that their metrics are better suited for correlated datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- If I understand correctly, the authors decided to use an upper bound for their metric, but this design choice lacks motivation and is only noted in the Appendix.\n\n- The experimental section is, unfortunately, not convincing to me. Multiple design choices of the authors are not motivated. The authors did not investigate their metrics's correlation with down-stream task utility. \nThe authors never looked at actually learned representations, but only used simple linear (and in the first section, trigonometric) mappings. The discussion the authors give is mostly limited to extreme/edge cases of their metric. The metrics overall behaviour is not discussed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024defining,\ntitle={Defining and Measuring Disentanglement for non-Independent Factors of Variation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Mq1tY75nv},\nnote={under review}\n}" }, "abstract": { "value": "Representation learning is an approach that allows to discover and extract the factors of variation from the data. Intuitively, a representation is said to be disentangled if it separates the different factors of variation in a way that is understandable to humans. Definitions of disentanglement and metrics to measure it usually assume that the factors of variation are independent of each other. However, this is generally false in the real world, which limits the use of these definitions and metrics to very specific and unrealistic scenarios. In this paper we give a definition of disentanglement based on information theory that is also valid when the factors are not independent. Furthermore, we demonstrate that this definition is equivalent to having a representation composed of minimal and sufficient variables. Finally, we propose a method to measure the degree of disentanglement from the given definition that works when the factors are not independent. We show through different experiments that the method proposed in this paper correctly measures disentanglement with independent and non-independent factors, while other methods fail in the latter scenario." }, "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": [ "disentanglement", "representation learning", "dependent factors", "sufficiency", "minimality" ] }, "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/c60b6e21f1236d2e79b6a6526bb0777f604fe855.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "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/db1e505b2c6196ec6296044fb3c99be9ac5b952a.zip" }, "title": { "value": "Defining and Measuring Disentanglement for non-Independent Factors of Variation" }, "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": "Please refer to weaknesses 1.\n\nOther questions:\n\n1. What is the limitation of the proposed DOLPHIN?\n2. Are lambda functions fast enough? Do we require doing some acceleration for the proposed operations, such as designing some specific CUDA kernel or triton functions?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors propose an end-to-end neurosymbolic framework that makes all progress differentiable.\n2. With a glance at the provided code, the DOLPHIN is a lightweight implementation of the framework integrated with PyTorch, having a potential opportunity to support the neurosymbolic community.\n3. The evaluation on 13 benchmarks across 5 neurosymbolic tasks show the advantage of the proposed DOLPHIN." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes DOLPHIN, a scalable neurosymbolic learning framework that integrates symbolic reasoning efficiently within deep learning models. Unlike existing frameworks that struggle with CPU-GPU data transfer bottlenecks, DOLPHIN enables symbolic programs to operate as a set of inheritanted PyTorch nn.module, allowing efficient GPU-based differentiation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. It seems DOLPHIN only supports neurosymbolic programs with deterministic symbolic processes. For example, if HWF task requires the neural part to predict both numbers and operators (+,-,*,/), the symbolic part cannot be programmed with the Apply function. How DOLPHIN deal with this situation?" }, "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 see \"weaknesses\" above -- in particular my confusion with the DTKP provenance." }, "rating": { "value": 6 }, "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 describes the basic execution and batching mechanism clearly, at least for the DAMP provenance. The Dolphin framework does seem to be an improvement over SOTA in terms of basic usability for certain classes of neuro-symbolic programs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce Dolphin, which is a pytorch-friendly framework for performing neuro-symbolic computations. The neural component of a computation is assumed to be a model, such as an MNIST digit classifier, which outputs a discrete set of symbols (e.g the digits 0-9), with a probability attached to each of them. The symbolic component is a python program which runs a computation over the symbols. The result of symbolic execution is a pytorch tensor, representing final output probabilities for each possible result, which is end-to-end differentiable with the neural components. The authors apply Dolphin to several of neuro-symbolic benchmarks, and show that it is faster than competing frameworks.\n\nDolphin essentially works by running the symbolic program for every possible combination of input symbols, and tracking the probability of each combination. The symbolic program is executed on CPU, but Dolphin evaluation will merge different traces of the program which have the same output into batches. The probabilities can then be computed using batch operations that are GPU-friendly, as well as being end-to-end differentiable with pytorch.\n\nThe authors also provide two different mechanisms, which they call provenances, for tracking probabilities. The DAMP provenance tracks all probabilities, while DTKP tracks only the top-K proofs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors spend a lot of time talking about the easy parts of the problem, and fail to adequately discuss the hard parts. As a result, they gloss over two glaring weaknesses that I see with using this approach to solve anything other than cherry-picked trivial problems. \n\nThe first issue is the combinatorics. When evaluating a function f(A,B), where A and B are distributions over symbols, evaluation must evaluate f(a,b) for every possible combination of symbols { a | a \\in A }, and { b | b\\ in B }. Depending on the exact problem, this can easily lead to an combinatorial explosion in the number of possible outputs. The authors test their code on the \"sum of MNIST digits\" problem, where the combinatorics are reasonable; even given 20 digits, there are at most 181 possible answers. If they were to instead try the \"product of MNIST digits\", which is a tiny change to the code, then the number of possible outputs would balloon, and the technique would likely fail. \n\nThe second issue is control flow. As a symbolic computation, the \"sum of digits\" has no loops or branches, and thus is trivially easy to batch. The authors mention that they support recursive computations, but those generally require a branch to terminate the recursion, and often have divergent control flow. In the presence of branches, different traces of the program take different paths, and no longer cleanly batch together. \n\nThe usual solution (e.g. in CUDA programs) is that when evaluation encounters a branch, it splits the batch of traces into a then-branch and an else-branch, and then merges the traces together again afterwards. Without merging, the traces will continue to diverge on subsequent branches, until each trace is operating independently at batch size 1, and the benefits of parallelism are lost. \n\nMerges happen at the join points in a control-flow graph, which requires the underlying library to build a control-flow graph. Alternatively, since there are only two batched operations (conjunction and disjunction), the authors could first construct an (unbatched) DAG of operations, and then merge/batch together independent nodes of the DAG after the fact, in the style of Looks et al. \"Deep learning with dynamic computation graphs,\" or Neubig et al. \"On-the-fly operation batching in dynamic computation graphs.\"\n\nHowever, the authors make no mention of any machinery to analyze control-flow, build control-flow graphs, or otherwise auto-batch in the presence of divergent control flow. In fact, they do not even provide a discussion or examples of how to write recursive computations with their library at all, despite claiming that it is possible. \n\nMy main objection with both of these issues is that the authors simply don't discuss these problems at all, when IMO they are very clearly major limitations that affect the kind of programs that Dolphin is able to run. \n\nA further weakness of the writing itself is that the authors do not do a good job of explaining the DTKP provenance, which seems like it's quite important. I have several criticisms here. First, it is possible that choosing only the top-K proofs after each operation will address the combinatorics issue, which would be a big deal. However, I'm uncertain, because the authors gloss over combinatorics problem altogether without discussion. Second, the authors claim that their mechanism for merging DTKP tags is equivalent to weighted model counting, but this claim is wholly unsubstantiated. I didn't really understand the formula in Table 1 at all, including how infinities get into the tags. At the very least, the authors should provide a detailed discussion of DKTP in the appendix, ideally with a proof of how it relates to WMC, if space within the paper itself is an issue. \n\nFinally, the authors mention that wrt. to the HWF task, \"the DTKP-AM provenance is more effective than DAMP since the tags in DAMP provenance lack the structure needed to capture the semantics of the symbolic program.\" This statement seems important, but really requires further explanation; I don't understand it at all. Providing HWF as a worked example (perhaps in the appendix) would be valuable to anybody who actually wants to use Dolphin. \n\nErrors:\n\nLine 310: \"Its conjuction operation is defined as the addition of probabilities, and its disjunction is defined as the multiplication of probabilities.\" Unless my understanding is way off base, shouldn't this be the other way around? For independent observations, p(A and B) means multiplying p(A) and p(B)? That's what the authors show in Table 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": 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": "- It would be better to provide a breakdown of the training time (e.g., according to Figure 1) to justify the major efficiency improvement of Dolphin.\n- In Figure 5, for small tasks that all competitors could converge within the time limit, why Dolphin has a better accuracy?\n- In Figure 6, there is a trade-off between accuracy and training time for different provenances (DAMP vs. DTKP-AM). How should we select the most suitable one in practice?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- A new framework for neurosymbolic learning, namely Dolphin, is developed.\n- Dolphin shows superior performance compared to existing frameworks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work presents Dolphin, a brand new framework designed to enhance the scalability of neurosymbolic learning. Dolphin allows developers to write differentiable symbolic programs in Python, utilizing PyTorch for end-to-end GPU acceleration. The framework conveys flexible programmability, end-to-end differentiability, scalability, and tunability, which are essential to handling complex programs and large datasets effectively. Experimental results demonstrate that Dolphin significantly outperforms existing frameworks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My major concern is that the paper is not well structured and hard to follow. In the introduction section, the authors criticized existing frameworks that they must use a separate CPU-based backend and suffer from the slow inter-process data transfers. For me, this is implementation-specific, and it does not drive me to the reasons why we should redesign the entire framework. Although the authors further discuss the challenges in lines 52-67, I find it rather irrelevant to the aforementioned limitation of slow inter-process data transfers.\n\nMoreover, as a new framework, there lacks an overview to depict the layered structures. This prevents readers from having a general picture. It is hard to tell why the designs/implementations could realize the core principles. Additionally, I cannot map the core principles to the challenges discussed in the introduction, either." }, "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": "- Line 213: “Tags are tensors that represent relative likelihoods”. Relative with respect to what? Relative likelihoods (as I understand it) represent a fraction between two likelihoods.\n- The experimental section doesn’t really explain how the CLUTRR and Mugen tasks are solved by Dolphin. E.g. what are the Dolphin programs used here? I think it could be useful to at least include these in the Appendix.\n- I found the naming of “Distribution” unclear. Unless I misunderstand it, a “Distribution” isn’t a probability distribution? (E.g. the Filter operation can remove probability mass.)\n- How did you perform hyperparameter tuning? Did you tune the baselines in a similar fashion? Given that Table 2 compares total training time, better hyperparameters also affect the reported runtime.\n- Why is there such a pronounced accuracy difference between Dolphin and Scallop in some experiments? From what I understand, the provenances like DAMP are essentially inherited from Scallop, so similar accuracy in Scallop should be possible (although not with the same runtime of course).\n\nMinor comments:\n\n- The paper mentions that an “NVIDIA GeForce RTX 4352” was used for the experiments for all experiments (besides CLUTRR). Is this a typo? I’m not aware of the existence of this specific model, and could not find anything about it on the internet. In contrast, the Appendix mentions that a GeForce RTX 2080 Ti was used.\n- For Table 2, what is the unit of time? I assume this is in seconds, but I couldn’t find this anywhere.\n- For Table 2, what is the provenance used for Dolphin? I assume this is DTKP-AM, but I couldn’t find this anywhere.\n- Line 518: “these methods are point solutions”. What do you mean by \"point solution\"?\n- The brackets on the citation on line 107 are wrong.\n - Figure 6 bottom would be more clear with a log y-scale.\n- Several citations refer to the arXiv preprint instead of the conference publication (e.g. for neurASP, CLUTRR, and NeuralLog)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The problem tackled in the paper (scalable and efficient neurosymbolic learning) is relevant and an important issue for the broader neurosymbolic community. \n- Overall, I feel that the paper is rather well-written and includes several useful figures and examples, resulting in a clear presentation of the ideas. \n- The design of Dolphin seems more accessible to a deep learning audience compared to many existing neurosymbolic systems. Notably, it does not require knowledge of e.g. ASP or Prolog (as is the case for NeurASP or DeepProbLog) and can be integrated easily into a deep learning library such as PyTorch or Tensorflow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Many neurosymbolic frameworks have been introduced in recent years, which typically execute their symbolic component on the CPU. This limits their scalability, due to the inferior CPU performance and data transfer latency between the CPU and GPU. The paper introduces Dolphin, a neurosymbolic framework that is fully implemented by parallel tensor operations, and hence can be run on the GPU using conventional deep learning libraries such as PyTorch. The experiments indicate that Dolphin exhibits considerable speed-ups compared to existing frameworks such as Scallop." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Novelty**. The key contribution of the paper - speeding up a neurosymbolic framework by tensorizing it and running on GPUs - is certainly not a new idea. One example of prior work is the LYRICS framework by Marra et al. (2019), which also uses tensor operations to perform the symbolic operations in parallel on the GPU (see e.g. Figure 1 in their paper). Logic tensor networks (Badreddine et al., 2022) and Tensorlog (Cohen, 2020) are some additional examples. These frameworks often also support different provenances / semirings / t-norms. The parallelization of neurosymbolic methods with expressive probabilistic semantics is more challenging, but also here there is plenty of existing work (see e.g. Darwiche (2020) or Dang et al. (2021)). Unfortunately, the paper does not mention prior work on parallelized neurosymbolic learning, nor how it is different from these existing methods.\n\n**Semantics**. It is not clear to me what exact semantics Dolphin aims to achieve. The first provenance (DAMP) is essentially fuzzy semantics (which already has been shown to be easily parallelizable, e.g. Badreddine et al. (2022)). On the other hand, “Apply” mostly brute-force enumerates models meaning the necessary independence assumptions for probabilistic sematics can often hold. (c.f. the MNIST-experiment). The second provenance is the top-k semiring, which is less trivial to parallelize. However, the proposed solution of adding the different proofs destroys the top-k semantics (lower-bounding the WMC). This also results in the introduction of clamp operations, which could lead to zero gradients.\n\n**Language**. A distinction from existing methods is that Dolphin introduces its own set of programming primitives (apply, filter, etc.). Previous neurosymbolic frameworks have typically built on an existing language, e.g. Datalog for Scallop, ASP for NeurASP, ProbLog for DeepProbLog, etc. However, there is no justification for the choice of programming primitives. How does its expressivity relate to existing systems such as Scallop? Why wasn’t an existing language chosen? In my opinion, a lot of different choices could have been made (e.g. why do you need ApplyIf instead of just Apply + Filter?).\n\n**Experiments**. I was surprised that the IndeCateR baseline achieved such low accuracy, given that the experiment seems to be the same as in the IndeCater paper, where the reported results are much better. I just tried out the original IndeCateR implementation myself, and I could replicate the MNIST-addition (L) on my machine in 2 minutes. In contrast, the paper reports a timeout after 10 hours. The accuracy also reaches 86.8%, as opposed to less than 10% in the paper (I'm not sure how the paper reports accuracy if it times out). As the code for the baselines is not included in the supplementary material, I hope the authors can clarify these discrepancies. There are additional issues in the experimental section, e.g. there is no mention of hyperparameter tuning, c.f. the questions section. \n\nLastly, the performance of Dolphin is claimed to be state-of-the-art but I’ve seen several systems get better results on the considered benchmarks (the comparison is hard as actual numbers are not reported, and only bars). To give just some examples, Orvieto et al. (2023) report 94% for Pathfinder, and Manhaeve et al. (2021) report near-perfect accuracies for CLUTRR. I want to stress that I don’t think state-of-the-art results are necessary, but if they are claimed this should be properly supported.\n\nIn summary, the concerns about the novelty of the paper combined with the experimental evaluation issues unfortunately mean I cannot recommend acceptance.\n\n\n**References**\n\nBadreddine, S., Garcez, A. D. A., Serafini, L., & Spranger, M. (2022). Logic tensor networks. *Artificial Intelligence*.\n\nCohen, W., Yang, F., & Mazaitis, K. R. (2020). Tensorlog: A probabilistic database implemented using deep-learning infrastructure. *Journal of Artificial Intelligence Research*, *67*, 285-325.\n\nDang, M., Khosravi, P., Liang, Y., Vergari, A., & Van den Broeck, G. (2021). Juice: A julia \npackage for logic and probabilistic circuits. In *Proceedings of the AAAI Conference on Artificial Intelligence*.\n\nDarwiche, A. (2020). An advance on variable elimination with applications to tensor-based computation. In *ECAI.*\n\nManhaeve, R., Dumančić, S., Kimmig, A., Demeester, T., & De Raedt, L. (2021). Neural probabilistic logic programming in DeepProbLog. *Artificial Intelligence*, *298*, 103504.\n\nMarra, G., Giannini, F., Diligenti, M., & Gori, M. (2019). Lyrics: A general interface layer to integrate logic inference and deep learning. In *Joint European Conference on Machine Learning and Knowledge Discovery in Databases*.\n\nOrvieto, A., Smith, S. L., Gu, A., Fernando, A., Gulcehre, C., Pascanu, R., & De, S. (2023, July). Resurrecting recurrent neural networks for long sequences. In *International Conference on Machine Learning* (pp. 26670-26698). PMLR." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024dolphin,\ntitle={Dolphin: A Programmable Framework for Scalable Neurosymbolic Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3NFtzhFbYM},\nnote={under review}\n}" }, "abstract": { "value": "Neurosymbolic learning has emerged as a promising paradigm to incorporate\nsymbolic reasoning into deep learning models.\nHowever, existing frameworks are limited in scalability with respect to both\nthe training data and the complexity of symbolic programs.\nWe propose Dolphin, a framework to scale neurosymbolic learning at a fundamental level by mapping both forward chaining and backward gradient propagation in symbolic programs \nto vectorized computations.\nFor this purpose, Dolphin introduces a set of abstractions and primitives \nbuilt directly on top of a high-performance deep learning framework like \nPyTorch, effectively enabling symbolic programs to be written as PyTorch modules.\nIt thereby enables neurosymbolic programs to be written in a language like Python that is familiar to developers and compile them to computation graphs that are amenable to end-to-end differentiation on GPUs.\nWe evaluate Dolphin on a suite of 13 benchmarks across 5 neurosymbolic tasks that combine deep learning models for\ntext, image, or video processing with symbolic programs that involve multi-hop \nreasoning, recursion, and even black-box functions like Python `eval()`.\nDolphin only takes 0.33% -- 37.17% of the time (and 2.77% on average) to train these models on the largest input per task compared to baselines Scallop, ISED, and IndeCateR+, which time out on most of these inputs.\nModels written in Dolphin also achieve state-of-the-art accuracies even on the largest benchmarks." }, "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": [ "neurosymbolic learning", "scalability", "vectorization", "differentiable 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/cf4c62a90d33ee275022e0526bfad31afdcf52f4.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": { "value": "/attachment/4e67292b6302dea2fa54ea29e0429646ae1d865d.zip" }, "title": { "value": "Dolphin: A Programmable Framework for Scalable Neurosymbolic 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": 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\" 1." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Originality:\n- The paper introduces a novel in-layer compression approach, departing from conventional pre-LLM compression methods. This represents a fundamental shift in how vision tokens are handled in MLLMs.\n- The adaptive compression strategy that operates within different LLM layers is an innovative solution to the high-resolution image processing challenge.\n\n2. Technical Quality:\n- The work provides solid theoretical foundations by analyzing token compression through the lens of low-rank approximation in self-attention mechanisms.\n- The authors conduct detailed empirical studies to demonstrate why early-layer compression is suboptimal, supporting their approach with concrete evidence.\n- The technical approach is well-motivated through empirical observations about token importance varying across layers.\n\n3. Solid Experiments:\n- The method achieves substantial practical improvements, reducing training/inference time by 20% and vision tokens by 60% while maintaining competitive performance.\n- The 6.2% performance improvement over existing compression techniques represents a significant advancement in high-resolution image processing for MLLMs.\n- The experimental validation is comprehensive, spanning multiple model sizes (0.5B to 7B parameters) and various benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the challenge of processing high-resolution images in multimodal large language models (MLLMs) by introducing ACT-IN-LLM, a novel adaptive compression strategy for vision tokens. Unlike existing pre-LLM compression methods that reduce tokens before LLM processing, ACT-IN-LLM performs compression within different LLM layers through an adaptive compression module (ACM). The method selectively compresses key and value tokens in the self-attention mechanism while retaining all tokens across layers, guided by each layer's final token that encodes the complete multimodal context. The authors provide theoretical analysis demonstrating that their key-value compression approach achieves better low-rank approximation compared to existing compression techniques. Experimental results across various LLM sizes (0.5B to 7B parameters) show that ACT-IN-LLM achieves a 6.2% improvement over existing compression methods while maintaining competitive performance with non-compression models, reducing training/inference time by approximately 20% and vision tokens by 60%." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Methodological Clarity and Analysis:\n- The paper lacks clear explanation of how attention weights across multiple heads are handled in their analysis. This is crucial for understanding their token importance assessment methodology, as different aggregation methods (averaging across heads, selecting specific heads, or analyzing heads separately) could lead to different conclusions about token importance.\n- The analysis of attention weight distributions between different types of tokens (vision-to-vision, vision-to-text, text-to-vision) is missing, which could provide deeper insights into the token compression mechanism.\n- The authors could strengthen their analysis by including entropy measurements of attention weights for different tokens, which would provide quantitative support for their token selection strategy.\n\n2. Technical Presentation:\n- The paper introduces concepts like \"high-resolution\" and \"low-resolution\" tokens without first establishing the context of LLaVA's AnyRes visual encoding scheme. This may create confusion for readers not familiar with the underlying visual encoding mechanisms in multimodal LLMs." }, "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": 4 }, "primary_area": null, "questions": { "value": "1. As you are progressively shrinking the ratio $r_i$, $r_j$ and $r_p$, how do you determine where and how much should you shrink?\n\n2. Why do you use the attention map from the previous layer to guide vision token compression instead of the current layer's attention map?" }, "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 well-written and easy to follow, offering a clear mathematical proof to demonstrate the theoretical effectiveness of the proposed method.\n\n2. The proposed approach is both intuitively and mathematically sound.\n\n3. The experimental results are thorough, complemented by a comprehensive ablation study." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel Adaptive Compression Module (ACM) designed to dynamically reduce the number of high-resolution image tokens for key/value during the forward pass. The ACM leverages attention maps to identify and retain the most relevant high-resolution tokens, preserving only the top k tokens for key/value. Experimental results demonstrate the efficiency and effectiveness of this approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The evaluation benchmark is somewhat limited. Adding more benchmarks, such as visual grounding benchmarks and other non-text-related high-resolution benchmarks like V* Bench, would facilitate more comprehensive evaluations." }, "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. Based on the results presented, all query tokens are retained, while key and value tokens are compressed in the self-attention mechanism. If the value tokens are compressed, the number of tokens outputted from the attention block should match the number of compressed value tokens. Then why are all image tokens still preserved when entering the final LM head? Please explain this in detail.\n2. Please provide a detailed explanation of how your strategy reduces computational load across various components." }, "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 provides a valuable perspective on why early token deletion should be avoided.\n2. The writing is generally smooth, and the presentation of figures and tables is visually appealing, contributing positively to the overall presentation of the paper.\n3. The motivation section reads coherently and introduces the problem that needs to be addressed in a natural manner." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposed ACT-IN-LLM method improves Multimodal Large Language Models (MLLMs) by adaptively compressing vision tokens within different layers, unlike traditional methods that reduce tokens before reaching the LLM. This approach preserves all tokens throughout layers, selectively compressing only key and value tokens in self-attention to maintain critical information while reducing computational load." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper dedicates significant space to comparisons with traditional methods that compress tokens before using LLMs. However, it overlooks an essential baseline, FastV [1], which also compresses image tokens within the LLM itself and allows for direct comparison of training results. This omission makes the paper less convincing.\n2. The paper claims \"reducing training/inference time,\" but does not provide any data demonstrating training time reduction.\n3. The proposed strategy appears usable without training; therefore, it would be beneficial to include results showing inference acceleration without additional training.\n4. The existence of Table 3 is quite awkward: first, there are numerous gaps in the table, and second, the training data is entirely different, making these models incomparable.\n[1]An Image is Worth 1/2 Tokens After Layer 2: Plug-and-Play Inference Acceleration for Large Vision-Language Models, https://arxiv.org/abs/2403.06764" }, "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": "Thanks for the authors' valuable exploration in this area. I have several concerns, and if these can be addressed, I would like to raise my rating.\n\n1. The reported compression rate seems relatively low (the proposed method achieves 83% of the full model's performance and 94% of its memory usage according to Table 2). Would it be possible for the authors to provide results with a higher compression ratio (around 60%, for example) to more effectively demonstrate the advantages of the proposed method?\n2. It appears that FastV[1] is not included in the comparisons. Could the authors consider providing comparisons with FastV, particularly at higher compression ratios (such as around 60%)?\n3. The current reliance on text-guided token selection may have limitations. Have the authors considered incorporating the complexity of visual information into the compression strategy?\n\n[1] https://arxiv.org/pdf/2403.06764" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "By constructing a unified formulation of compression methods and analyzing low-rank approximations, the paper provides a robust theoretical framework supporting its method. Experimental results convincingly demonstrate that ACT-IN-LLM outperforms existing pre-LLM compression and interaction-based methods, achieving competitive accuracy while significantly reducing token count and computational costs. Notably, the proposed method shows good scalability, with consistent gains observed across larger model sizes and datasets. These strengths suggest that ACT-IN-LLM offers a practical, efficient solution for high-resolution MLLM applications, and the work is both well-structured and empirically solid." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper examines the limitations of token compression strategies in MLLMs when processing high-resolution visual inputs. It presents ACT-IN-LLM, an approach designed to address these limitations by adaptively compressing visual tokens across LLM layers, contrasting with existing methods that apply compression before token input to the LLM. The authors claim this layer-wise, interaction-guided compression effectively preserves essential visual information, improving model accuracy while reducing computational load. Experiments indicate significant performance gains over prior compression strategies and competitive results with non-compression methods, highlighting ACT-IN-LLM’s potential to enhance high-resolution MLLM efficiency and scalability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Relying solely on text-guided token selection may limit the model's adaptability, as it could overlook the inherent complexity of visual information itself. Without considering factors like scene detail or object density, the compression strategy might miss important visual nuances, potentially affecting performance across diverse tasks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024actinllm,\ntitle={{ACT}-{IN}-{LLM}: Adaptively Compression Vision Tokens in {LLM} for High-Resolution Multimodal Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Ofy2jNsNL},\nnote={under review}\n}" }, "abstract": { "value": "High-resolution inputs empower Multimodal Large Language Models (MLLMs) to capture intricate visual details, thereby enhancing comprehension. However, the self-attention mechanism’s quadratic complexity poses significant computational and memory challenges as image resolution increases, particularly with long-vision tokens. Existing approaches generally alleviate these issues by reducing vision tokens before feeding them into LLMs. Although efficient, this Pre-LLM compression strategy fails to match the performance of models utilizing all tokens, particularly on high-resolution benchmarks. Our experiments reveal that the performance gap arises from this strategy’s limitation in selecting important visual tokens in early LLM layers, leading to the irretrievable loss of critical information. To overcome these challenges, we propose a new strategy that Adaptively Compresses vision Tokens within different LLM layers, named ACT-IN-LLM. Our innovative approach retains all tokens throughout the layers to ensure no vital information is lost while compressing key and value tokens in the self-attention mechanism, to reduce computational costs. The layer-wise compression of ACT-IN-LLM is guided by the interaction information between vision and text tokens, leading to more accurate selections. Our theoretical analysis and extensive experiments demonstrate the effectiveness of ACT-IN-LLM, showing a 6.3% improvement over existing token compression techniques. It also achieves the competitive performance with non-compression methods, while reducing training/inference time by ∼ 20% and vision tokens by ∼ 60%." }, "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": [ "Multimodal Large Language Models; High-resolution; 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/078f5730c544160d26a27af1352e3a5063d39057.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": "ACT-IN-LLM: Adaptively Compression Vision Tokens in LLM for High-Resolution Multimodal 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": "* In the tables with the numbers for metrics please highlight the best numbers (bold)\n* What hardware was used for training the model? Training time, memory requirements. Add to A.4\n* In theory the method should work for objects that traditionally have challenging reflectance properties such as hair or fur. I am not sure if hair and fur were part of the training dataset, but it still might be interesting to see if it works." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Pseudocode in A.1clarifies all misunderstandings from the text descriptions. I can not stress enough how much I like this way of communicating the pipeline.\n* The approach uses a diffusion process instead of a multi-view optimization, which is exceptionally fast in direct comparison.\n* Trained on a rich dataset with a huge variety of lighting conditions.\n* Light representation (input) is a simple environment map. Changing lighting is therefore very simple (no complicated neural light representation)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a method to generate a relighable 3D representation using 3DGS for the geometry and a diffusion based model to get the illumination dependent appearance parameters for those Gaussians.\nThe geometry is predicted in form of per pixel Gaussians from the sparse input views (4 - 16) in a single Transformer forward step. The tokens of the geometric representation is concatenated with HDR features extracted from the illumination given as environment map and the noise target views. After denoising the tokens for everything except the input gaussians are discarded. The remaining tokens are decoded into the appearance (SH) of the Gaussians. The Gaussians are then rasterized into any novel view. The diffusion model is trained such that the lighting of the rendered Gaussians should match the lighting of the input environment map. During inference this environment map and the target view camera pose can be arbitrarily be chosen, thus a scene can be reconstructed from sparse views and then relit." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Knowing only even a fraction of the research and work that has gone into disentangling the ambiguity between lighting and shading it rubs me the wrong way to read something that suggests it solved it without actually addressing the core problem. The method does not really decompose a scene into geometry, lighting and shading and is not usable if the use case would require extracting or editing reflectance properties. The way I see it this paper does relighting by decomposing a scene into geometry and appearance, however this is very different to what methods which explicitly extract reflectance and illumination do. The problem statement is profoundly different if you have to produce a explicit, reflectance representation which represents some underlying physical property of the surface, compared to just estimating the product of light and reflectance. I don't think much has to be changed to show respect for this difference: In 2.1 it should be mentioned that the method models the appearance under arbitrary illumination without an explicit reflectance representation. In the introduction the claim that the ambiguity between shading and lighting, is overcome should be phrased more carefully or be clarified. As far as I understand this paper estimates appearance as the integrated product between the unknown shading and a given lighting in form of a view dependent SH. This is really great work, but should not be confused with reflectance decomposition." }, "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": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- originality-wise: I appreciate the proposed end-to-end transformer-based architecture for relightable 3D object reconstruction. \n- quality-wise: when taking into the efficiency of the proposed approach, I think quantitatively the performance is good.\n- clarity-wise: the paper is well-written and easy to follow.\n- significance-wise: the task of relightable 3D object reconstruction is vital for many downstream applications, e.g., AR/VR and robotics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the task of reconstructing relightable 3D objects from sparse images. For this, the authors extend the idea of GS-LRM to incorporate the generation of a relightable appearance. Specifically, they feed the geometry feature from GS-LRM through a diffusion process based on the transformer architecture. The transformer attends to the target illumination and outputs appearance tokens. The combination of the appearance tokens from the newly added transformer and the geometry tokens from the original GS-LRM transformer concludes the generation of Gaussian Splats. Experiments on the Stanford-ORB and Objects-with-Lighting real-world benchmark as well as the TensoIR synthetic benchmark demonstrate the effectiveness of the proposed approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "## Concerns about the architecture\n\nI think the authors missed an important question to answer: why do we choose the current architecture? Essentially, the mechanism of discarding many tokens (L231) is wasting the model's capability. \n\nThe root question is why do we need to use self-attention instead of cross-attention? Why do the environment maps and the denoised images need to be treated the same as the appearance tokens, especially since only the appearance tokens will be used for the Gaussian Splats rendering?\n\n## Not enough understanding of the current model\n\nEven though with the current architecture, I do not think the authors provide enough analysis. Specifically, have the authors visualized the attention maps of the transformer? What does the transformer learn? Does it attend to some specific areas in the environment map that cause the specular effects in the rendering? How does the transformer attend to those denoised images?\n\n## Concerns about the qualitative results\n\nIn Fig. 3.(a), the produced Pepsi can's color is quite different from the ground truth. A similar thing happens to the gnome's colors. Further, the characters on the Pepsi can are quite blurry compared to NVDiffRec-MC / ground-truth. Additionally, in Fig. 3.(c), the RelitLRM produces quite different shows from the ground truth. However, the shadows are correctly predicted by both InvRender and TensoIR. \n\nWhy is this the case? Have the authors carefully studied the causes? Whether increasing the number of views will help? Will 16 views mitigate these issues as the authors state that \"performance saturates around 16 views\" (L525)? If even 16 views cannot resolve the issue, what are the intrinsic shortcomings of the proposed model?\n\nI hope the authors can provide a systematic analysis for a good understanding of the model.\n\n## Missed important baselines\n\nI think the authors missed several quite related as well as important baselines, e.g., [a, b]. They both use the idea of diffusion-based relighting model to tackle the task of relightable reconstruction.\n\nEspecially IllumiNeRF [b], which directly tackles the task of relightable object reconstruction and competes on the benchmark of Stanford-ORB and TensoIR. Frankly speaking, [b] outperforms the proposed approach on both benchmarks quantitatively:\n- PSNR-H / PSNR-L / SSIM / LPIPS: 24.67 / 31.52 / 0.969 / 0.032 (RelitLRM) vs 25.42 / 32.62 / 0.976 / 0.027 ([b] on Stanford-ORB)\n- PSNR / SSIM / LPIPS: 29.05 / 0.936 / 0.082 (RelitLRM)) vs 29.709 / 0.947 / 0.072 ([b] on TensoIR)\n\n[a] A Diffusion Approach to Radiance Field Relighting using Multi-Illumination Synthesis. EGSR 2024.\n\n[b] IllumiNeRF: 3D Relighting without Inverse Rendering. ArXiv 2024.\n\n## Concerns about the novelty claim\n\nOn L088, the authors claim the first contribution as \"a regression-based geometry reconstructor followed by a diffusion-based appearance synthesizer\". Though I appreciate the end-to-end transformer architecture, I may not be convinced that the idea is entirely novel since the above-mentioned IllumiNeRF has already proposed an almost exact idea. I would recommend modifying it to correctly position the work.\n\n## Missed training details\n\nWhat is the hardware setup required to train the model? How long does it take to complete the training, hours or days?\n\n## Incorrect description about the benchmark\n\nIn L356, the authors state that the Stanford-ORB benchmark has \"60 training views and 10 test views per lighting setup\". This is not true. Please correct it.\n\n## Typos\n\nL290: \"is involves\" -> \"involves\"" }, "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 the paper, the authors provide the rendering performance. However, I cannot find the training time. Please provide more specifics on the training setup, such as the number of GPUs used and total training hours. \n- A more thorough analyais and discussion of why performance plateaus between 8 to 16 views would enhance the paper's quality.\n- Please provide quantitative evaluations of the extracted geometries.\n- In the object insertion image (Figure 1(c)), how is illumination in the scene accounted for? Did you sample on the position of the object to capture the surrounding environment and incorporate the environment map into your model? Additionally, how do you account for the indirect light from another object produced by your RelitLRM in the scene?\n- The method still takes 2 to 3 seconds to render. In contrast, the geometry and textures obtained from other methods can be rendered using many real-time rendering techniques. Moreover, in current industrial applications, it is challenging to abandon triangle meshes and textures. Therefore, this method cannot be considered a viable solution for 3D digital assets. However, if this approach could be extended to scene relighting, its applicability could be significantly broadened.\n- Is the number of Gaussians predicted by the model sufficient for capturing different topologies and structures across diverse objects?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The method models geometry and rendering independently, akin to NeRF’s modeling architecture. It first constructs geometry from multi-view images, then generates the rendering by combining geometry features with environmental maps.\n- The method is practical as it only requires 4-8 views as input. It is able to effectively reconstruct relightable objects without per-scene optimization.\n- On both synthetic and real-world datasets, the method offers competitive relighting results, requiring much fewer input views than other methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method trained end-to-end on synthetic multi-view renderings of objects under varied, known illuminations. The approach is able to generate high-quality Gaussian splatting representations of 3D objects under novel illuminations from sparse input images (4-8 views) captured in unknown, static lighting conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- In the paper, the authors claim that performance plateaus at around 16 views. However, as shown in Table 5, there is only a marginal improvement in image quality on the TensoIR-Synthetic dataset when increasing from 8 views to 16 views. \n- The novelty of the approach needs clearer articulation. The authors state that their method differs from Neural Gaffer (Jin et al., 2024) by not being limited to single-view inputs. However, this advantage seems to stem from the use of GS-LRM (Zhang et al., 2024a). It is important to clarify how their application of diffusion for relighting distinguishes their method from existing techniques.\n- This method separates geometry from rendering, but the paper does not show results of the decomposed geometry. It is unclear how good or bad the quality of the reconstructed geometries is." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose \\emph{\\method}, a Large Reconstruction Model (LRM) for generating high-quality Gaussian splatting representations of 3D objects under novel illuminations from sparse (4-8) posed images captured under unknown static lighting." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024relitlrm,\ntitle={Relit{LRM}: Generative Relightable Radiance for Large Reconstruction Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3Oli4u6q3p},\nnote={under review}\n}" }, "abstract": { "value": "We propose RelitLRM, a Large Reconstruction Model (LRM) for generating high-quality Gaussian splatting representations of 3D objects under novel illuminations from sparse (4-8) posed images captured under unknown static lighting. Unlike prior inverse rendering methods requiring dense captures and slow optimization, often causing artifacts like incorrect highlights or shadow baking, RelitLRM adopts a feed-forward transformer-based model with a novel combination of a geometry reconstructor and a relightable appearance generator based on diffusion. The model is trained end-to-end on synthetic multi-view renderings of objects under varying known illuminations. This architecture design enables to effectively decompose geometry and appearance, resolve the ambiguity between material and lighting, and capture the multi-modal distribution of shadows and specularity in the relit appearance. We show our sparse-view feed-forward RelitLRM offers competitive relighting results to state-of-the-art dense-view optimization-based baselines while being significantly faster. Our project page is available at: https://relitlrm.github.io/." }, "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": [ "Relightable reconstruction", "Inverse Rendering", "Generative Relighting" ] }, "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/8a5a5db26b59913c565de14b69879b782caab1c1.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": "RelitLRM: Generative Relightable Radiance for Large Reconstruction 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": 2 }, "primary_area": null, "questions": { "value": "I will try and cluster my questions in sensible groups.\n\n1. _Theoretical Understanding_:\n - Can you provide theoretical insights into when WC sampling should outperform SE sampling?\n - How does the optimal sampling ratio change with model size and task complexity?\n - What are the key factors that determine the success of weak-to-strong improvement?\n\n2. _Methodology_:\n - How would the results change with more sophisticated filtering strategies?\n - Could you provide more details about the specific prompting strategies used?\n - How sensitive are the results to the choice of temperature and sampling parameters?\n\n3. _Generalisation_:\n - What characteristics of a task make it more/less suitable for WC sampling?\n - How would the results scale to even larger model sizes?\n - What is the relationship between FPR and final model performance?\n\n4. _Practical Implementation_:\n - How would you recommend implementing this in scenarios without ground truth?\n - What modifications would be needed for different domains or tasks?\n - Could you provide more detailed guidance on optimal sampling strategies for different scenarios?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. _Originality_:\n - Introduces a novel compute-matched sampling framework with clear mathematical foundations\n - Proposes a new \"weak-to-strong improvement\" training paradigm that challenges conventional wisdom\n - Provides a fresh perspective on the compute-quality trade-off in synthetic data generation\n\n2. _Experimental Rigour_:\n - Comprehensive evaluation across multiple dimensions:\n * Multiple model pairs (both open and closed models)\n * Various compute budgets and training paradigms\n * Different dataset sizes and difficulty levels\n - Thorough ablation studies that isolate the impact of coverage and diversity\n - Both human and automated evaluation of false positive rates\n - Clear validation of results through transfer learning (Functional MATH)\n\n3. _Practical Impact_:\n - Demonstrates significant cost savings potential (0.15x cost for comparable or better performance)\n - Shows consistent improvements across model sizes (7B to 27B)\n - Provides actionable insights for practitioners\n - Results particularly relevant given the trend of improving smaller models\n\n4. _Technical Depth_:\n - Rigorous mathematical formulation of compute-matching\n - Analysis of traed-offs between coverage, diversity, and error rates\n - Ablation studies support main claims\n - Clear empirical validation of theoretical framework" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper challenges the common practice of using strong but expensive (SE) language models to generate synthetic training data, proposing instead that using weaker but cheaper (WC) models may be more compute-optimal. The authors introduce a \"compute-matched sampling\" framework that enables fair comparison between WC and SE models by accounting for their relative compute costs. At a fixed compute budget, this framework shows that one can generate P_SE/P_WC more samples from a WC model than an SE model. The authors evaluate this approach across multiple model pairs (Gemma2 9B/27B and Gemini Flash/Pro), tasks (primarily mathematical reasoning), and training paradigms (knowledge distillation, self-improvement, and a novel \"weak-to-strong improvement\"). They assess the generated data along three key dimensions: coverage (problems solved), diversity (unique solutions per problem), and false positive rate (correct answers with incorrect reasoning). The results consistently show that training with WC-generated data outperforms SE-generated data when properly compute-matched." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. _Theoretical Foundation_:\n - Lacks formal analysis of when WC sampling should outperform SE sampling\n - No theoretical bounds on the optimal sampling ratio\n - Missing analysis of the relationship between model size and optimal sampling strategy\n - Limited exploration of failure modes and their characteristics\n\n2. _Methodology Limitations_:\n - Heavy reliance on ground truth for filtering solutions\n - Limited exploration of alternative filtering strategies\n - FPR evaluation methodology could be more robust (50 human samples probably insufficient)\n - Some key implementation details relegated to appendices\n\n3. _Generalisation Concerns_:\n - Primary focus on mathematical reasoning tasks\n - Limited exploration of other domains (coding results show context-dependency)\n - Unclear scalability to larger model sizes\n - Performance on more complex reasoning tasks not fully explored\n\n4. _Practical Considerations_:\n - Deployment challenges in scenarios without ground truth not fully addressed\n - Resource optimisation strategies could be explored more\n - Limited discussion of integration with existing training pipelines\n - Cost-benefit analysis could be more comprehensive across different 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": 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. Although both low and high budgets are studied, could you please provide the results of an extremely high budget where the cost is not an important factor? This should be indicative of diverse data scales.\n2. Despite the train-test splits of MBPP, this paper only trains models on MBPP and tests them on HumanEval. The testing results on MBPP are expected to be provided for a more comprehensive understanding.\n3. Writing：\n- 1. All the ref links are invalid.\n 2. l101, l104: \"i.e.\" should be \"i.e.\".\n 3. l103: grammar error for \"we supervise finetune\".\n 4. l109: use \\citet{} for \"(Zelikman et al., 2024;Singh et al., 2023).\"." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The research question is significant, focusing on performance comparison of data sampled from WC and SE models, respectively.\n2. The findings are impressive, challenging the traditional belief that data from a strong model is better for finetuning models.\n3. The evaluation settings are diverse, demonstrating the effectiveness and robustness of this method, despite only the Gemma series models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper revisits the trade-offs between generating synthetic data using a stronger but more expensive (SE) model versus a weaker but cheaper (WC) model, and finds that at a fixed sampling compute budget, finetuning LMs with data from a WC model can consistently outperform data from a SE model in multiple settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper centers exclusively on the Gemma series, and it is essential to extend the analysis to the Llama series to demonstrate the robustness of the conclusions.\n2. While the paper aims to highlight the lower computational cost of the WC model for data synthesis (particularly important for large-scale data generation), all the experiments are conducted on relatively small datasets. This discrepancy undermines the overall contribution of the paper.\n3. Compared to the SE model, the WC model can be regarded as a more diverse yet lower-quality variant. Therefore, it is crucial to compare it with techniques designed to enhance output diversity. Specifically, if adjusting the sampling temperature of the SE model consistently results in performance degradation relative to the WC model, this suggests that the WC model provides a superior quality-diversity trade-off compared to merely increasing the sampling temperature." }, "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": "It seems that the difference in data quality between the WC and SE models becomes larger at lower budgets. Is it possible that the WC and SE models generate data of similar quality when the budget is very high?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well-structured and clearly written, making the methodology and results easy to follow.\n\nThe experiments are well-executed and provide convincing evidence of the benefits of the proposed approach.\n\nIt addresses a critical issue in synthetic data generation, offering a valuable contribution to this area of research." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the novel observation that generating synthetic data using a weaker but cheaper (WC) model is more effective than using a stronger but more expensive (SE) model. The authors demonstrate that, under the same budget, data generated by WC models tend to have higher coverage and diversity, though with a corresponding increase in false positive rates. Additionally, they show that models fine-tuned on WC-generated data consistently outperform those trained on SE-generated data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The conclusion may not hold when using models from different companies. Based on my experience, under the same budget, data generated by a larger model like Qwen2.5 7B could outperform that of a smaller one like Gemma2 2B.\n\nThe paper could benefit from experimenting with more complex reasoning tasks, such as tree search algorithms, and using a reward model to evaluate the quality of the generated 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": "* Your distillation setup is limited to finetuning. One setup where it would be more realistic to not have enough budget is pretraining. Do you have any results on this? I of course do not expect to pretrain a network until convergence during the rebuttal, but it would already be helpful if you could show the first couple of iterations just to make sure the worse data (higher FPR) does not seem to converge to a much worse model.\n* I'd be interested in sample-matched figures. The figures where I'd be most interested in a sample-matched comparison are Figures 4c and 5c. This would allow finding out if a small model can successfully improve a larger model, which would challenge beliefs in the field.\n* Just to go sure: In the self-improvement setups, you keep training a model iteratively on its own generations from the current parameters? Or do you mean that you finetune a \"fresh\" 7B model using an already converged 7B model?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The paper investigates not only knowledge distillation, but also a self-improvement setup\n* In Figures 4 and 5, it is interesting that training on model-generated solution paths (one per question; at least in the \"27B low compute\" setup) gives a better performance than training on human-provided solution paths (also one per question)\n* Figure 7 carries an interesting finding: Despite training on the data generated by the small model which has more errors, the ultimate trained model does not have more errors in its reasoning. This implies that the additional data mitigates its lower quality, which might add evidence to the discourse beyond the setup studied in this paper.\n* The writing and flow of experiments is mostly clear" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates whether it is better to (self-)distill from a Gemma-27B LLM or to distill three times more finetuning data from a three times smaller Gemma-9B model. It finds that the three times more data of the smaller model, despite including more errors, leads to a higher performance of the finetuned student model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* In 21 of the 22 Figures, the paper hinges on \"matching the compute\", i.e., being allowed to generate 3 times more data when using the 3 times smaller LM. This confounds two factors of variation, making it hard to interpret the findings. This is the main weakness of the paper. One idea to improve on this weakness would be to test out distilling 1, 3, and 9 samples from the large LLM and 3, 9, and 27 samples from the small LLM (instead of the current 1+10 vs 3+30), so that there are both overlapping settings with a matched number of samples and with a matched compute.\n* In the only figure where the small LLM is compared to the large LLM without this advantage (Figure 20 in the appendix), the large LLM produces better training data. It can be expected that if we use the large LLM to generate enough data until the student model converges, it will make a better distilled model. Thus, the only real application of the proposed method is when we do not have enough budget to produce enough data to converge. For the finetuning setup of the paper, that would amount to not being able to generate data for 8k-12.5k questions. This is a setup with limited applicability in practice. It would increase the contribution (score) of the paper to investigate problems where budget limits are hit more frequently in practice, like pretraining, see also my question below.\n* Relative and absolute increases are reported inconsistently. E.g., in Figure 3b the fact that the proposed small model finds 11 instead of 5 solution paths per question (when it is allowed to generate 3 times more paths in total) is reported as a 125% increase (line 268), whereas the fact that 24% of its solutions paths are wrong compared to 17% of the large model is reported as a 7% increase (line 310). This inconsistency becomes problematic when reporting the increase on percentage numbers (e.g., line 258), where it is unclear whether this is a relative or absolute increase. Keeping the reporting consistent would increase both the presentation and the scientific soundness scores.\n* The paper only evaluates Gemma (/Gemini) models. It would help judge the generalization of the claims (and increase the contribution score) to test it out on at least one other LLM, like a Llama model.\n* The datasets are very limited to two math datasets, limiting the contribution. As above, more datasets would help judge the range of applicability, especially whether it also works on non-math and non-reasoning datasets.\n* The paper does not compare to baselines, despite citing multiple closely related approaches\n* The method still requires annotated data, because the LLM-generated data needs to be filtered out if it does not match the GT. It would increase the applicability of the score (and thus the contribution score) if there would be an ablation without filtering, i.e., answering whether the unfiltered erroneous data from the smaller model can still train a better model.\n\nSmall notes that did not influence my score and don't need to be rebuttled, I just note them to make the camera-ready better:\n* The first paragraph of Section 3 could be shortened; it's message (in Equation 1) is just \"if a model has x times more parameters, it takes x times longer to generate\".\n* typo in line 54, \"filters\"\n* typo in line 103 \"we supervise finetune\"\n* typo in line 151, \"consists\"\n* typo in line 157, \"for training student LM\"\n* typo in line 241, \"that where\"\n* The references exclusively list arxiv versions of the papers, not their actual published versions\n* The reference .bib file should best use double brackets for \"{{OpenAI}}\", \"{{The Llama Team}}\", to prevent the ill formatting in line 483 (\"Team, 2024; Anthropic, 2024; AI, 2024\")" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024smaller,\ntitle={Smaller, Weaker, Yet Better: Training {LLM} Reasoners via Compute-Optimal Sampling},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=3OyaXFQuDl},\nnote={under review}\n}" }, "abstract": { "value": "Training on high-quality synthetic data from strong language models (LMs) is a common strategy to improve the reasoning performance of LMs. In this work, we revisit whether this strategy is compute-optimal under a fixed inference budget (e.g., FLOPs). To do so, we investigate the trade-offs between generating synthetic data using a stronger but more expensive (SE) model versus a weaker but cheaper (WC) model. We evaluate the generated data across three key metrics: coverage, diversity, and false positive rate, and show that the data from WC models may have higher coverage and diversity, but also exhibit higher false positive rates. We then finetune LMs on data from SE and WC models in different settings: knowledge distillation, self-improvement, and a novel weak-to-strong improvement setup where a weaker LM teaches reasoning to a stronger LM. Our findings reveal that models finetuned on WC-generated data consistently outperform those trained on SE-generated data across multiple benchmarks and multiple choices of WC and SE models. These results challenge the prevailing practice of relying on SE models for synthetic data generation, suggesting that WC may be the compute-optimal approach for training advanced LM reasoners." }, "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 and small language models", "reasoning", "math", "compute-optimal", "sampling", "supervised finetuning" ] }, "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/af167d3c7bc7f6070d6d302d1551976ede8917bd.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": "Smaller, Weaker, Yet Better: Training LLM Reasoners via Compute-Optimal Sampling" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]