Datasets:

babytreecc

/

ICLR2025review

like 7

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "I’d like to rate the current submission reject due to limited technical contributions and lack of convincing experiments. The paper needs significant changes including new experiments and possibly methodological improvements in justifying the practical use behind the proposed 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 well written and easy to follow.\n- Interesting topic -- making linear attention efficient is of great interest to the research community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper attempts to make linear attention efficient and useable by reducing the memory consumption and introducing an alternative mechanism to dropout. Results show the usefulness of the proposed approaches, while maintaining the linear scaling in N in both wall-clock time and memory usage." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The results are somewhat unsatisfactory, with only limited to a very small model. More experiments and analysis are needed to justify the effectiveness of the proposed method.\n- What about practical usage? Can this method start from the pre-trained LLMs and directly convert them to efficient linear attention? Also, can it start from a LLM with linear attention and make it efficient through little amount of training or finetuning?\n- Can this method scale to large models? Say, billion scale models. Authors at least perform experiments using 1-3B models.\n- What about hardware efficiency of the proposed changes to the linear attention?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Same Weakness" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "In terms of practicality, reducing memory cost is very crucial. If this work maintains comparable performance or minimizes the performance drop, it has impactful potential." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses two major limitations in current linear attention mechanisms. First, although existing approaches aim to reduce computational complexity to linear time, they still practically require O(ND²). To address this, this work proposes a method to lower the memory usage to O(ND). By deriving and computing the attention layer gradient analytically, this paper achieves this efficiency without relying on a differentiable programming library, thus avoiding the need to store variables for backpropagation. This work uses Factorization in both forward and backward passes and validates the reduced memory and time complexity in the context of attention layers and large language model training. Additionally, because linear attention doesn’t inherently compute an attention matrix, dropout cannot be directly applied. To overcome this, this work introduces an alternative mechanism that emulates the effect of dropout, with its effectiveness confirmed through an ablation study." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "### Lacks of benchmarks\nAlthough it shows memory and latency experiments by increasing token length, it didn't show common LLM benchmarks such as MMLU and GLUE or long-context LLM benchmarks such as n streaming books (PG19), Long Context Understanding (LongBench), and book summarization (Booksum)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 does the proposed method compare with those RNN-like methods such as Mamba and RetNet?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The Motivation is clear.\n\n- The theoretical derivation is detailed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript proposes two techniques to improve the practical efficiency of linear attention. One is the factorization of matrix multiplication, and the other is a modified dropout technique. Toy experiments show the effectiveness of the two techniques." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall, this manuscript is not ready for publication. My concerns are listed as follows:\n\n- The technical novelty is limited. The first proposed method, matrix multiplication multiplication, is a standard operation in linear attention methods [1]. I see some novelty in the second dropout technique, but its effectiveness is validated only in small-scale experiments.\n\n- The experiments are not convincing enough. One small model (Pythia-14M) is adopted, and only a training curve is presented. To make the results more convincing, it is recommended to conduct experiments on larger-scale models (both language and vision models should be included) and well-known benchmarks. The comparison should be the about the trade-off between efficiency and performance.\n\n- The presentation needs substantial improvements. Some examples are listed as follows:.\n - In the introduction, the authors categorize linear attention methods into two lines of work: Sparsified/localized Attention, and Kernel Separation. However, no references are cited here.\n - In the equations (e.g. eq1), the index subscriptions $i,n,j$ should not have bold fonts.\n - In Sec. 3.1, the FLOPs calculation example could be given by normal matrix presentations instead of a 3x3 toy example.\n - The derivaton in Sections 3&4 are too detailed and might distract readers from the main method. It is recommended to put detailed derivation process in the appendix, and present the major method in the main text.\n\n\n[1] Flatten transformer: Vision transformer using focused linear attention. ICCV, 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": "1. In section 3.1, the final operation we need to calculate is \n\n[a_1, a_2, ..., a_n]^T \\times [b_1, b_2, ..., b_n] \\times [d_1, d_2, ..., d_n]^T\n\nIf we compute the first \\times first, there will be n^2+n^2 multiplications and n*(n-1) accumulations. If we compute the second \\times first, there will be (n+n) multiplications and (n-1) accumulateions.\n\nSo the point sec 3.1 made it that, change the operation order will save FLOPs. Do I miss something?\n\n2. The new dropout alternative seems complicated. It seems no experiments discuss its complexity. Will it take much inference latency?\n\n3. Overall, I give my initial rating as \"marginally below the acceptance threshold\". I hope the authors can solve my concerns as I mentioned in weakness and questions section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The topic of saving the memory footprint of linear attention is interesting. This is important while it seems few researcher have noticed.\n\n2. The forward and backward derivation is also very clean.\n\n3. An alternative for dropout is proposed, and the experiments in LRA and Tiny Stories datasets verifies its effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focus on reduce the space complexity of vanilla linear attention. To be specific, the authors change the order of some matrix multiplication in the forward pass of linear attention when the kernel is linear. The authors also provide a detailed derivation of the forward pass and backward pass. Because linear attention does not have a explicit attention map in the forward pass, the authors propose an alternative for the dropout regularization mechanism. Some testing experiments shows this method are indeed save the memory footprint, but the experiment is not abundant." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) After the forward and backward modification, althought the computational seems equivalent compared with linear attention, will the proposed method achieve the same performance is still unclear. The authors can check the RepVGG paper, which shows that equivalent computational may not produce the same performance.\n\n(2) The paper only change is computing order of Linear Attn, and proposed a new dropout alternative. I think the novelty is limited.\n\n(3) Lack of experiment.\n\n (3.1) for the time and memory scaling experiments (sec. 5.1). The authors only show one fixed number of head H, fixed head dim D, varying token length result experiment, and, one fixed token length N, varying head dim D experiment. I think the author should show more results on the H, D choices, since this experiment is training free, and whether the experiment results would show the same trend. Additionally, the only chosen one is weird. Since in ViT-Base H=12 and D=64, in DiT-S H=6 and D=64, I cannot find a popular model with H=16 and D=32.\n\n (3.2) For LLM experiment in Sec. 5.2, the author does not express the reason why they choose this model and this dataset. And they only show the loss curve (also the orange curve is not trained as long as the blue one is). No evaluation metric provided.\n\n (3.3) For the ablation study in Sec. 5.3, I do not get the point why the author do not provide these accuracy results in main results, but put them in ablation. The training hypermeter is also not clear. I have also check the LRA results in other papers (e.g. S4, by Albert Gu in ICLR 2022, https://arxiv.org/pdf/2111.00396). The results of LRA in the submission are far lower than this 2 years ago paper. And both the Softmax result and Linear(alt. drop.) results have different numerical range compared with the results in S4 paper.\n\n(4) The experiment hyper-parameter is not described. That may be the reason why it cannot match the results in S4 paper. I think the authors should follow others setting, or explain why they choose a different setting.\n\n(5)Minors:\n\n (5.1) wrong formula in line 90: f(x) = exp(q \\cdot k / root(D)), while there is no x in the expression.\n\n (5.2) a latex bug in appendix D.1 (line 860) have not be fixed." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a method to reduce the memory complexity of Kernel Separation in Transformer attention from $O(ND^2)$ to $O(ND)$, where $N$ is the number of tokens and $D$ dimension per attention head. We also introduce an alternative dropout mechanism." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024towards,\ntitle={Towards Making Linear Attention Usable},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y59zhBNKGZ},\nnote={under review}\n}" }, "abstract": { "value": "The original Transformer attention mechanism, based on Softmax, has time and memory complexities of $O(N^2D)$ and $O(N^2)$, where $N$ is the number of tokens and $D$ the dimension per attention head. As current LLM applications trend towards processing larger token sequences, and Transformers gain popularity in image, video, and audio processing, addressing this quadratic cost becomes imperative. Since the introduction of Transformers, numerous approaches have been proposed to linearize this scaling. One such method is Linear Attention, which captures all-to-all token pair attention in $O(ND^2)$ time. However, its drawback lies in its high memory footprint of $O(ND^2)$. While Linear Attention has shown promise in small-scale benchmarks, the high memory demand has prevented Linear Attention to be studied in context of large benchmarks and practical use cases. In this work, we demonstrate how to reduce the memory complexity to $O(ND)$ by approaching calculations from a novel perspective. Additionally, since Linear Attention does not compute the attention matrix directly, it precludes the use of traditional dropout. To address this, we introduce an alternative dropout mechanism. Our study confirms linear scaling in both wall-clock time and memory usage. We also compare our method with Flash Attention and conduct an ablation study on our proposed dropout alternative." }, "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": [ "Linear Attention", "Kernel Separation", "Transformers", "Memory Reduction" ] }, "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/d8f7f3dd6f560ff14c6886d10f6246651f21ea96.pdf" }, "presentation": null, "primary_area": { "value": "infrastructure, software libraries, hardware, systems, 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": "Towards Making Linear Attention Usable" }, "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": "- With the work SIREN in mind, I wonder if the Gaussian basis is the best choice?\n- The idea is quite similar to 3D Gaussian splatting in rendering. However, there are differences. One easy way to explain is that, we can think of the volume rendering equation as a specific type of governing physics equation. Then the success of applying parametric mixed Gaussian in one specific type of equations doesn't mean it is the best idea for others. For example, there can be long-range interactions for certain PDEs, while Gaussian functions are local and lack such representation capability. I would like to have the authors' opinions on such comments." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- A theoretical proof of the universal approximation capability for the proposed method is provided.\n- The experiments are diverse, the baselines are sufficient, and the ablation study is detailed.\n- The writing is overall clear and detailed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors use deep Gaussian mixture models as parametric mesh representations to replace MLP in PINN, so that it performs better for high-frequency and non-linear components. Experiments for PDEs such as Allen-Cahn, Helmholtz, Nonlinear diffusion, etc., are conducted compared with several previous PINN derivatives, showing that the proposed method achieves sota solution accuracy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Though diverse and rich, the experimental settings are all relatively simple, and lack comparisons to traditional methods.\n- Correct me if I am wrong, but I think only the solution accuracy is provided, without the computation time?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- I would be interested to see how the authors address the weaknesses above. Could authors provide additional evidence or context to support the novelty? \n- Have the authors conducted any experiments to demonstrate how the method alleviates spectral bias?" }, "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. The code and experiments are of good quality. It provides a more concise and more parameter efficient feature embedding method compared to previous parametric grid representations method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose the Physics-Informed Gaussian (PIG) that learns feature embeddings of input coordinates, using a mixture of Gaussian functions. Compared to PINN, this paper applies Gaussians to input coordinates and makes the mean and covariance matrix learnable. The locality of Gaussians encourages the model to capture high-frequency details by effectively alleviating spectral bias." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The primary concern lies in the originality of this method. Though it’s considered as a feature embedding, mathematically, it is in the same spirit of a KAN layer with RBF. Also, it is not uncommon to consider Gaussian embedding in ML. \n\n- The paper states that the approach \"dynamically adjusts the positions and shapes of the Gaussians during training\" with parameters \"iteratively updated throughout the training process.\" The claims in the paper give an impression of adaptive sampling method based on the training. However, the approach does not actually take into account the residual or the loss function landscape. It would be better to provide a more rigid description on this. Also, in the experiments, the method is compared with different PINN approaches for each PDE, and the authors report the results without conducting experiments by themselves again. Though “the sensitivity of PINN variants to hyperparameters complicates fair comparisons”, it’s still necessary to conduct experiments with same setup and number of parameters to ensure a fair comparison. The average error of PIG is generally higher, but the authors only highlight the best of PIG. PIG is claimed to “enable more accurate and efficient approximations”, but the efficiency is not clear. Hence, it seems a lot of conclusions in this paper are overstated.\n\n- I would like to see how the spectral bias can be alleviated, either through experimental evidence or analytical justification.\n\n- For the 2D and 3D cases, what is the mathematical formulation for the Gaussian embedding? How is the parameter being initialized? Additionally, how are the boundary conditions (BC) being constrained?\n\n- The paper mentions that \"computational costs per iteration of our method are significantly lower than JAX-PI.\" A detailed report of these computational costs would be beneficial." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- (Line 245-) \"Similar to the previous parametric grid methods, which obtain feature embeddings by interpolating only neighboring vertices, this locality encourages the model to capture high-frequency details by effectively alleviating spectral bias\": Is it possible to mathematically prove that \"this locality encourages the model to capture high-frequency details by effectively alleviating spectral bias\" in some sense? This is just a question out of curiosity.\n\n- Is the title of Section 3.2.2 relevant?\n\n- Why is tanh used in the architecture (it is fine if there is no specific reason)? What about sin or swish activations?\n\n- Is there any issues about mode collapses ($\\mu_i \\rightarrow \\mu_j$ for all $i$ and $j$ and/or $\\sigma_i \\rightarrow 0$ or $\\infty$)?\n\n- I think the update scales of $W$ and $\\theta$ differ significantly from those of $\\mu_i$ and $\\sigma_i$. The optimal updates for $\\mu_i$ and $\\sigma_i$ likely depend on the scale of the simulation. Did the authors encounter any issues related to this?\n\n- It may be off-topic, but is the proposed method applicable to high-dimensional PDEs, such as $d\\gtrsim100$? How significant is the computational cost?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The paper is well-written and easy to follow. I enjoyed reading it.\n\n- Motivation is clear.\n\n- Comparisons with previous works are well-discussed.\n\n- The proposed model is supported by both theoretical and empirical evidence.\n\n- Error bars are provided, enhancing the reliability of the experimental results.\n\n- Code is provided for reproducibility." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Physics-Informed Gaussians (PIG) is proposed, an efficient and accurate PDE solver that utilizes\nlearnable Gaussian feature embeddings and a lightweight neural network, counteracting the problems inherent in previous static parametric grid approaches. PIG model achieves competitive accuracy and faster convergence with fewer parameters compared to state-of-the-art methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- (Minor) Figures should be vector images to avoid pixelization.\n\n- While I found no significant weaknesses, there may be room to further expand the discussion on the novelty and its potential impact on future research." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I may miss something but, is there any reason why choose Gaussian functions as feature embedding? Would other embedding functions fail in the proposed framework?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The strength of this paper is to overcomes the limitations of traditional PINNs. The proposed method maintains the benefits of PINNs, improves performance in PDE approximation, and can be applied to various PDEs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Physics-Informed Gaussians (PIGs), which incorporates Gaussian embeddings in the PINNs framework, to eliminate the inductive biases of neural networks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "As mentioned in Discussion and Limitations, theoretical understanding of the convergence properties is lack, although the universality of PIG is discussed in Section 3.3." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024pig,\ntitle={{PIG}: Physics-Informed Gaussians as Adaptive Parametric Mesh Representations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y5B0ca4mjt},\nnote={under review}\n}" }, "abstract": { "value": "The approximation of Partial Differential Equations (PDEs) using neural networks has seen significant advancements through Physics-Informed Neural Networks (PINNs). Despite their straightforward optimization framework and flexibility in implementing various PDEs, PINNs often suffer from low accuracy due to the spectral bias of Multi-Layer Perceptrons (MLPs), which struggle to effectively learn high-frequency and non-linear components. Recently, the parametric mesh representations have been investigated as a promising approach to effectively eliminating the inductive biases of neural networks. However, they often require very high-resolution grids and a large number of collocation points to achieve high accuracy while avoiding overfitting issues. In addition, these are limited by the fixed positions of the mesh parameters, hindering their ability to approximate complex PDEs. To overcome these limitations, we propose Physics-Informed Gaussians (PIGs), which combine feature embeddings using Gaussian functions with a lightweight neural network. Our approach uses trainable parameters for the mean and variance of each Gaussian, allowing for dynamic adjustment of their positions and shapes during training. This adaptability enables our model to optimally approximate PDE solutions, unlike models with fixed parameter positions. Furthermore, the proposed approach maintains the same optimization framework used in PINNs, allowing us to benefit from their excellent properties. Experimental results show the competitive performance of our model across various PDEs, demonstrating its potential as a robust tool for solving complex PDEs." }, "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": [ "Gaussians", "Physics-informed Deep 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/ca455a58a9fa1df1bea38e54158e4710bf6f2244.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/a4ae8d6a4f529c34d8e978c55c597eaff9f4ad8e.zip" }, "title": { "value": "PIG: Physics-Informed Gaussians as Adaptive Parametric Mesh 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": { "value": "We would like to sincerely thank all reviewers for their valuable feedback, and we have decided to withdraw the submission for further improvement." }, "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": 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": "* I personally appreciate the architecture that supports multimodal perception and generation. However, it is true that vision and audio alone do not fully encompass the concept of Any-to-Any; other modalities are also prevalent. It appears that the current architecture lacks sufficient scalability for additional modalities. Introducing new modalities often requires re-training the LLM, which poses a limitation that may not be conducive to building a truly Any-to-Any architecture.\n\n* I would like to understand the nature of the interactions between modalities. For instance, how does modeling and learning within the audio modality affect that of the image modality?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper represents a commendable attempt in the field of multimodal integration. The proposed Vehicle for Enhancing Harmony (VEH) offers valuable insights into facilitating the interaction between perception modality encoding and decoding. Overall, the manuscript is well-written and provides a thorough performance evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduc X-VILA, an omni-modality model designed to extend the capabilities of large language models by incorporating image, video, and audio modalities. X-VILA achieves cross-modality understanding, reasoning, and generation within one model. And related capabilities are evaluated on extensive benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* **Confusion in Structural Design**: According to the authors, the Embedding Highway design is intended to be applicable across various modalities. However, it is only applied to the visual modality, which appears to contradict the authors' objective of constructing a truly Any-to-Any multimodal model.\n\n* **Novelty of the Model**: To my knowledge, the architecture that employs modality-independent encoding and decoding, integrated through a LLM, has been extensively explored in several prior works (e.g., Next-GPT [1]). This architecture does not seem to demonstrate clear novelty. It would be beneficial for the authors to systematically compare their work with these existing studies to highlight its innovative aspects.\n\n[1] Wu S, Fei H, Qu L, et al. NExT-GPT: Any-to-Any Multimodal LLM[C]//Forty-first International Conference on Machine Learning." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Why use Imagebind as all modality encoders? It seems all tasks performs bad, If we change \"unbind encoders\", such as clip for vision, whisper/beats for audio, the performance could definitely be better. Imagebind only sounds better instead of performance better.\n2. Number of compared methods are too small. Only NextGPT? \n3. \" As depicted in Figure 4, we have identified two key capabilities that have surfaced Long-context cross-modality generation. X-VILA exhibits an impressive capacity for comprehending and combining diverse concepts from multiple iterations of input. Consequently, it produces natural and coherent output, as suggested by the users.\" I don't think the context is long enough to be called \"long-context\" in the example of Figure 4.\n4. The first stage training, so-called \"encoder-LLM-Decoder alignment training\" utilizes the \"X-text pairs\" from academic datasets as in prior work, so what does the decoder train? It is so confused." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.The instructional tuning dataset may be helpful for multi-modality generation research.\n2. Proposed method surpasses Next-GPT on compared benchmarks \n3. Proposed Visual Embedding Highway sounds make sense." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose a multi-modality large language model which can model the understanding and generation between audio, image, video and text. It also propose a X-to-X generation instructional tuning dataset. The method is evaluated on multimodal generation benchmark and some general visual/audio QA/caption benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Novelty is limited. Unifing text and visual generation have been studied for a lot, such as Cogview or Emu series. Adding audio modality also have been researched by AnyGPT or NextGPT. So this paper seems more like some performance incresements istead of some new insights, which is not excited enough to me.\n2. Architecture design does not have new insights. Multimodal encoders + projectors + LLM + diffusion decoders are a common way for unifying multimodal understanding and generation. Even though authors emphasize that diffusion models are tuned together which make difference from earlier works, but the architecture still seem not novel and elegant enough. In addition, the visual high way design seems serving for edit-like tasks targetedly instead of a general design. \n3. Performances on visual QA or audio caption benchmarks are too weak." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Is there a planned timeline for releasing the training code and datasets for X-VILA?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The introduction of the Visual Embedding Highway (VEH) to preserve and enhance visual features represents a significant advancement over existing methods in visual alignment.\n2. The paper provides a large X-to-X dataset for multi-modal instruction, which is a valuable resource for advancing research in multi-modal modeling.\n3. The authors plan to open-source X-VILA and the dataset, promoting academic and engineering research in multi-modal foundation models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces X-VILA, a model capable of handling various input and output modalities, including video, image, language, and audio. X-VILA employs a two-step alignment mechanism involving textual alignment and a novel visual alignment component called the Visual Embedding Highway (VEH). This mechanism addresses common challenges in multi-modality alignment, such as loss of visual information. The authors contribute a large-scale X-to-X multi-modal instruction dataset to support further research in multi-modal models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Missing Example Component: The example in Figure 1 lacks the \"video\" input component, which disrupts the clarity and completeness of the illustration.\n2. High Computational Requirements: The model’s training process is resource-intensive, potentially limiting accessibility and reproducibility, particularly for smaller research teams." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 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": "The introduction of the Visual Embedding Highway (VEH) module to preserve visual details is an original contribution that enhances the model's performance in visual tasks. \n\nThe paper is of high quality, as evidenced by its thorough methodology, comprehensive experiments, and rigorous evaluation. \n\nThe paper is well-structured and clearly written." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces X-VILA, an omni-modality model designed to extend the capabilities of large language models (LLMs) by incorporating image, video, and audio modalities. X-VILA achieves cross-modality understanding, reasoning, and generation by aligning modality-specific encoders with LLM inputs and diffusion decoders with LLM outputs. The key contributions of the paper are:\n\n- It presents a new family of any-to-any modality chat LLMs capable of conducting multi-modality conversations, understanding signals from different modalities, and generating content in various formats, including video, audio, image, and text.\n- The paper introduces a new X-to-X multi-modality instruction tuning dataset, which has proven effective for cross-modality alignment. \n- To address the issue of visual information loss, a visual alignment mechanism with a visual embedding highway module is introduced, which allows visual features to bypass the LLM, enhancing the correspondence of visual content between input and output stages.\n- X-VILA demonstrates emergent properties across modalities even in the absence of similar training data, showcasing abilities like long-context cross-modality generation and new types of cross-modality ability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- How do the authors ensure the data quality of the proposed X-to-X dataset, and what is the specific process for its creation?\n\n- Why is the additional module VEH implemented solely in the visual component?\n\n- Compared to Next-GPT, the method in this paper utilizes a larger instruction tuning dataset. Could the authors provide experimental results demonstrating the impact of scaling up the instruction data?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nye2024xvila,\ntitle={X-{VILA}: Cross-Modality Alignment for Large Language Models},\nauthor={Hanrong Ye and De-An Huang and Yao Lu and Zhiding Yu and Wei Ping and Andrew Tao and Jan Kautz and Song Han and Dan Xu and Pavlo Molchanov and Hongxu Yin},\nyear={2024},\nurl={https://openreview.net/forum?id=y5G1BfV7Am}\n}" }, "abstract": { "value": "We introduce X-VILA, an omni-modality model designed to extend the capabilities of large language models (LLMs) by incorporating image, video, and audio modalities. By aligning modality-specific encoders with LLM inputs and diffusion decoders with LLM outputs, X-VILA achieves cross-modality understanding, reasoning, and generation. To facilitate this cross-modality alignment, we curate an effective interleaved any-to-any modality instruction-following dataset. Furthermore, we identify a significant problem with the current cross-modality alignment method, which results in visual information loss. To address the issue, we propose a visual alignment mechanism with a visual embedding highway module. We then introduce a resource-efficient recipe for training X-VILA, that exhibits proficiency in any-to-any modality conversation, surpassing previous approaches by large margins. X-VILA also showcases emergent properties across modalities even in the absence of similar training data. The project will be made open-source." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Hanrong_Ye1", "~De-An_Huang1", "~Yao_Lu13", "~Zhiding_Yu1", "~Wei_Ping1", "~Andrew_Tao1", "~Jan_Kautz1", "~Song_Han5", "~Dan_Xu4", "~Pavlo_Molchanov1", "~Hongxu_Yin2" ] }, "authors": { "value": [ "Hanrong Ye", "De-An Huang", "Yao Lu", "Zhiding Yu", "Wei Ping", "Andrew Tao", "Jan Kautz", "Song Han", "Dan Xu", "Pavlo Molchanov", "Hongxu Yin" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multi-task learning", "vision-language models", "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": { "value": "ye|xvila_crossmodality_alignment_for_large_language_models" }, "pdf": { "value": "/pdf/30af5cc7feabdffec617f6dc7e5d1af6366c4472.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": "X-VILA: Cross-Modality Alignment for Large Language 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": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Could the author better clarify the key novelty of GOLD compared to existing works? What distinguishes the usage of the generator in GOLD from previous works?\n- The training procedure has two stages: one stage involves fixing the GNN and training the LGM, and the second stage involves fixing the LGM and training the GNN. Is it possible to combine these two stages using a gradient reversal layer?\n- Further ablative studies on $L_{DReg}, L_{Unc}, L_{EReg}$ (removing each of them or applying each individually) may help to better demonstrate the effectiveness of the proposed new divergence regularization (Eq. 12).\n- What modifications would be needed to extend the framework beyond node-level detection, such as to graph-level OOD detection tasks?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- *No OOD Data Required*: Synthesizes pseudo-OOD data through adversarial training, alleviating the need for real OOD samples.\n- *Implementation Flexibility*: Supports both LDM and VAE variants, offering trade-offs between performance and computational efficiency.\n- *Strong Empirical Performance*: Outperforms non-OOD methods and matches/exceeds methods using real OOD data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a framework designed to detect out-of-distribution data without requiring pre-existing OOD datasets or pre-trained models. The proposed GOLD framework includes:\n- A latent generative model to generate synthetic embeddings that imitate in-distribution embeddings from a GNN.\n- A GNN encoder and OOD detector to classify in-distribution data and maximize energy divergence between in-distribution and synthetic embeddings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Using generative models to generate samples for downstream tasks has been widely adopted in previous methods. For example, generated samples are commonly used in continual learning for experience replay. It seems that the use of generators in GOLD applies the same concept to different tasks. Energy-based detection is also widely used. The major contribution seems to lie in a new divergence regularisation (Eq.12).\n- Further ablative studies on divergence regularisations may be needed to better reflect their effectiveness.\n- The presentation needs to 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": 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. Regarding the generation process of pseudo-OOD samples based on ID data, if these generated samples are overly close to the ID distribution, it could lead to confusion in the OOD detector, potentially causing ID samples to be misclassified as pseudo-OOD. Could the authors elaborate on whether any specific techniques were applied during training to control this distributional proximity?\n2. Additionally, I noticed that the model exhibits substantial variance in the FPR95 metric on the Amazon and Coauthor datasets. Could the authors clarify whether this variance is linked to the aforementioned distributional control during the generation process?" }, "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 method with notable novelty, particularly in how it handles the generation of OOD (Out-of-Distribution) samples based on ID (In-Distribution) data. This approach demonstrates creative problem-solving and provides a potentially valuable contribution to OOD detection research.\n2. The experimental section is comprehensive, covering multiple datasets and providing a range of performance metrics. This thorough evaluation supports the robustness of the method and suggests that it may perform effectively across diverse scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the GOLD framework, an innovative implicit adversarial learning approach for graph OOD detection that bypasses the need for additional OOD data or pretrained generative models. GOLD employs an alternating optimization scheme, utilizing a latent generative model and a GNN encoder to produce pseudo-OOD embeddings for training, enhancing robustness against OOD instances. Extensive experiments across five benchmark graph datasets demonstrate the framework’s competitive performance compared to state-of-the-art OOD detection methods, even without true OOD data. Minor issues, such as a blank figure and lack of code availability, detract slightly from the work, but overall, the paper makes a valuable contribution to graph-based OOD detection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Certain issues in the manuscript reduce its overall clarity and precision. For example, Figure 1(d) appears to be blank, which may hinder understanding and interpretation of the paper’s content." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. What does subscript (i.e., [0], [1]) in Eq(10) mean?\n\n2. Have the authors tried other backbone models?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- GOLD’s adversarial latent generation is a novel approach that synthesizes pseudo-OOD data without auxiliary datasets, making it efficient and broadly applicable.\n\n- GOLD’s effectiveness is demonstrated through comprehensive experiments on five benchmark datasets, showing its robustness across various graph types and OOD scenarios.\n\n- The implicit adversarial objective and energy-based detection approach lead to a clear divergence between ID and OOD embeddings, validated by experimental visualizations.\n\n- This paper is well-structured." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents GOLD, a novel framework aimed at enhancing Out-of-Distribution (OOD) detection in graph neural networks (GNNs) without relying on external OOD data. GOLD introduces an implicit adversarial training pipeline where a latent generative model (LGM) is trained to generate embeddings that mimic in-distribution (ID) data, while an OOD detector is optimized to increase divergence between ID embeddings and these synthetic pseudo-OOD embeddings. It effectively simulates OOD exposure, helping the model distinguish OOD nodes in graph data. Extensive experiments demonstrate that GOLD outperforms state-of-the-art methods in OOD detection across several benchmark datasets without the need for real OOD data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper has no obvious weaknesses except for the training computational cost induced by the pseudo-OOD data. However, this cost increase is acceptable. And the authors have also discusses this issue in POTENTIAL LIMITATIONS 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": 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": "See weakness." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The idea of implicit adversarial learning is novel.\n2. The experiments are solid and comprehensive, especially the improvements achieved on the FPR95 dataset is impressive.\n3. The paper is easy to follow and well-structured. The writing is good. The theoritical proofs are provided." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes the GOLD, a novel framework for graph OOD detection that employs implicit adversarial learning to generate synthetic OOD instances without the need for pre-trained models or additional OOD data. By utilizing a latent generative model to produce embeddings mimicking in-distribution data, which are then differentiated from actual in-distribution embeddings by a graph neural network encoder and an OOD detector, the method is able to effectively simulate the OOD exposure. The framework is evaluated on five benchmark graph datasets, demonstrating superior OOD detection performance without using real OOD data compared with the state-of-the-art OOD exposure and non-exposure baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I didn't see obvious weaknesses. Here I provide some suggestions:\n\nCould you provide a more in-depth analysis of the performance achievements on the FPR95 dataset, including an analysis from the perspective of the dataset's characteristics? Additionally, could you explain why similar significant improvements were not observed on other datasets (although the improvements on other datasets are also quite good)?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024gold,\ntitle={{GOLD}: Graph Out-of-Distribution Detection via Implicit Adversarial Latent Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y5einmJ0Yx},\nnote={under review}\n}" }, "abstract": { "value": "Despite graph neural networks' (GNNs) great success in modelling graph-structured data, out-of-distribution (OOD) test instances still pose a great challenge for current GNNs. One of the most effective techniques to detect OOD nodes is to expose the detector model with an additional OOD node-set, yet the extra OOD instances are often difficult to obtain in practice. Recent methods for image data address this problem using OOD data synthesis, typically relying on pre-trained generative models like Stable Diffusion. However, these approaches require vast amounts of additional data, as well as one-for-all pre-trained generative models, which are not available for graph data. Therefore, we propose the GOLD framework for graph OOD detection, an implicit adversarial learning pipeline with synthetic OOD exposure without pre-trained models. The implicit adversarial training process employs a novel alternating optimisation framework by training: (1) a latent generative model to regularly imitate the in-distribution (ID) embeddings from an evolving GNN, and (2) a GNN encoder and an OOD detector to accurately classify ID data while increasing the energy divergence between the ID embeddings and the generative model's synthetic embeddings. This novel approach implicitly transforms the synthetic embeddings into pseudo-OOD instances relative to the ID data, effectively simulating exposure to OOD scenarios without auxiliary data. Extensive OOD detection experiments are conducted on five benchmark graph datasets, verifying the superior performance of GOLD without using real OOD data compared with the state-of-the-art OOD exposure and non-exposure baselines. 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": [ "Graph Neural Network", "Out-of-Distribution Detection" ] }, "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/d6f9e37d478eeacddcdbc2225a3f39eb5b4158a6.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "GOLD: Graph Out-of-Distribution Detection via Implicit Adversarial Latent 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": 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": "Q1. Can this approach work reliably in the Zero-shot coordination (unseen agents) setting? \n\nQ2. Does the success rate on Overcooked increase with more time steps? \n\nQ3: Is it possible to utilize a different LLM with a lower cost to analyze the ReAd-S of this method in the Overcooked-AI setting? Are LLMs other than GPT-4 able to perform low-level control in Overcooked where the action space involves only navigation and interaction? \n\nQ4: Why isn't ReAd-S compared with other methods using the Llama-3.1-70B model?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "S1: The method introduces a method to eliminate the need for direct physical interaction with the environment, instead utilizing a pre-trained critic to approximate an advantage score and internally replan the action. \n\nS2: The authors provide an extensive theoretical justification and guarantee for advantage decomposition in the multi-agent setting and empirically demonstrate the success of their method on two different benchmarks in multiple settings \n\nS3: Their method is able to generate direct low-level actions without the need for intermediate high-level action plans which is a significant advantage over previous methods\n\nS4: The method does not require access to exact probability of the sampled action from LLMs and can be used with closed-source LLMs" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new method for grounding LLMs for collaborative embodied multi-agent applications that utilizes a critic to score advantage values of actions based on the actions of partner agents. The critic provides advantage score for action predictions of the LLM to eliminate the need for direct physical interaction with the environment at inference time. They provide a detailed theoretical justification for the method and demonstrate superior grounding and coordination in the RoCo-Bench and Overcooked-AI environments compared to LLM-based baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1: Lack of discussion about generalization to unseen partners. In practical coordination scenarios, partner agents might not employ the same algorithms as the ego-agent. It is essential for coordination methods to be robust to unseen partners.\n\nW2: Limited test coverage on Overcooked AI benchmark - it looks like Reflexion and React have 0 % task completion rate. Experiments should consider more time steps as it might turn out that although Reflexion or React take more time, they could have a higher success rate. Additionally, Overcooked also has a multiple delivery objective, as performance in the second delivery might be improved by better time management during the first delivery. Testing only one delivery is insufficient.\n\nW3: The method utilizes a critic model that is trained in the environment. Whereas the baselines are all training-free approaches which might be an unfair comparison.\n\nW4: ““Choosing w at the current state s signifies all agents take no actions, then the next state s′ = s and the agents receive shared reward r(s, w) = 0 since w bring no changes to the environment.” is a strong assumption as wait states can lead to negative rewards in the environment depending on the task configuration.\n\nW5: The Overcooked-AI environment should be evaluated with ReAd-S to ensure the reliability of the method in the decentralized 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": 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": "The paper presents an innovative approach by leveraging MARL theories to enhance collaborative behavior in LLM-based agents within multi-agent environments. The theoretical foundation is well-developed, allowing readers to gain a deep understanding of the approach. Furthermore, the experimental results and their detailed analysis contribute valuable insights to the research community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel feedback mechanism, Reinforced Advantage feedback (ReAd), which combines the self-improvement abilities of large language models (LLMs) with advantage decomposition techniques from multi-agent reinforcement learning (MARL). This closed-loop feedback strategy with the introduce of a critic module enhances cooperation in multi-agent settings by training LLMs to learn a sequential advantage function using critic regression. Experiments on two multi-agent benchmarks indicate that ReAd outperforms existing feedback-based baselines and state-of-the-art methods, highlighting its effectiveness in promoting collaborative strategies among LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Since the paper takes insights from marl theories, including some marl methods as a baseline will further strengthen the robustness of the results\n\n- Even with the _Difficult Variants_ of RoCoBench, the proposed method has achieved a nearly 100 percent success rate, so it seems the task may still be too easy or oversimplified.\n\n- The experimental results showing a 0% success rate for all baselines on the overcooked-ai benchmark need further explanation. The original overcooked-ai paper reported RL baselines with non-zero performance—why aren’t they included here?\n\n- As noted by the authors, there is a simplification of the visual perception and the generalizability of ReAd to more complex scenarios remains uncertain." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 argued that multimodal VLMs such as GPT4o don't always parse the scene reliably. Have you tried query GPT4o with both the agent prompt (shown in Fig 1) and image input and use GPT4o as the planner instead of the image parser? \n\n2. Have you thought about using another LLM as a critic?\n\n3. How does ReAd-J compare to existing RL baselines on Overcooked?\n\n4. It looks like none of the other LLM baselines succeeded in a single episode in Overcooked. This looks strange. Could you explain why they failed? Is it because they didn't complete the task within a fixed number of timesteps or they actually never completed the task? How was the timestep limit set? Does the results change if you increase the limit?\n\n5. Did you try the models on the original RoCoBench? Why create a new dataset instead of using the original one, which has been tested/used by many existing studies?\n\n6. Do other LLM baselines (MindAgent, React) have agent communication in your DV-RoCoBench experiment?" }, "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 incorporating some critic in the LLM planner is promising and can mitigate hallucination among the LLMs. It's an interesting and important problem for the AI/LLM community.\n\nThe paper is well written and easy to understand. The experiment is well done and the analysis is thorough and detailed. \n\nI particularly appreciate the extensive discussions and analyses presented in the appendix, as they openly address the model's robustness and limitations. This level of detail not only strengthens the credibility of the research but also provides valuable insights for future work in this area.\n\nOverall, I applaud the authors for this well-written work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel algorithm Reinforced Advantage feedback (ReAd), which leverages LLM for multiagent planning. The paper tests the algorithm on two domains, RoCoBench and Overcooked. The results show that ReAd outperformed existing LLM-based baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Model: The main weakness of the proposed algorithm is its limited generalizability. It seems like the critic NN is trained over hundreds of plans/trajectories in each domain. If this is the case, then this model becomes very domain-specific and would need a new critic for each new domain. That defeats the purpose of using LLM as a planner, which is meant to be fast and easily adaptable to different domains so you don't have to build any domain-specific model. I wonder if you can use LLM (maybe with few-shot prompting) as a critic?\n\n2. Baseline choice: The paper only considered LLM baselines. Since the model is essentially doing RL, why not include RL baselines as well? If you are already training the critic to estimate some reward functions, then why do you need the LLM planner at all? Or at least you can try a version where you use RL for planning and LLM for communication among agents. There're also a lot of existing work on Overcooked AI. Does your model outperform those baselines?\n\n3. Baseline implementation: If I recall correctly, LLM agents such as MindAgent and React don't have communication modules. I encourage the authors to include more details on how these baselines are implemented. If these baselines don't actually allow communication among agents in DV-RoCoBench, then it's not a fair comparison.\n\n4. Evaluation: The paper would be much stronger if it would include a human experiment/evaluation, which was used in MindAgent and RoCo.\n\n5. Presentation: It would make the results more readable if you can put ReAd-J and ReAd-S next to each other in Fig 3." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In Line 38, I wonder about the author's mentions of \"lack of task-specific knowledge\" instead of domain-specific or environment-specific knowledge. LLMs possess commonsense actionable knowledge (i.e., knowledge to perform tasks) but have never learned the characteristics of the actual environment where this knowledge should be applied.\n\n2. What is the size of dataset \\mathcal{D}?\n\n3. Rather than directly fine-tuning the LLM, it uses prompting to employ the LLM as an advantage function optimizer. How is the policy iteration implemented where (i) the LLM (behavior policy) is fine-tuned using the advantage function learned from dataset \\mathcal{D}, and (ii) this fine-tuned policy becomes the behavior policy to collect new data? Does dataset collection by the behavior policy occur only once?\n\n4. What is the neural network architecture of the advantage function?\n\n5. Are environment observations converted to text before being passed to the LLM agent?\n\n6. Does using the Advantage function (or Q-function) directly as a policy significantly decrease performance? I'm curious about how a conventional discrete action space RL approach would perform, where Q-values are calculated for all actions and the action with the highest Q-value is selected." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Grounding LLM's reasoning prowess in physical environments is a highly promising and crucial field that can help leverage continuously evolving LLMs for real-world problem solving.\n\n2. Addressing interaction costs in embodied environments is a challenging yet practical problem that can resolve various factors, including safety issues and LLM inference costs.\n\n3. The paper is well-organized and written clearly without confusion, making it easily accessible to readers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work discusses grounding LLM's common sense reasoning capability in physical environments. Specifically, it explores methods to not only generate appropriate actions but also effectively increase interaction efficiency when using Multi-Agent LLM agents in embodied environments. Based on the theoretical background of multi-agent systems, the authors utilize a learned advantage function as immediate feedback for LLM agents. The LLM is prompted to generate actions with high advantage values, and if the generated action plan has low advantage values, it is designed to perform replanning autonomously. The superiority of this framework is demonstrated through experiments across multiple environments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main concern is the novelty of the contribution. Applying MARL's theoretical aspects to embodied multi-agent collaboration appears quite straightforward, and seems independent from the authors' stated goal (as written in Line 77) of enhancing LLMs' reasoning capabilities.\n\n2. While the introduction addresses interaction efficiency as a major issue in prior work, the data collection efficiency of the proposed learning method should also be discussed. This TRPO-style on-policy learning approach is generally known to be less data-efficient than off-policy RL algorithms (even in single-agent settings).\n\n3. I am concerned if the comparison between baselines is fair. For instance, was the same amount of environmental interaction required to create ReAd's training dataset also provided to ReAct and Reflexion?\n\n4. If data collection through interaction is possible, I think the current environmental tasks are relatively simple. For example, using sparse rewards (success/fail) attached to the end of collected dataset trajectories for retrieval-augmented task planning could enhance LLM reasoning more simply and efficiently than the proposed method. Please refer to paper [1]. Could the authors include [1] as a baseline in their experimental comparisons?\n\n[1] LLM-Planner: Few-Shot Grounded Planning for Embodied Agents with Large Language Models, CVPR 2023" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present a novel and theoretically supported feedback for closed-loop LLM planning in multi-agent collaboration." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024towards,\ntitle={Towards Efficient {LLM} Grounding for Embodied Multi-Agent Collaboration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y5tkxH7kxQ},\nnote={under review}\n}" }, "abstract": { "value": "Grounding the reasoning ability of large language models (LLMs) for embodied tasks is challenging due to the complexity of the physical world. Especially, LLM planning for multi-agent collaboration requires communication of agents or credit assignment as the feedback to re-adjust the proposed plans and achieve effective coordination. However, existing methods that overly rely on physical verification or self-reflection suffer from excessive and inefficient querying of LLMs. In this paper, we propose a novel framework for multi-agent collaboration that introduces Reinforced Advantage feedback (ReAd) for efficient self-refinement of plans. Specifically, we perform critic regression to learn a sequential advantage function from LLM-planned data, and then treat the LLM planner as an optimizer to generate actions that maximize the advantage function. It endows the LLM with the foresight to discern whether the action contributes to accomplishing the final task. We provide theoretical analysis by extending advantage-weighted regression in reinforcement learning to multi-agent systems. Experiments on Overcooked-AI and a difficult variant of RoCoBench show that ReAd surpasses baselines in success rate, and also significantly decreases the interaction steps of agents and query rounds of LLMs, demonstrating its high efficiency for grounding LLMs. More results are given at \\url{https://read-llm.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": [ "LLM planning", "Large Language Models", "Multi-Agent Collaboration" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/a30f6c123365d3adf6a7eecad21feaf4d53e6503.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": "Towards Efficient LLM Grounding for Embodied Multi-Agent Collaboration" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Given the increasing use of Large Language Models (LLMs) and Deep Reinforcement Learning (DRL) in quantitative finance, does the team have plans to integrate these techniques into QuantBench in future versions?\n\n2. Would the authors consider expanding the platform to include risk-evaluation metrics? Could the author provide insights into the technical feasibility of incorporating stress testing in QuantBench, perhaps by simulating market shocks based on historical crisis periods?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Originality: The paper tried to provide a standardized, open-source environment specifically designed for quantitative finance, which supports quant trading tasks from factor mining to order execution, representing an integration of existing modeling methods with advanced data processing and evaluation mechanisms.\n\n2. Quality: The quality of the work is OK, providing comparison across multiple dimensions, including model architecture, training objectives, and validation strategies.\n\n3. Clarity: The paper is well-organized and explains the design and objectives of QuantBench.\n\n4. Significance: The platform currently focuses on established techniques and common performance metrics, providing a useful standardized framework for evaluating models in quantitative finance, and its significance might be impactful within the industry." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents QuantBench, an industrial-grade benchmark platform designed to standardize and enhance AI research in quantitative finance. The platform integrates diverse data sources, including market, fundamental, relational, and news data, and providing a layered approach that spans alpha factor mining, modeling, portfolio optimization, and order execution. It also supports multiple data resolutions, from quarterly to tick-level data, facilitating a range of quant tasks and enabling multi-scale strategies. \n\nQuantBench categorizes models based on architectural design and training objectives, evaluating temporal models for time-series data and spatial models for relational data. It also examines the impact of training objectives, including classification, IC, MSE, and ranking losses, finding that different objectives yield varied results based on model architecture.\n\nExperiments demonstrate that model updating frequency, validation set construction, and ensemble methods significantly impact model performance, particularly in managing alpha decay and reducing overfitting, which suggests future directions in continual learning, robust validation strategies, and ensemble diversity to further improve predictive accuracy and model stability.\n\nQuantBench aims to provide a standardized and open-source platform that fosters collaboration and bridges the gap between academic research and practical industry applications in AI-driven quantitative investment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper's originality may be somewhat limited in terms of introducing novel methodologies or fundamentally new problem formulations. Although this application to quant finance is useful, the overall concept of a benchmarking platform is not entirely new. Moreover, the paper is limited by its focus on classical machine learning and deep learning algorithms, without including more recent advancements in AI specifically tailored to quantitative finance, such as Large Language Models (LLMs) and Deep Reinforcement Learning (DRL).\n\n2. The paper’s focus on predictive accuracy and return metrics is valuable, but it lacks depth in evaluating risk management, a crucial aspect of quantitative trading. Since QuantBench aims to align with industry standards, it would benefit from incorporating a broader set of risk metrics beyond Sharpe ratios—such as maximum drawdowns, downside risk, or conditional value-at-risk (CVaR)—to provide a more comprehensive assessment of model robustness.\n\n3. The platform currently focuses on established techniques and common performance metrics, may not fully address the rapidly evolving needs of practitioners and researchers looking for innovative solutions in quantitative trading." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Dear authors, \n\nWe consider the work interesting and relevant. Nevertheless, we would like to point out certain improvement opportunities.\n\nGENERAL COMMENTS\n\n(1) - \"The results in Table 2 indicate that DNNs generally outperform tree-based models in IC\" -> The authors only compare an LSTM model against an XGBoost model. We consider the claim to be too strong for the referenced results. To strengthen the claim, we encourage the authors to conduct experiments with additional tree models (e.g., random forest, LightGBM, CatBoost, ...) and contrast them to the results for deep learning models mentioned in Table 3.\n\n(2) - We encourage the authors to include some statistical tests to assess whether the differences among results are statistically significant. This could be an important step in the benchmark when analyzing the results.\n\n(3) - The authors mention three task-agnostic metrics (robustness, correlation, decay) but it is unclear where they were reported and assessed when considering the results obtained.\n\n(4) - The authors provide a timeline of models covered by QuantBench in Figure 3. Nevertheless, the experiments do not cover some of the most recent ones. Would it be possible to include them?\n\nTABLES\n\n(5) - All tables reporting metrics: (a) provide arrows next to the metric names (up/down) indicating whether a higher/lower result is better; (b) bold the best results, (c) align numbers to the right, to make differences of magnitude evident to the reader.\n\nMINOR COMMENTS\n\n(6) - \"charaterstic-sorted\" -> \"charateristic-sorted\"\n\n(7) - \"Wikidata’s inclusion yielded minimal improvements, possibly because the information is already widely known and exploited by other market participants.\" -> Please provide some additional context to understand how the fact that other market participants could exploit such information affects the results or the data used to train the model.\n\n(8) - \"QuantBench is designed to provide a holistic view of the full quant research pipeline, rather than focusing on specific techniques such as reinforcement learning\" -> It seems that the current version of the proposed QuantBenchmark focuses only on a specific type of methods (e.g., no reinforcement learning methods were considered). We encourage the authors to provide a more detailed perspective on what is currently supported and the ambition to support in the future. Are there some kinds of models that will not be considered in the future (e.g., reinforcement learning)?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- the authors propose a novel benchmark for artificial intelligence methods on quantitative investment\n- they compare the proposed benchmark with existing SOTA benchmarks\n- they perform a comprehensive evaluation of machine learning methods related to quantitative investment" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce a new industrial-grade benchmark with three goals in mind: (a) alignment with quantitative investment industry practices, (b) flexibility to integrate artificial intelligence algorithms, and (c) provide a pipeline that covers the whole quantitative investment process. They also perform an empirical study that reveal some new research directions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- the authors make some claims that are not backed by substantial experimentation\n- the authors should include some assessment regarding the statistical significance of the results obtained" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "None" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "see above" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides a benchmark platform for AI methods in quantitative investment. It offers as strength \"standardization with quantitative industry practices\", which is not a strength in general. It might be an advantage for transferability of results, but in general industry standards can restrict innovation and are often rather complex. Another strength argued for is the integration of various AI algorithm, which should be considered a minimum requirement rather than a key strength. The last strength, i.e., full-pipeline coverage is more interesting, as other benchmarks often only provide data, putting the burden on pre-processing on the researcher, which also makes reproducibility more challenging. The paper focuses on non-GenAI (though it does assess transformers), which is ok, though it is certainly a restriction, as from personal experience I know a number of investors that use, e.g., ChatGPT for the better or worse. That is, the approach of feature engineering is very heavily embedded in the platform (e.g., they provided extracted features from news), though overall receiving less and less attention in AI. Also the data is only vaguely described in Section 2 and 3, leaving the reader puzzled about the basic composition of the benchmark. \n\nA major concern is that the paper fails to discuss any ethical (and regulatory) concerns, such as bias (or compliance, which is huge topic in industry, which the seem to aim at)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "see above" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "None." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to the question listed in the weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors effectively convey the importance and necessity of having a benchmark platform like QuantBench in the realm of quantitative finance. \n\nThe abstract and introduction are well-structured and clearly articulate the research goals and the authors' approach, providing a solid foundation for understanding the study's scope and significance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose the QuantBench, which is a platform claimed to boost quant research efficiency by eliminating burdensome preprocessing tasks, enabling researchers to focus on algorithmic innovations. They state that it serves as a bridge between academia and industry, facilitating the implementation of advanced algorithms in investment strategies and enhancing practical applications of academic research. Through standardization, they believe that QuantBench can improve communication between these sectors, accelerating AI advancements in quantitative investment. They also conducted multiple empirical studies using QuantBench to explore key research directions, including addressing quant data's distribution shifts, evaluating graph structures, and assessing the real-world applicability of neural networks versus tree models. The outcomes of these studies highlight the needs for robust modeling approaches, such as training models with inherent diversity or implementing causal learning methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) The benchmark incorporates models that do not include some state-of-the-art (SOTA) approaches. Although the authors present a clear roadmap for model development in Figure 3, the experiments only include a limited selection of earlier models. Plus, there are more recent related work which can be included in the model comparison, like [1]. Could you please clarify why the SOTA methods are excluded from your model comparisons? The absence of these comparisons leaves the derived insights less compelling. \n\n[1] Wang, S., Yuan, H., Zhou, L., Ni, L. M., Shum, H. Y., & Guo, J. (2023). Alpha-gpt: Human-ai interactive alpha mining for quantitative investment. arXiv preprint arXiv:2308.00016.\n\n(2) While Section 3 of the paper extensively describes the data collection process for the benchmark, the specifics regarding the data inputs for each model type outlined in Section 6.2 remain ambiguous and potentially inconsistent. It is noted that the authors claim their selected models utilize textual feature inputs; could you provide more detailed explanations of how models categorized under 'Tabular' incorporate these textual features? Additionally, the paper mentions that graph models use graphical features as partial inputs, yet it lacks essential details. It is crucial to include information on whether other comparative models also utilize graphical features, how they do so, and where one can find details about the graph construction to assess the quality of your graph generation.\n\n(3) And model categorization in Table 3 lacks strong reference. Would you please provide the corresponding academic reference for your model classification?\n\n(4) Section 6.4 appears to lack crucial details, as the authors do not provide the necessary context for alpha decay, such as its typical usage and the commonly significant levels in the finance domain. Additionally, although multiple models are discussed in the previous section, it is unclear which model is used to create the visualization in Section 6.4. The authors have not clearly specified this. It would be beneficial for the clarity and completeness of the analysis if these points were addressed in the text.\n\n\n(5) The shadow color choices used to represent 1 standard deviation for the two types of returns could be improved by using different colors. Currently, it is challenging for readers to distinguish between them. Using distinct colors would enhance clarity and make it easier to differentiate the returns visually.\n\n(6) Given that the paper presents multiple studies to claim the applied scenarios of QuantBench from various perspectives, the derived insights are dispersed and challenging to follow. It is essential to include a conclusion section that consolidates these insights into a summarizing paragraph. This section should discuss how QuantBench can enhance real-world investment strategies and identify the unresolved questions that persist. Such a conclusion will provide a comprehensive overview, highlighting the platform’s practical implications in the industry and directing future research towards addressing the remaining challenges." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A systematic benchmarking platform for AI-driven quantitative investment" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024quantbench,\ntitle={QuantBench: Benchmarking {AI} Modeling for Quantitative Investment},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y6wVRmPwDu},\nnote={under review}\n}" }, "abstract": { "value": "The field of artificial intelligence (AI) in quantitative investment has seen significant advancements, yet it lacks a standardized benchmark aligned with industry practices. This gap hinders research progress and limits the practical application of academic innovations. We present QuantBench, an industrial-grade benchmark platform designed to address this critical need. QuantBench offers three key strengths: (1) standardization that aligns with quantitative investment industry practices, (2) flexibility to integrate various AI algorithms, and (3) full-pipeline coverage of the entire quantitative investment process. Our empirical studies using QuantBench reveal some critical research directions, including the need for continual learning to address distribution shifts, improved methods for modeling relational financial data, and more robust approaches to mitigate overfitting in low signal-to-noise environments. By providing a common ground for evaluation and fostering collaboration between researchers and practitioners, QuantBench aims to accelerate progress in AI for quantitative investment, similar to the impact of benchmark platforms in computer vision and natural language processing." }, "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 learning", "quantitative investment" ] }, "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/552d6354a282b59c5939750f7e01443e403a90fc.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/240b79b6213ec4802a7816bb5d23fdedbb656cbf.zip" }, "title": { "value": "QuantBench: Benchmarking AI Modeling for Quantitative Investment" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- [Q1] When fusing the graph and geometry features prior to feeding them to the unified encoder, is there any mechanism used to distinguish tokens from the 2D graph and those from the 3D geometry?\n- [Q2] The overall pipeline highly resembles MultiMAE [B] in computer vision, where a non-uniform masking strategy was used by sampling probabilities from a Dirichlet distirbution, then masking each modality with its corresponding probability. Have the authors considered non-uniform masking probabilities as such?\n- Small typo in Line 271: did you mean \"concatenate\" instead of \"contact\"?\n\n[D] Bachmann et al., MultiMAE: Multi-modal Multi-task Masked Autoencoders. ECCV 2022." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- [S1] **Simplicity of approach.** Each component of MolCoMA is simple and is presented clearly such that the paper is easy-to-follow.\n- [S2] **Interesting Topic.** Pretraining GNNs that are generalizable to various molecular property prediction tasks is a well-studied problem, and MolCoMA could be a good addition to try in the molecular pretraining literature for practitioners." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes MolCoMA, a self-supervised molecular pretraining framework that integrates information from both 2D topology and 3D geometry in a fine-grained manner. The architecture is consisted of two modality-specific encoders, leading to a unified cross-modal encoder that leverages a complimentary masking strategy to mitigate representational overlap between the two modalities. 2D and 3D modalitity-specific decoders then perform feature reconstruction and conformer denoising, respectively, in addition to a cross-modal node-wise feature reconstruction task that aims to recover 2D features from the 3D features. Experiments on MoleculeNet and QM9 datasets show that MolCoMA outperforms previous work on diverse molecular property prediction tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Unclear motivation and intuition.** Despite complementary masking being the main contribution of this paper, the intuition behind the approach is hard to understand.\n - [W1] In general, a more difficult pretraining task is expected to bake in more useful knowledge into neural networks as seen with using hard negatives in contrastive learning literature [A]. Based on this intuition, it seems counterintuitive that complementary masking outperforms random/uniform masking as the ground-truth targets can easily be found based on nodes from the other modality during pretraining. In other words, how can we expect the 2D-specific encoder to perform well on downstream tasks with no 3D conformers, if it was pretrained towards relying on 3D-information, which is not present during finetuning?\n - [W2] Throughout the paper, the authors mention 2D and 3D information \"each possess unique representational strengths\" (Line 210), yet claim that uniform/random masking used in previous work \"result in feature redundancy due to the overlap\" (Line 207), which makes the main motivation for complementary masking self-contradictory. If 2D and 3D information indeed contain unique information, shouldn't it be the case that random masking should be enough since they express different knowledge? But if there is large representational overlap, that means complementary masking makes the task too easy for the model, which brings us to the point of W1 above. Either way, the intuition behind the proposed method needs further discussion and clarification.\n\n- **Insufficient justification on method design choices.** The proposed architecture and objective are not justified properly and requires further discussion.\n - [W3] The unified encoder composed of vanilla Transformer layers fails to preserve symmetry within the data distribution, which is crucial in ensuring generalizability and robustness. Specifically, SE(3) transformations $g(\\cdot)$ on the 3D geometry $\\mathbf{R}$ are not respected for 3D noise prediction (i.e., $f_{\\mathbf{\\theta}}( \\mathbf{X}, \\mathbf{E}, g(\\mathbf{R}) ) \\neq g(f_{\\mathbf{\\theta}}( \\mathbf{X}, \\mathbf{E}, \\mathbf{R})$) as vanilla attention is not equivariant to SE(3) roto-translations. This equivariance property could be enforced (1) approximately via data augmentation [B] or (2) exactly by replacing the attention mechanism [C], but MolCoMA discusses neither of these, making the architecture design less reliable. \n - [W4] The cross-modal reconstruction objective seems ill-defined, in the sense that the model is trained to map 3D features that are responsible for predicting the ground-truth noise (thereby depends on the noise added to the 3D conformer), to 2D features that are stable regardless of the noise. In effect, this could result in a suboptimal trade-offs for the 3D denoising task, yet this discussion is only done vaguely in Lines 324-332.\n - [W5] Lastly, the final objective (Equation 9) involves a weighted sum of three distinct loss functions, without any guidance on how the weights should be set. It would be interesting to test how the performance varies with different weights (other than $\\alpha_1 = \\alpha_2 = \\alpha_3 = 1$ case), possibly leading to insights on how each modality contributes to molecular property prediction.\n\n[A] Robinson et al., Contrastive Learning with Hard Negative Samples. ICLR 2021.\\\n[B] Quiroga et al., Revisiting Data Augmentation for Rotational Invariance in Convolutional Neural Networks. AISC 2019.\\\n[C] Fuchs et al., SE(3)-Transformers: 3D Roto-Translation Equivariant Attention Networks. 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": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- What is the difference between $H(l)$ and $H^{(l)}$ in Eq (4)?\n- How to choose the hyperparameters provided in Table 5? Since molecular representation learning frameworks are often sensitive to the hyperparameters, the hyperparameter search strategy is also important." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- MolCoMA achieves strong performance in two standard molecule benchamrks, MoleculeNet and QM9.\n- The idea of unified encoding with complementary masking is simple yet effective. Although I think this is not new in the self-supervised learning community, e.g., [1-2], but the idea has been under-explored in the molecule domain yet, so it is somewhat novel.\n\n[1] MixMAE: Mixed and Masked Autoencoder for Efficient Pretraining of Hierarchical Vision Transformers \\\n[2] Mixed Autoencoder for Self-supervised Visual Representation Learning" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces MolCoMA, a molecular representation learning framework that leverages two modalities of molecules: 2D topology and 3D geometry. Specifically, MolCoMA pretrains 2D and 3D molecule encoders (with another unified encoder) through masked auto-encoding with complementary masking. This masking strategy encourages the model to learn modality-specific characteristics more effectively, filling in the gaps between the complementarily-masked modalities. As a result, the pre-trained 2D and 3D encoders via MolCoMA achieve state-of-the-art performance on 2D downstream tasks (MoleculeNet) and 3D downstream tasks (QM9)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The method is not comprehensively described.\n - What are the data transformations, $T\\_\\text{2D}$ and $T\\_\\text{3D}$?\n - Is the unified encoder a vanilla bidirectional transformer? Does it simply take the mixed representations as an input sequence?\n - Is the output sequence of the unified encoder directly passed into the 2D and 3D decoders?\n - What is the Simple GNN Tokenizer? Is it randomly initialized or a pretrained model? Due to the stop-gradient operation, the tokenizer is not optimized, so it requires a detailed description.\n - When is the complementary masking applied? In Figure 1, it appears that masking is applied before modality-specific encoding. However, in Section 3, it seems that masking is applied right before unified encoding.\n - In cross-modal reconstruction, $h\\_\\text{2D}$ and $h\\_\\text{3D}$ appear to be modality-specific representations that do not rely on the unified encoder. Is this correct? (I assume $h\\_\\text{2D}$ and $h\\_\\text{3D}$ are only defined in L252-L263).\n - What is meant by avoiding the use of visible geometric tokens in L333? It is difficult to understand how cross-modal reconstruction works and why visible tokens should be avoided.\n - Overall, there are several confusing notations, and many parts are not clearly explained. I strongly recommend the authors use concrete mathematical notations to reduce confusion.\n2. Some concerns on experimental results.\n - Some baselines use different pretraining datasets compared to this paper. For example, MoleBERT uses 2M molecules from the ZINC15 database, and 3D-EMGP and GraphMVP are pretrained on the GEOM dataset containing 50K–100K molecules. However, MoleCoMA uses PCQM4Mv2 with 3.4M molecules. This difference in pretraining datasets could significantly impact performance, so I wonder if MoleCoMA is still effective with a different (smaller) dataset.\n - The paper lacks analysis of the effectiveness of the unified encoder. Since the authors claim the unified encoder as a contribution of this paper, providing such analysis is crucial.\n - One way to interpret the quality of molecular representations is molecule retrieval based on the learned representations. If the retrieval finds chemically similar molecules well, one could infer that the learned representations are chemically informative and useful for downstream 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": 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. Can the attention map of the Unified Encoder be visualized to examine the interactions between modalities?\n2. Can the outputs of two unique encoders, which are not in the same representation space, be directly merged and encoded together without alignment?\n3. What impact does the structure of the GNN Tokenizer used as a 2D Decoder have on the results? For example, GCN, GAT, GIN, Graph Transformer.\n4. What impact do the three sub-objectives in Equation 9 have on performance individually? It would be helpful to provide an ablation analysis of the three sub-objectives." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The motivation is understandable and reasonable, as previous work on multi-modal molecular representation learning has overlooked the cross-modal complementarity at the atom level.\n2. The proposed method is technically sound, as similar approaches have already been proposed and validated in image-text multi-modal representation pre-training.\n3. The experimental results on MoleculeNet and QM9 look promising, and the proposed method has been compared with the latest baseline models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a multi-modal molecular representation pre-training framework, MolCoMA, to facilitate fine-grained interactions of intrinsic features across modalities. The framework employs two modality-specific encoders to capture unique characteristics of each modality and a unified encoder with a complementary masking mechanism to integrate these features. By optimizing in-modal and cross-modal reconstruction losses, this framework learns robust 2D and 3D molecular representations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The datasets used for evaluation in the experiments are relatively limited. For 2D, the evaluation is only conducted on classification datasets from MoleculeNet, without testing on regression datasets. For 3D, the evaluation is only conducted on QM9, without testing on datasets like MD17 or GEOM-Drugs.\n2. The Unified Encoder is a vanilla self-attention module that does not take structural inductive bias into account.\n3. Some details of the method are not clearly described, such as the implementation of the cross-modal reconstruction function in Equation 8." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Overall, the contribution of this paper seems incremental and the experimental evaluations are not fully convincing in several aspects. Please refer to weaknesses for 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 proposed method interacts the 2-D information and 3-information, which is reasonable.\n2. The proposed method achieves promising results in different tasks.\n3. The paper is well-written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel molecular representation learning framework, termed MolCoMA, which demonstrates promising performance on the MoleculeNet and QM9 datasets. MolCoMA leverages complementary information from both topological and geometric representations while facilitating cross-modal interaction. Notably, it employs a well-designed masking technique that obscures atoms across different modalities in a complementary manner. By integrating various types of molecular information, the proposed method effectively addresses a range of downstream tasks, including biological and quantum applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of novelty. The interaction between 2d-topology and 3d-geometry has already the explored in previous studies, such as GraphMVP [a], Transformer-M [b]. \n\n2. MolCoMA is pre-trained on PCQM4Mv2, while Mole-Bert is pre-trained on ZINC and GraphMVP is pre-trained on GEOM. Therefore, the comparison is not fair. To verify the effectiveness of the proposed method, the pre-training datasets of different methods should be the same.\n\n3. More benchmarks on 3D tasks should be conducted, such as MD17 and LBA. Moreover, the proposed method should be compared with recent works SLIDE [c] and Frad [d].\n\n4. Why is complementary masking better than other strategies? The masked information of one modality can be revealed by another modality. The authors should conduct more experiments or theoretical analyses between different masking strategies and verify why complementary masking is effective.\n\n5. To verify the effectiveness of the MolCoMA, it would be better to conduct experiments on more architectures such as EGNN [e]\n\n[a] PRE-TRAINING MOLECULAR GRAPH REPRESENTATION WITH 3D GEOMETRY, ICLR 2022.\n\n[b] ONE TRANSFORMER CAN UNDERSTAND BOTH 2D & 3D MOLECULAR DATA, ICLR 2023.\n\n[c] Sliced Denoising: A Physics-Informed Molecular Pre-Training Method, ICLR 2024.\n\n[d] Fractional Denoising for 3D Molecular Pre-training, ICML 2024.\n\n[e] E(n) Equivariant Graph Neural Networks, ICML 2021." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Complementary masking strategy enhances atom-level multi-modal molecular representation" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024molcoma,\ntitle={MolCo{MA}: Complementary Masking Strategy for Promoting Atom-Level Multi-Modal Molecular Representation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y7Ud3RAPT8},\nnote={under review}\n}" }, "abstract": { "value": "Molecular representation learning, which captures the fundamental characteristics of chemical compounds, is crucial for AI-driven drug discovery. Methodologies exist that integrate various modalities (e.g., 2D topology and 3D geometry) and develop robust representations. However, current multi-modal fusion strategies either align embedding space through independent models separately, thereby overlooking complementary information, or bridge modalities at a coarse-grained level, failing to capture inherent correlation. To facilitate fine-grained interactions of intrinsic features across modalities, this study presents MolCoMA, an innovative pretraining framework for Molecular representation, employing a unified encoder that leverages Complementary Masking mechanism. Specifically, we first employ two distinct encoders to capture the unique characteristics and structures inherent in different modalities. We then utilize a unified encoder accompanied by a customized complementary masking strategy to seamlessly integrate information, mitigating overlap and similarity between 2D and 3D representations. Finally, we incorporate a cross-modal reconstruction module to enhance fine-grained interactions at the atomic level. Extensive experiments demonstrate that our model outperforms existing molecular pretraining methods across both 2D and 3D benchmarks. This finding underscores the effectiveness of our approach to fusing information between modalities." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multi-modal Fusion", "Molecular Pretraining", "Molecular Representation 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/39dc003a2444c2b363bfd7bb7ab32962a65d326a.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": "MolCoMA: Complementary Masking Strategy for Promoting Atom-Level Multi-Modal Molecular Representation" }, "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": 4 }, "primary_area": null, "questions": { "value": "See above" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper proposes a framework to learn interactive behaviors of agents in 3D worlds, which I believe is an interesting and promising direction. With the rise of embodied AI, this paper offers a novel approach to leveraging real-world videos for understanding interactions between agents and their environment.\n\n2. This paper is clearly written and well-presented, with an impressive demo that effectively showcases the benefits of the proposed framework.\n\n3. The proposed methods are logical for building the entire system and have been thoroughly validated through experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework called Agent-to-Sim to learn the interactive behaviors of 3D agents in a 3D environment from casually captured videos. Specifically, a coarse-to-fine registration method is developed for persistent and complete 4D representations. A generative model of agent behaviors of agent behaviors is trained to enable the generation of agent's reactions to the observer's motions. Interesting demos show the effectiveness of the proposed system." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In terms of behavior generation, there is extensive existing work on generating realistic human motions based on intentions or trajectories, with most approaches utilizing parametric models like SMPL. Given that the agent in the demo appears somewhat unrealistic, I was curious why parametric models weren't used for agent representation. For example, there are parametric models available for animals, such as [1,2].\n\n[1] Zuffi et al.: 3D Menagerie: Modeling the 3D Shape and Pose of Animals, CVPR 2017\n\n[2] Zuffi et al.: Lions and Tigers and Bears: Capturing Non-Rigid, 3D, Articulated Shape from Images, CVPR 2018\n\n2. Regarding the egocentric perception: is the egocentric visual/video data encoded as well? It is mentioned \"a latent representation from a\nlocal feature volume around the agent\", but an agent like a cat can not see the scene behind it. Besides, the proposed egocentric encoding to transform the world to the egocentric coordinates to avoid over-fitting as well as past encoding to capture the past motion sequence seem to have been proposed and well discussed in [3], which also predicts the motions from spatial controls.\n\n[3] Jiang et al.: EgoPoser: Robust Real-Time Egocentric Pose Estimation from Sparse and Intermittent Observations Everywhere, ECCV 2024\n\n3. The proposed system is trained and tested on the collected dataset. However, based on the information provided in the paper, this dataset appears quite small and lacks diversity, as it includes only 4 agents, 3 scenes, and has a limited duration. Given the availability of numerous large datasets featuring moving agents and varied environments, incorporating these existing datasets for evaluation would strengthen the findings. Additionally, I am curious about the generalization capability of the proposed method, as the experiments were conducted by training on only 22 videos and testing on a single remaining video. Do the training and testing data contain the same scenes? If so, this might lead to potential overfitting issues.\n\n4. I am interested in the potential applications of the proposed systems beyond the general areas mentioned in the paper, such as driving, gaming, and movies. It would be valuable to explore specific existing tasks to demonstrate how the system can provide practical benefits." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 address the questions raised in paper weaknesses in the rebuttal. Please address the lacking details of the method description and the raised shortcomings of the evaluation." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The proposed approach to combine 4D scene reconstruction and training diffusion models for behavior prediction seems novel and is interesting.\n- The paper is mostly well written and is easy to follow.\n- The paper presents two contributions: 1. 4D scene reconstruction including body pose estimation of an agent using NeRF-based representations. 2. A hierarchical diffusion model for behavior prediction in the 4D scene representation.\n- The related work section provides a compact overview of the state-of-the-art for related fields.\n- The experimental evaluation demonstrates improvements over previous methods for 4D scene reconstruction and agent motion prediction. It also demonstrates improvements over several ablations of the method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Agent-To-Sim (ATS), a framework for learning models of agent behavior conditioned on the observer (ego-perspective) and the static background scene. The scene and the agent are represented using neural radiance field in a canonical state. Additionally, the dynamic motion of the observer and the agent as well as the agent's body deformation is represented using a sequence of SE(3) poses. The latter is achieved by representing the motion as a set of 3D Gaussians in the agent frame and mapping locations by blend-skinning. The NeRF representations are combined and rendered in the respective poses using ray integration. The dynamic scene reconstruction is obtained from video sequences using bundle-adjustment like optimization, initialized with a trained camera pose regressor. Behavior prediction for the agent is trained using a 3-stage diffusion model which subsequently predicts the goal sequence, the path of the agent, and finally its body pose sequence. The approach is evaluated for camera registration, 4D reconstruction, and behavior prediction accuracy quantitatively on one test video sequence for a cat. It is also compared with previous methods as baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper is too packed by presenting two contributions at once and by this important details (see below comments) are missing which hinder reproducibility of the approach.\n- l. 199ff, it is unclear how the rendering works. How are the NeRF representations of scene and agent combined in detail? Please provide further information either in main paper or supplemental material.\n- The paper references directly to several Figures in the appendix which is attached to the main paper without noting that this is supplemental material. Please indicate clearly in the main paper text that they are supplemental material when referencing the figure. Currently, it seems the paper would try to circumvent the page limit. \n- What is the orientation representation for regression in Eq. 5 ?\n- It is not clear why there is a need for a joint optimization of poses and representations T and D as indicated in Eq 6, because the video sequence is annotated with camera poses from a different (unspecified) system (l. 225). Please clarify which other method is used to obtain the camera motion trajectory. Why is it necessary to optimize the poses in Eq. 6 ? How would the approach perform without this post-optimization? Please add an ablation study. What is the point of training a neural localizer if the camera poses are known from a different system? \n- How are the agent root and body part poses intiialized for Eq. 6? \n- How are agent and static scene segmented? \n- l. 251ff, Is the training and swapping of beta performed across various scenes (i.e. different types of rooms) or only within specific scenes? \n- l. 263, why was T*=5.6s chosen and how does the performance of the method depend on the horizon length? This should be evaluated as supplemental results.\n- Fig. 2, it is not clear what the thin black arrows mean (e.g. between the motion fields and between score map and goal/path). What is the motion field depicting? Where is this output in the proposed method?\n- l. 304ff., the note on generalization to different actions seems problematic. The datasets used in the experiments seem small. The biggest dataset (cat) has only 22 training videos of a few minutes each. The anecdotal result pointed to in the supplemental material (Fig. 11) might be a rare example. Please discuss and revise the claims.\n- Eq. 11, what is \\Psi_s ? Please define.\n- l. 314ff, the text mentions trajectories, but the equation only transforms a single pose (\\xi^a). Please clarify.\n- l. 318ff, please explain why the approach uses the observer trajectory perceived by the agent and not the other way round. Please also provide an evaluation if this makes a difference. \n- l. 360, a segmentation method is mentioned,but it is not stated how it is used.\n- l. 401. GT relative camera pose is obtained by some unspecified process. How does it work in detail and how accurate is this process? Please provide this description in the supplemental material.\n- l. 428ff / Tables 4 and 5: the motion prediction is conditioned on ground-truth goal and path for the evaluation. However, this should also be evaluated for regressed goal and path to fully evaluate the proposed method without such oracle information.\n- How were ground-truth goal and path annotated in the data ?\n- It seems the paper lists several datasets (bunny, human, etc), but only shows results for the cat videos.\n- If possible, the datasets used in the evaluation should be made publicly available for future comparison with the proposed method. Is it planned to release the datasets publicly ? Please also specify how this data was collected. Are there potential personal data protection issues for releasing the data?\n- The datasets used for evaluation are rather small and, hence, results seem anecdotal and significance is unclear. The method should be evaluated quantitatively on more and longer sequences and a larger variety of agents (e.g., 10-20 different animals and human subjects) and scenes (e.g. 10-20 different rooms) to increase significance of results and to provide further insights into the performance and limitations of the proposed method. Also the variety of daily activities should be increased and annotated. Cultural and gender diversity of subjects should also be considered when designing a benchmark dataset. With larger datasets, also the dependence of the performance of the method on the amount of available training data could be evaluated. \n\nFurther minor comments:\n- p. 2, please add a label/caption to the lower figure and reference it from the text.\n- l. 76ff, the notation for \\sigma, c, \\psi and the variants c_s/a, \\sigma_s/a, and \\psi_s/a is not accurate. Are the subscripts referencing parts of the variables without subscript? What kind of mathematical structure are they? \n- l. 210 \"buil\" => \"build\"\n- Fig. 2 caption \"fully\" => \"full\"\n- Eq. 11, \"X_w\" should be \"X^w\"\n- l. 369, the reference to Eq 5 should point to Eq 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": 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. I assume Equation 1 and 2 are through NERF context. it is not clear to train them, how to get the poses to train them?\n2. How to train Equation 2 when agent is in motion. Elaboration is needed.\n3. A picture is needed to show the representation of agent. It is not clear what is meant by root and how that is identified from video automatically for any morphology. \n4. Satio et. al. uses SMPL morphology where getting the skinning is understood. Here the agent can be anything. So the skinning part needs more elaboration to understand by the reader. What is meant by \"bones have time varying centers\" and how this impact the skinning? Need picture with explanation.\n5. What do we mean by set of 3D Gaussian as bones? How they are fitted in video? How they are initialised?\n6. Scene specific neural localiser regresses the camera poses, how to get the GT poses? In the following paragraph it is written that \"to learn neural localiser we first capture a walkthrough video and build a 3D map..\" how to build this 3D map? where the poses are coming from? Is this video is the same we are dealing for 4D reconstruction? Lots of confusion here.\n7.If the above 3D is in place , we can sample R*, T* and solve Eqn 5. But what is theta there and what is the importance of that parameter? Agent can move while R* can be fixed. This is not well understood.\n8. Dynamic 3DGS also require poses. How to get them? Equation 5 needs more illustrative explanation.\n9. Because the writing of registration section is bit convoluted questions are arising for Behavious representation and Goal generation. Why score based method was needed for generating goal? VAE could be good enough?\n10. Scene observer and past encoding is a complex design. When the scene is already encoded as a MLP(Eq 1,2) it is not clear why another encoding? Or am I missing something?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper shows an original pipeline to reconstruct 4D representation from multiple videos over longer timespan like a month. This is a complicated problem and the method solves this with good qualitative and quantitative results. \nThe choice of representation in every stages is carefully thought through and use of the body of knowledge is good. For example, the study by Banani et.al, which shows that large image models have good 3D and viewpoint awareness, is exploited nicely in this paper.\nThe decoupling of static and dynamic structure and registration is well designed and merged to solve the complicated relationships exist in a long scene with different variation in layout or color over time." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method for learning interactable behaviour of agent from casual video. To this end, the reconstruct the video (agent, scene and observer) in 4D. The representation has canonical structure in 3D for time independent element using NERF and a time varying structure (observer and agent). The paper proposes a coarse to fine registration of agent and environment into a canonical 4D space. This 4D representation further helps in behaviour representation further used in understanding the behaviour of agents and goal conditioned trajectory generation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Despite the strength, the paper writing is not very good for a reader. Many places the approach is not well understood. In many places argument is placed without much conviction, for example, neural localiser finds more robust than geometric correspondence while being computationally efficient is not well grounded through technical principles or empirical evidences. The ego perception of the world and scene, observer, and past encoding needs more elaboration in writing to make them understand. It is not clear how the path is generated given a goal? is this a shortest path or any other path. More doubts are raised as questions below.\n\nBecause the writing, required illustrations, and the flow of thought is bit convoluted, I am tending towards the low score of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I have some questions on the paper writing:\n\n1. In eq.8, both \\sigma and \\epsilon indicate the noise? since it is mentioned in line 303 that \\sigma is a noise value. If it is the case, please clarify: Does \\epsilon indicate the variance of the noise? \n2. In line 310, how large is the volume queried from the 3D feature volume?\n3. What is the definition of the generated path? Is it the root translations in the scene？" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I appreciate the careful 4D agent-to-sim pipeline design, including canonical and time varying structure design, the camera localization and the optimization procedure. It enables the improvement of the interactive behavior simulation results as verified in the rendering quality and Interactive behavior prediction. \n\n1. the simulated behaviors or animations of cats and humans looks realistic and interacts with the scene geometry in a physics-plausible way, no obvious penetration between the agent and the scene geometry in the video demos. \n2. The hierarchical behavior simulation model, from goal, path to body poses, and egocentric perception features in this paper are verified to be effective in improving the accuracy of the behavior prediction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a 4D reconstruction method and environment-aware behavior generation method for 3D agents. It represents the canonical structures for the scene and agents by MLPs as in NeRF, and represents the time-varying structure with camera pose, root pose and bone transformations at each frame. These structures are all learned from casually captured videos. After recovering the structures, three diffusion models are trained to generate goal, path and body poses conditioned the 3D feature field of the scene. \n\nThe main contribution of this paper is a complete pipeline to reconstruct and simulate interactive behaviors of 3D agents in a dynamic scene." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. It is not clear how to collect the training data for neural camera localization. Does the camera move along with the agent or fixed during capturing？ If it is fixed, there are only a few camera poses that can be used in the training. If it can move, that means we already have an algorithm to obtain accurate camera poses for training. \n2. The symbol \\sigma is used to represent density in Sec. 3.1 and noise in Sec. 3.3, a kind of misleading when reading the paper." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Given monocular videos collected across a long time horizon (e.g., 1 month), we build interactive behavior models of an agent grounded in a 3D environment." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024agenttosim,\ntitle={Agent-to-Sim: Learning Interactive Behavior Model from Casual Longitudinal Videos},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y80D4IojuY},\nnote={under review}\n}" }, "abstract": { "value": "We present Agent-to-Sim (ATS), a framework for learning interactive behavior models of 3D agents in a 3D environment from casually-captured videos. Different from prior works that rely on marker-based tracking and multiview cameras, ATS learns natural behaviors of animal and human agents in a non-invasive way, directly from monocular video collections. Modeling 3D behavior of an agent requires persistent 3D tracking (e.g., knowing which point corresponds to which) over a long time period. To obtain such data, we develop a coarse-to-fine registration method that tracks the agent and the camera over time through a canonical 3D space, resulting in a complete and persistent spacetime 4D representation. We then train a generative model of agent behaviors using paired data of perception and motion of an agent queried from the 4D reconstruction. ATS enables real-to-sim transfer of agents in their familiar environments given longitudinal video recordings (e.g., over a month). We demonstrate results on pets (e.g., cat, dog, bunny) and human given monocular RGBD video collections captured by a smartphone." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "dynamic 3d reconstruction; multi-video registration; motion generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/8b9663f15b6e43ab90c953d7cd0be864f04ebd82.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/3f51894728dacc7925fdc0f0df2264a11856c83f.zip" }, "title": { "value": "Agent-to-Sim: Learning Interactive Behavior Model from Casual Longitudinal Videos" }, "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": "1. Could the paper provide more explanations about how the \"true\" labels for evaluation is obtained? If people cannot provide accurate high-quality labels, how do we evaluate machine performance on these hard tasks? Is there a chicken-and-egg problem here?\n\n2. Could the authors comment on how the reported results compare with human performance?\n\n3. What's the definition for these \"hard\" tasks? One precise piece of information mentioned in the paper is that people are \"less than 90% accurate\", but an accuracy of 90% does not appear too hard to me.\n\n4. Addressing my other main concerns in the \"Weaknesses\" section would be helpful." }, "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 poses an important and interesting question about what the best methodology (i.e., what type of data to collect and what training method to use) is in this challenging regime of \"hard\" tasks where human data is unreliable or expensive. Improving LLMs' ability to generalize to these tasks has lots of potential in making positive societal impacts.\n\n2. The paper presents a wide range of experimental settings and results, which provide useful information for future practitioners working on improving LLMs on these hard tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers the task (termed \"scalable oversight\") of training LLMs on problems that human judgment is unreliable or expensive. The paper conducts experiments that vary {the type of large models, the datasets and associated tasks, the training methods}. Using results from these experiments, the paper observes different regimes where it can be more effective to only collect high-quality data, low-quality data, or a mix of both." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I appreciate the efforts and empirical results by the paper, I feel that this paper looks more like an engineering technical results for the following reasons:\n\n1. Scientific rigor:\n\n(a) Please include error bars on all reported plots. If such errors are not available, please provide discussions on the statistical significance of the reported observations.\n\n(b) The experimental observations can be made more precise. For example, while the paper claims that \"$256-$512\" is the mixed regime in Figure 1, this appears very task dependent in Figure 2.\n\n(c) The discussion on limitations can be more in depth. For example, a distinction can be made on low/high-quality data provided by human vs. smaller LMs (as in the experiments). Does the paper specifically study training data provided by LMs? Does this generalize to low/high-quality data provided by people or not? Also in the experiments, the smaller LMs are trained on other data (\"we generate weak labels using small LMs that have been finetuned on the task\"), which confounds the conclusions made in the paper about low/high-quality data in the experiments.\n\n2. Unsubstantiated contributions:\n\n(a) The second claimed contribution includes \"microeconomic assumptions\", but as far as I can tell, this is not investigated in the paper. Appendix B appears unfinished. \n\n(b) I also think the third claimed contribution is very limited. While research framing can be a valid contribution, in this case, the problem of data usage and training methods is a well-known research topic and extensively studied in the past. The Pareto frontier is essentially a piece of empirical result that provides some practical evaluation and guidance on what training methods work better in what settings.\n\n2. Lack of context:\n\nI find the background knowledge introduced in this paper insufficient for audience beyond LLM practitioners. For example, what are the tasks \"BoolQ, HellaSwag, SciQ, Cola, ...\"? What's \"LoRA\" for supervised finetuning? What does it mean to \"early-stop based on validation AUROC\"? Why are \"Qwen 1.5-0.5B and Qwen1.5-4B\" chosen for the initial set of experiments? What's a new \"head\" for training? What's the procedure for \"few-shot prompting\" and what are \"in-context examples\"?\n\n3. Insufficient literature review:\n\nThe literature review only provides citations on very recent advances specifically tied to LLM development. I personally find it insufficient without giving credits or contextualizing the research question with respect to other lines and domains of work. Two of such domains are:\n\n(a) The field of crowdsourcing primarily concerns how to make use of noisy human data, which is closely relevant to the paper.\n\n(b) The field of economics computation literature concerns data valuation and pricing, which is also closely relevant to the paper.\n\n4. Clarity (minor):\n\n(a) In the result section, the paper mentions \"quality of weak labels\". What does this mean? I thought weak labels mean low quality in this paper.\n\n(b) I suggest the paper using different terms than \"quantity-dominant\" and \"mixed\". These terms are confusing because the quantity-dominant regime cannot afford enough weak labels, and the mixed regime can afford enough weak labels.\n\n(c) The wording \"scaling results\" in Section 4.1 is vague. What results does this refer to?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Why choose NLP classification task as the one to test this tradeoff? Will the result holds for other tasks? If this is unclear, what should the main takeaway for the reader of the work to be?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The question on the trade-off between quantity and quality is interesting; and the finding that no pure strategy (few high quality data and much low-quality data) is clear. The results are supported by extensive experimentation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explored different ellicitation techniques on improving an NLP binary classification tasks under some fixed budgets. Under this budget, one can either have a lot of low quality data, or a few high quality data, or a mixed between the two. Using these tasks, the author explored the pareoto frontier between mixing the low and high quality data, finding that no pure strategy dominates." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My biggest concernis that the insights of the paper is not very clear. As it is true that there is a trade-off between quantity and quality, it is unclear how the specific finding in this paper can be generalizable to tasks beyond NLP binary classifications showcased in the paper. In this case, it might make sense to come up with a scaling law that captures this tradeoff." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- How and why were the tasks binarized? Does binarization meaningfully affect the results? \n- It seems like the gains from using only weak labels are generally quite limited. How does this relate to the results from Burns et al (2023)? \n- How do you determine that \"256 and 1024 high-quality finetuning examples do not reliably elicit knowledge from Llama-3-8B, but elicit most of Llama-3-70B’s knowledge.\"?\n- \"The quality of in-context examples is inconsequential, while the quality of SFT examples matters substantially.\"\n - Is this refering to a specific regime? The top/bottom left of figure 4 seems to show a big difference, at least for 2/8 in-context and few SFT samples. \n- Is there a typo in the caption of Figure 3? The accuracy of the weak labels displayed in the figure appears to be larger than 70.2%. Or is this the difference between training/test accuracy?\n- \"We use 3 random seeds, except for training runs where the smaller stage takes less than or equal to 10 examples, in which case we use 7 random seeds\"\n - How do you get less than 10 examples when the minimum budget is at 16?\n- Regarding the GPT-4 results, how well does GPT-4 perform on the tasks in a zero-shot manner?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The research question on quality-quantity tradeoffs is interesting. \n- While quite specific to the weak-to-strong setting, the experimental setup seems to be well-suited for the research question. \n- The experimental results cover multiple tasks and models and are somewhat comprehensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "- This paper investigates the quality/quantity tradeoff when finetuning an LLM on cheap low-quality and expensive high-quality labels. \n - The low quality labels are produced by a small LLM that has been finetuned to the task at hand (on a separate datasets), while the high-quality labels appear to be ground-truth labels. \n - Assuming that the cost of low-quality labels is ten times lower than the cost of high-quality labels, the authors find three regimes, depending on the available budget: For small budgets, performance increases in the fraction of low-quality labels used for training. For large budgets, performance instead decreases when more low-quality labels are used. Lastly, for intermediate budgets, performance first increases and then decreases in the percentage of low-quality labels.\n- Experiments are done on multiple (binary) NLP classification tasks, and include different training strategies, sometimes combined with few-shot prompting. \n- Ablations are conducted using different models, both for label generation and finetuning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper appears to have been submitted in an unfinished state\n - Different microeconomic assumptions are mentioned as part of the contributions and discussion, but I did not find any discussion of these in the main text (and very little in the appendix).\n - There is a range of minor issues that are unproblematic on their own, but add up: \n - The paper uses the wrong template (ICLR 2024 rather than 2025)\n - The paper containe multiple instances of questions in odd places that appear to be todos by the authors. For example, \"All datasets?\" at the end of the references, and \"What about scenarios where the cost you invested into training up your labelers or understanding a problem has externalities for other training runs etc?\" in Appendix B. \n - A bunch of citations are missing details like the arxiv identifier\n - Figures 1 and 2 are presentented in double column without any apparent reason, making things more difficult to read. \n - The logical flow of the writing could be improved at times. For example: \"the role of pretraining in weak-to-strong generalization\" is mentiond before weak-to-strong generalization is introduced, making the remark difficult to understand. \n - There is also room for improvement in terms of \"local\" writing clarity. For example:\n - \"Results broken down by each of the three datasets can be found in Appendix figures 6, 7, and 8, suggesting that the results hold across tasks and weak label qualities.\"\n - What result is supposed to \"hold\" here? I can see that the pareto-fronts have somewhat similar shapes, but what does that imply?\n - The description of the quantity-dominant setting is difficult to parse (perhaps in part due to the dual-column structure).\n- Apart from the paragraph on scalable oversight, the related work section is relatively sparse and could benefit from incorporating existing work on sample values [1] and quality/quantity tradeoffs in machine learning [2,3,4].\n- While this might be hindsight bias speaking, I do not find the main result, that a mixture of higher quality and lower quality labels is optimal, particularly surprising. For example, while [3] is not fully comparable, in part because high quality labels are selected actively, that work also finds a mixture of higher and lower quality labels to be optimal for training. \n\n\n[1] Torralba, Antonio, and Alexei A. Efros. \"Unbiased look at dataset bias.\" CVPR 2011. IEEE, 2011.\n\n[2] Sheng, Victor S., Foster Provost, and Panagiotis G. Ipeirotis. \"Get another label? improving data quality and data mining using multiple, noisy labelers.\" Proceedings of the 14th ACM SIGKDD international conference on Knowledge discovery and data mining. 2008.\n\n[3] Chen, D., Yu, Z., and Bowman, S. R. Clean or annotate: How to spend a limited data collection budget. arXiv preprint arXiv:2110.08355, 2021.\n\n[4] Crammer, Koby, Michael Kearns, and Jennifer Wortman. \"Learning from data of variable quality.\" Advances in Neural Information Processing Systems 18 (2005)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We compare learning methods and optimally trade off the quantity and quality of labels to elicit knowledge from a pretrained model in a labeling-cost constrained setting." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024balancing,\ntitle={Balancing Label Quantity and Quality for Scalable Elicitation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y8TjnkdWNA},\nnote={under review}\n}" }, "abstract": { "value": "Scalable oversight studies methods of training and evaluating AI systems in domains where human judgement is unreliable or expensive, such as scientific research and software engineering in complex codebases. Recent work in this area by Burns et al. (2023) suggests that Language Models (LMs) pretrained on internet-scale corpora exhibit an inductive bias toward producing correct answers, even when finetuned on error-prone labels produced by a smaller language model. This suggests that massive pretraining combined with finetuning on imperfect human labels may be a solid baseline method for scalable oversight. In the real world, however, label quality is not fixed: practitioners face a quantity-quality tradeoff when generating finetuning data. In this paper, we explore the microeconomics of the quantity-quality tradeoff on binary NLP classification tasks used in Burns et al. (2023). We find that there are three regimes of eliciting classification knowledge from pretrained models using supervised finetuning: quantity-dominant, quality-dominant, and a mixed regime involving the use of low- and high-quality data together to attain higher accuracy at a lower cost than using either alone. We explore sample-efficient elicitation methods that make use of two datasets of differing qualities, and establish a Pareto frontier of scalable elicitation methods that optimally trade off labeling cost and classifier performance." }, "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": [ "Scalable oversight", "Alignment", "Safety", "Few-shot learning", "Eliciting latent knowledge", "Weak-to-strong 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/d7d785a08dc058fb74bcf8d3237f484481fefbf6.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": "Balancing Label Quantity and Quality for Scalable Elicitation" }, "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": 1 }, "primary_area": null, "questions": { "value": "Both the statements and technical aspects should be refined for rigor. A thorough revision is recommended before resubmission." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The topic is interesting and valuable. The authors commendably attempt to combine physics-driven weather data with observation-based remote sensing data to explore the correlation between these two modalities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents an approach that aims to integrate physics-driven weather data with observation-based remote sensing data to investigate the correlation between these two modalities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some opinions and assertions in the paper lacks experimental or theoretical support. For example:\n * Line093: The claim that \"temporal flexible architecture is useful for generalizing across downstream tasks\" is presented without the necessary theoretical backing or experimental validation.\n * Line096: The assertion that the embedding created by the proposed architecture is richer than those using reconstruction tasks is only substantiated by results from the authors' framework, lacking comparison with other methods and relevant theoretical derivation.\n2. Unclear Methodology and training detail description.\n * The description of the methodology lacks formalization, rendering the training process unclear.\n * There is insufficient detail regarding the network architecture and training procedures such as the hyper-parameters, training and fine-tuning epoch number, batchsize, and etc.\n * The language employed in the paper lacks academic rigor, featuring informal expressions (e.g., using \"let us look at …”), grammatical errors, improper punctuation, formatting issues (e.g., using $\\times$ instead of x), incorrect capitalization, and various typos. A revision of the writing is strongly recommended.\n3. Deficiencies in Experimental Design. \n * The baseline comparisons in the experimental part can only be classified as ablation studies, as they do not include comparisons with other foundational models.\n * In the downstream task validation, comparisons are limited to the authors' ablation models, without benchmarking against state-of-the-art models in the relevant application field.\n * Given that many pretrained ViT models (even without fine-tuning in the RS domain) achieve commendable performance in remote sensing tasks, the authors should also compare their results with those of pretrained models.\n * There is a lack of experimental design concerning the model's generalization capabilities. As a foundational model, the experiments are supposed to involve more types of downstream tasks and cover remote sensing data from more satellites.\n4. Suboptimal Experimental Results. The visual results in this paper appear to be of limited utility, and there is also a lack of comparison with results from other MAE-based or contrastive learning-based foundation models." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Why is causal encoding applied to the fused multimodal embeddings, while Bi-LSTM is used for the weather data? Logically, weather data should also be unidirectional.\n\n2. In Figure 6, the non-restored areas in the SM-VSF results exhibit mosaic artifacts and appear inconsistent with the original image. Does this method affect regions of the image outside the restored areas?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The authors applied variable-step prediction, commonly used in time series forecasting, to the pretraining task of a remote sensing multimodal foundational model and achieved performance improvements in two downstream tasks. This is a preliminary exploration of a remote sensing multimodal foundational model in spatiotemporal prediction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a foundational framework for exploring the causal relationships between temporal textual data and remote sensing imagery. The authors first use a Bidirectional LSTM to model the temporal text, followed by a Unidirectional Transformer to extract the causal relationships between the textual and image features. Ultimately, this framework can generate corresponding remote sensing images based on weather data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The impact of weather on the environment is a highlight of this paper and should be its most compelling section. However, this topic is only briefly mentioned in the abstract and introduction. Unfortunately, no subsequent experiments are conducted to discuss the specific effects of weather on remote sensing imagery, which is quite regrettable. The explanation of the causal relationship between weather and imagery seems to be solely connected to the use of the causal Transformer model, which feels somewhat tenuous. I suggest that the authors present several cases to demonstrate how different weather models affect the prediction outcomes or conduct an ablation study to isolate the contribution of weather data to the model's performance.\n\n2. The model's performance is limited. Specifically, in Figure 8, I even feel that the MM-MAE results visually appear superior to those of SM-MAE. The SM-MAE results seem somewhat more blurred. Additionally, the images in Figures 6 and 8 are quite similar. I would encourage the authors to provide more qualitative analysis to demonstrate their method's advantages better. I suggest that the authors conduct a detailed analysis of challenging individual cases, or visualize the learned representations to illustrate how they capture weather-related information.\n\n3. The entire paper seems to rely on only two quantitative analysis experiments, presented in Tables 1 and 2, which feel somewhat insufficient. The authors have designed a complex encoder composed of multiple components, but no ablation studies have been conducted to justify the selection of each component. This makes it difficult to be convinced that the author's choices are optimal. I suggest the authors include ablation experiments regarding the model architecture. For example, the additional experiments mentioned above." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. In light of Weakness 1, the authors are encouraged to clarify the motivation and considerations underlying the proposed method.\n2. Could you provide a detailed explanation of the design rationale behind your framework?\n3. CDL is widely recognized for its association with noisy labels. Do you implement any label correction on this data, and if so, what specific steps do you undertake?\n\nAdditionally, the authors are encouraged to address the concerns outlined in the Weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The concept of establishing a geoscience foundational model from the perspective of forecasting future spectral imagery is novel, and the utilization of weather information in this context has not been fully explored by the community before. This paper presents an intriguing solution tailored specifically to a foundational model for geoscience.\n2. Overall, the writing is satisfactory; however, there are some significant drawbacks that I will address later." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an innovative foundational model framework for geoscience, which undertakes the task of forecasting future spectral imagery using weather information and spectral imagery from previous timestamps. Through the construction of a MultiModal Variable Step Forecasting (MM-VSF) paradigm, the authors illustrate the superiority of MM-VSF in two downstream tasks: crop mapping and missing image prediction. When compared to MAE pre-training and single-modality scenarios, the proposed method demonstrates enhanced performance, thereby validating the significance of incorporating weather information in certain remote sensing tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The fundamental assumption may be flawed. While it is reasonable to assert that weather can be an influential factor in certain land use or land cover predictions, forecasting spectral imagery solely based on weather and previous spectral imagery clearly overlooks essential factors, such as human activities. For instance, in crop mapping predictions, farmers may choose not to cultivate any crops in a given year, a decision that is entirely independent of weather conditions. Similarly, the segmentation of urban functional areas is also not contingent upon weather; the construction of buildings occurs regardless of weather considerations. Therefore, the current pre-training task appears to be an inadequate choice for addressing a wide range of remote sensing tasks, rendering the fundamental assumption of this paper unconvincing.\n2. The presentation of the main methodology lacks clarity. For instance, the data format for the weather information is ambiguous, and it is unclear how the bi-directional LSTM processes this weather data. I suggest providing a formal description of both the input and output data to illustrate the data flow pipeline, as this would offer a clearer overview of the methodology. Furthermore, the rationale behind selecting ViT instead of SwinViT, as well as the choice of LSTM over GRU, is not explicitly addressed. It appears that the selection of these building blocks is somewhat arbitrary and lacks specific consideration within the framework's design.\n3. The experimental results presented are primarily an ablation study. In the main experimental section, the authors compare the proposed MultiModal Variable Step Forecasting (MM-VSF) model solely with various combinations of single modalities and MAE-based pre-training methods. The current results demonstrate only the necessity of the weather modality and the superiority of the Variable Step Forecasting over MAE. However, there is insufficient evidence to support the assertion that MM-VSF is a suitable choice for universal remote sensing interpretation tasks. I strongly recommend that the authors compare MM-VSF with well-established remote sensing foundational models, such as SkySense (CVPR 2024), SatMAE++ (CVPR 2024), and DeCUR (ECCV 2024), using recognized benchmarks such as BigEarthNet, fMoW, and DIOR. Additionally, the potential of contrastive learning as an avenue for developing remote sensing foundational models is not adequately addressed in this paper.\n4. The analysis of the learned causal relationships is notably absent. In the abstract and introduction, the learned causal relationship between weather and spectral imagery is presented as a significant contribution of this paper; however, a detailed analysis appears to be lacking. I would appreciate the inclusion of quantitative results or mathematical induction, as causality is a well-established mathematical concept, to further elucidate this aspect.\n5. Several important related works appear to be missing from the discussion. Regarding weather forecasting, the authors should consider including GraphCast (Science 2023) and Fuxi (Nature Communications 2024). Additionally, in the realm of contrastive learning, notable contributions such as DeCUR (ECCV 2024) and DINO-MM (IGRASS 2022) should also be acknowledged." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 does the variable step forecasting task handle significant time gaps in satellite data? For example, how does it manage situations where satellite imagery is missing for extended periods?\n2. Why the shape of weather data is 1x1？\n3. In the crop mapping task, how sensitive is the performance of MM-VSF to the temporal resolution of the input data? Would increasing or decreasing the frequency of the input images change the results significantly?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The introduction of a multimodal pretraining task adds a novel aspect to the existing methods in the geoscience foundation model landscape. The paper is generally well-structured. The ability to generalize across years and handle missing data offers practical utility in real-world settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a pretraining framework called MultiModal Variable Step Forecasting (MM-VSF). The proposed foundation model leverages multimodal data, specifically satellite imagery and weather data, to improve spatio-temporal remote sensing tasks like crop mapping and missing image prediction. The core idea is to pretrain the model using a forecasting task that captures causal relationships between the modalities, improving generalization and representation quality for downstream tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Insufficient experiments: My main concern about this paper is the lack of comprehensive experiments. The baseline models could be expanded, for example, by including comparisons with contrastive-learning-based methods to provide a more holistic evaluation.\n\n2. Lack of fine-grained ablation studies: The paper could benefit from more fine-grained ablation studies, such as experimenting with different masking ratios and strategies. This would help clarify the specific impact of these hyperparameters on model performance.\n\n3. Limited downstream tasks: The range of downstream tasks is insufficient. Including tasks like urban semantic segmentation mapping, which is highly relevant in the field of remote sensing, would strengthen the generalizability of the approach.\n\n4. Incomplete related work: The related work section is not sufficiently comprehensive and clear. There is a lack of discussion and comparison with relevant multimodal foundation models in remote sensing, particularly models that also utilize satellite and weather data [1-2]. Incorporating these references and analyzing their similarities and differences with the proposed approach would offer better context and positioning.\n\n5. Figures: Figures 1 and 2 have unclear text, which may hinder the understanding of key concepts. Improving the clarity of the text in these figures would enhance the overall presentation quality.\n\n[1] Ravirathinam, Praveen, et al. \"Combining Satellite and Weather Data for Crop Type Mapping: An Inverse Modelling Approach.\" Proceedings of the 2024 SIAM International Conference on Data Mining (SDM). Society for Industrial and Applied Mathematics, 2024.\n[2] Nedungadi, Vishal, et al. \"MMEarth: Exploring multi-modal pretext tasks for geospatial representation learning.\" arXiv preprint arXiv:2405.02771 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024towards,\ntitle={Towards a Knowledge guided Multimodal Foundation Model for Spatio-Temporal Remote Sensing Applications},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y8qBBbAdEv},\nnote={under review}\n}" }, "abstract": { "value": "In recent years, there has been an increased interest in foundation models for geoscience due to the vast amount of Earth observing satellite imagery. Existing remote sensing foundation models make use of the various sources of spectral imagery to create large models pretrained on the task of masked reconstruction. In this paper, we present a foundation model framework, where the pretraining task captures the causal relationship between multiple modalities. Our framework leverages the knowledge guided principles that the spectral imagery captures the impact of the physical drivers on the environmental system, and that the relationship between them is governed by the characteristics of the system. Specifically, our method, called MultiModal Variable Step Forecasting (MM-VSF), uses forecasting of satellite imagery as a pretraining task and is able to capture the causal relationship between spectral imagery and weather. In our evaluation we show that the forecasting of satellite imagery using weather can be used as an effective pretraining task for foundation models. We further show the effectiveness of the embeddings produced by MM-VSF on the downstream tasks of pixel wise crop mapping and missing image prediction of spectral imagery, when compared with embeddings created by models trained in alternative pretraining settings including the traditional single modality input masked reconstruction." }, "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": [ "Foundation model", "Spatiotemporal modelling", "Remote Sensing", "Knowledge guided" ] }, "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/0af57440475b30c9e88033babe54360898e2ab1c.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": "Towards a Knowledge guided Multimodal Foundation Model for Spatio-Temporal Remote Sensing Applications" }, "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": "In the paper, the author points out that for some complex scenes, since a large number of 3D Gaussians are needed for representation, the sorting process will exhaust the resources of mobile devices, making real-time rendering impossible, such as the bicycle scene in Mip-NeRF 360 dataset. However, I tried to visualize the bicycle scene using a gsplat-based webgl viewer on my phone (iPhone 15 Pro Max), and was surprised to find that it achieves real-time rendering, about 20-30fps. This seems to indicate that some optimizations to the 3DGS rendering pipeline (gsplat uses some tricks) can increase the rendering speed on mobile devices without more complex designs. I would like to know the author's opinion on my point of view and more analyses of the key differences that enable real-time performance on mobile devices. And they can test the performance of gsplat on Snapdragon® 8gen3 chipset and compare with their proposed method.\n\nHere is the link of gsplat webgl-viewer https://gsplat.tech" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The author provides a clear analysis of the tile-based sort in the vanilla 3DGS, explains its impact on rendering speed, and explains why it is difficult to implement on other devices.\n- The proposed method speeds up rendering while ensuring rendering quality, and has been verified on consumer mobile devices.\n- The proposed weighted sum approximation alleviates the *pop* phenomenon caused by inconsistent sorting results under different view-directions in the vanilla 3DGS." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a sort-free Gaussian Splatting pipeline, which reduces the computational overhead of rendering to adapt to scenarios with limited computing resources. The key insight of this paper is that the view dependent sorting in the vanilla 3DGS involves a lot of computational overhead and is difficult to port to mobile devices. Therefore, the authors proposed to use weighted sum and view-dependent opacity to approximate alpha-blending. Experiments demonstrate that proposed method achieves comparable rendering quality and faster rendering speed than the original 3DGS on mobile devices." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The proposed method gets rid of the costly sorting, but a more complex weight and view-dependent opacity are used. Although the author points out in the paper that a general scheme for learning scene representations is bound only by the mathematical constraints of the scene model. According to the experimental results in the paper, the use of simple weighted-sum and view-independent opacity cannot achieve satisfactory rendering results. Does this mean that for rendering processes that are not physically and optically, more complex models must be used, such as introducing more learnable parameters as in the paper?\n- In the conclusion part of the paper, the author mentioned some recent research on compact Gaussian Splatting. According to my understanding of these works, since the methods proposed in these works obtain more compact 3D Gaussian representations, they can reduce the sorting overhead and speed up rendering to a certain extent. Therefore, I think the author should add a comparison of rendering quality, speed and memory usage between the proposed method and these works, such as Compact 3DGS[1], LightGaussian[2] and Compressed 3DGS[3] in the experimental evaluation section.\n\n[1] Compact 3D Gaussian Representation for Radiance Field https://openaccess.thecvf.com/content/CVPR2024/papers/Lee_Compact_3D_Gaussian_Representation_for_Radiance_Field_CVPR_2024_paper.pdf\n[2] LightGaussian: Unbounded 3D Gaussian Compression with 15x Reduction and 200+ FPS https://arxiv.org/pdf/2311.17245\n\n[3] Compressed 3D Gaussian Splatting for Accelerated Novel View Synthesis https://openaccess.thecvf.com/content/CVPR2024/papers/Niedermayr_Compressed_3D_Gaussian_Splatting_for_Accelerated_Novel_View_Synthesis_CVPR_2024_paper.pdf\n\n[4]" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Does the method exhibit color shift artifacts during camera movement along the z-direction, as suggested by the concerns regarding linear weighting?\n- Since opacity is not capped at 1, how did you enforce the data range of final color, simply clamp?\n- How do other Gaussian attributes maintain their geometric meaning in this framework? Specifically: Does the mean of Gaussians still correspond to point cloud locations? How about depth extraction?\n- While traditional 3DGS implementations employ culling to manage computational complexity, does this approach require weighted summation of ALL Gaussians? What are the implications for scalability?\n- Please clarify the memory usage analysis: does it encompass only Gaussian attributes or include runtime memory requirements?\nWhat is the memory overhead of spherical harmonics coefficients for view-dependent opacity compared to vanilla 3DGS?\n- Regarding Table 2, could you elaborate on the starred data points showing significantly higher values? How does resource exhaustion due to sorting impact subsequent pipeline stages?\n- Given the potential for hardware-accelerated sorting methods specifically designed for 3DGS, possibly through hardware-software co-design, how might this work adapt to or complement such future developments?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper aims to address a fundamental bottleneck in 3DGS rendering by proposing a sort-free pipeline, distinguishing itself from previous approaches that primarily focused on reducing Gaussian counts.\n- The investigation of various weight derivation functions (direct sum-up, exponential weighted, and linear correction weighted) demonstrates a relative comprehensive experimental analysis effort.\n- Implementation and evaluation on mobile devices effectively demonstrates the practical benefits of the sort-free approach in resource-constrained environments.\n- The method appears to mitigate \"pop\" artifacts, a persistent challenge in volumetric rendering, providing an additional benefit beyond computational efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a novel rasterization backend for 3D Gaussian splatting (3DGS) that addresses the computational and memory overhead associated with depth sorting. Their primary contribution is replacing traditional alpha blending with a weighted sum rendering approach. The rendering is achieved through direct weighted summation of Gaussians, where weights are determined by a parameterized truncated linear function applied to the product of color and opacity. Additionally, they introduce view-dependent optimizable opacity for individual Gaussians, modeled using spherical harmonics.\nTheir weighted sum rendering (WSR) technique demonstrates comparable rendering quality to conventional 3DGS. To validate the advantages of their sort-free approach, the authors implement both vanilla 3DGS and WSR on the Vulkan API and evaluate performance on mobile GPU hardware, achieving a 1.23x speedup." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The theoretical foundation for replacing alpha blending with an algebraic function requires stronger justification. While the proposed representation achieves good results, it lacks the physical insights inherent in traditional alpha blending, which models light obstruction during propagation. The introduction of view-dependent opacity appears necessary for acceptable rendering quality, but the paper would benefit from a more rigorous analysis of why this combination works, rather than empirical validation alone.\n- The preference for linear weights over exponential weights raises questions about robustness. When considering camera movement along viewing rays, exponential weights (with β ≈ 1) should theoretically provide more consistent mixing ratios between Gaussians. Linear weights may introduce color shift artifacts due to distance-dependent mixing ratio variations. While the authors claims \"some artifacts remain visible\" then selected linear weight rather than exponential weight, these issues warrant more detailed examination.\n- The method's generalizability is limited by the transformation of opacity into a view-dependent \"black box\" variable. This fundamentally alters the role of opacity, which traditionally serves as a crucial signal for pruning and densification in many 3DGS applications, potentially limiting compatibility with existing and future research in this domain." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Can the authors provide some qualitative results on the Bicycle scene?\n2. Can the authors provide some videos showing the removal of the popping effect?\n3. Can the authors show some failure cases of the proposed approximation?\n\nI really want to give a positive rating to this paper, because I know it is very difficult to make a sort-free rendering to achieve the same quality as the sort-based version. However, the marginal speed improvement diminishes the significance of this paper. Furthermore, I hope the authors can demonstrate some failure cases of such approximation methods, as they are quite common in traditional sort-free algorithms. Doing this could give the reader a more comprehensive understanding of the proposed method." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "**Motivation**\n* The motivation of this paper is very clear. It clearly illustrates the drawbacks of the sorting requirement of 3DGS and the possible benefit of removing it. Sort-free transparent object rendering is also a well-studied direction in traditional rendering. Combining the algorithms from traditional computer graphics with 3DGS is well-motivated.\n\n**Method**\n* The proposed method transformed the traditional depth-based weighted rendering to a learnable version to better suit the per-scene optimization used in 3DGS, which provides significant performance improvement. \n* The proposed view-dependent opacity is an innovative approach that mimic the occlusion effect. Because this SH parameters are optimized together, it compensates for some performance loss from incorrect occlusion estimation without sorting.\n\n**Experiments**\n* The performance on normal desktop GPU demonstrates a small improvement in rendering speed and comparable rendering speed. In fact, it is already remarkable that a sort-free rendering algorithm can match the performance of a sorting-based rendering algorithm in a scene full of semi-transparent primitives.\n* The performance improvement on the edge device shows a more significant improvement because of the demanding sorting step." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a sort-free variant to the 3DGS algorithm. Because 3DGS is a semi-transparent representation, it requires sorting of all primitives which leads to higher running time, more hardware requirements, and lower compatibility on edge devices. Learning from the traditional sort-free rendering algorithm which replaces the sorting with an estimated weight from depth, this paper proposes to eliminate the sorting process of 3DGS completely. \n\nThis paper further proposes a view-dependent opacity to better simulate the occlusion effect. By learning spherical harmonics to represent the opacity, it allows the 3D Gaussians to appear and disappear at different angles, which could further mimic the occlusion effect.\n\nAs a result, this paper achieves a slightly higher rendering speed and lower GPU memory footprint, better compatibility on edge devices, and comparable rendering quality than 3DGS." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "** Motivation **\n* As shown in the paper, the sorting step is usually considered slow, but actually contributes to only 25% of the total rendering time. The real problem preventing 3DGS based method from matching the rendering speed of traditional methods is the alpha blending step. Although this paper removes the sorting step, it still needs the per-pixel alpha blending of Gaussians of calculated weights. As a result, the speed improvement is rather insignificant. \n* Rather than removing the sorting step, many works have proposed to removing the redundant Gaussians directly. These methods seem to provide much more significant speed boost. Of course these two method can be run in parallel, but it still weaken the importance of the proposed method. \n\n** Method **\n* The proposed method is largely borrowed from the traditional sort-free methods. I think this is acceptable because of the long history of sort-free rendering research in traditional computer graphics, it does negatively affect the novelty of the proposed method slightly.\n\n** Experiments**\n* As mentioned above, the improvement on standard desktop GPU is rather insignificant. Since sort-free rendering is essentially an approximation of the sort-based rendering, it might not be a good idea to choose this method for such a small improvement while there is a possibility of very bad performance on some scenes. \n* Continuing from the last point, sort-free approximations are usually inaccurate on small and thin structures. For example, I noticed that the proposed method on the bicycle scene in MipNeRF-360 dataset is somewhat lower than the original 3DGS method. This scene contains very thin and fine structures on the bicycle which could be a better illustration of whether the proposed method suffers from similar artifacts like the traditional methods. \n* Despite the significant rendering speed improvement on the edge devices, the paper lacks a detailed explanation of how the edge device running out of resources leads to such a big decrease in rendering speed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "1. In section 4.3 the authors reveal that view-dependent opacity is obtained by the addition of a new set of spherical harmonics related to the depth weighting introduced by the authors. In the experimental section, the authors show that this doesn't impact memory usage during rendering and a rationale given in the removal of the sorting step and the potential reduction in the number of Gaussians. However, this appears to be referring to instantaneous memory requirements. How does the introduction of the new set of spherical harmonics for each Gaussian affect post-training storage requirements of the radiance field? Is a greater storage footprint needed for the radiance fields produced by the proposed methods?\n2. Line 53 \"maintain rendering differentiable\" should be \"maintain rendering differentiability\".\n3. Line 141 \"Chris Wyman provide a good survey on these methods\" should be \"Chris Wyman provides a good survey on these methods\"\n4. Line 273, \"However, DIR-WSR does not work well for the complex scenes, \" should be \"However, DIR-WSR does not work well for complex scenes, \"\n5. Line 307 ends with an additional full stop." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This work solves a crucial problem in real-time rendering of Gaussian Splatting based radiance fields. By removing the need for depth-based sorting of the Gaussians before rendering, critical memory requirements are reduced enabling the implementation of the methodology on mobile devices or other applications with limited computational capacity. So I feel that the work represents a strong contribution to the field. Other strengths include:\n1. Clear and easy to read figures and tables clearly detailing the benefits of the proposed method.\n2. A detailed and clear explanation of the proposed method. It is clear what has been done, how this is motivated by previous literature.\n3. A strong experimental validation. Comparisons are made against well-recognized metrics and a the latest state-of-the-art methods. The experimental results support well the conclusions reached." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work attempts to solve a key problem with current Gaussian Splatting techniques, the computational load that occurs during rendering as a result of the non-commutativity of the opacity calculation needed during the rendering process. The authors achieve this by a novel weighting of term added to each Gaussian to replace the traditional opacity term. The weighting term allows for an order independent formulation of the computation of the final colour of the a rendered ray through the field. View-dependence is obtained by the introduction of a spherical harmonic based formulation to the new weighting term similar to that currently used in prior-art Gaussian Splatting methods for view-dependent colour. Performance is evaluated against prior-art techniques." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This work is of quite a high quality. I do not see any specific weaknesses in the methodology or the experimental results. There is one specific minor weakness in the experimental section wherein I would like further information on the memory requirements of the proposed method. I detail the exact question I have in the \"Questions\" section below.\nOne final minor issue is the presence of minor grammar errors in the text as I mention in the \"Questions\" section below." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024sortfree,\ntitle={Sort-free Gaussian Splatting via Weighted Sum Rendering},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y8uPsxR8PN},\nnote={under review}\n}" }, "abstract": { "value": "Recently, 3D Gaussian Splatting (3DGS) has emerged as a significant advancement in 3D scene reconstruction, attracting considerable attention due to its ability to recover high-fidelity details while maintaining low complexity. Despite the promising results achieved by 3DGS, its rendering performance is constrained by its dependence on costly non-commutative alpha-blending operations. These operations mandate complex view dependent sorting operations that introduce computational overhead, especially on the resource-constrained platforms such as mobile phones. In this paper, we propose Weighted Sum Rendering, which approximates alpha blending with weighted sums, thereby removing the need for sorting. This simplifies implementation, delivers superior performance, and eliminates the ``popping'' artifacts caused by sorting. Experimental results show that optimizing a generalized Gaussian splatting formulation to the new differentiable rendering yields competitive image quality. The method was implemented and tested in a mobile device GPU, achieving on average $1.23\\times$ faster rendering." }, "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": [ "Sort-free", "Gaussian Splatting", "Weighted Sum Rendering" ] }, "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/874753d2dcc0bb1e6c2c57fe99f48e16fc76d7d1.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/77c0762093213b39357d3ba6659d3f2f3c72043f.pdf" }, "title": { "value": "Sort-free Gaussian Splatting via Weighted Sum Rendering" }, "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": "Questions:\n- How do you make sure the objects in the image X are identified with the Auto Encoder? if some objects are missing the casual model cannot run properly right?\n- How does the baseline RAP work with LLama? LLama 8B does not take vision input.\n- L285: the description is confusing, why causal mapper is trained using image input?\n- Is LLM baseline just RAP? should just say RAP it in the table. Baseline LM is super confusing, there are many other simpler LLM prompting baselines in planning.\n\nTypos:\n- L269: p_\\phi(E_t | z_t) not p_\\phi(X_t | z_t)" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The integration of CRL and LLM planning is novel and interesting, it is straight forward to integrate it with multiple other LLM-based search algorithms, not only RAP.\n- The paper investigates the form of action representation in the casual world model, and provides detailed results on them." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a framework that combines LLMs with CRL to improve reasoning and planning tasks. The framework leverages both LLMs' common sense knowledge and CRL's \"look-ahead\" ability on causal structures in environments. During inference, LLMs use a causal world model as a simulator for generating and processing action and state descriptions in natural language. This can be well integrated with MCTS algorithm from previous work. Tests on causal inference and planning tasks reveal that this causally-aware approach outperforms traditional LLM methods in complex, long-term planning scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Reasoning via Planning is not the sota method, [1] combines LLM as a world model and LLM as a policy model, with MCTS, should be better than RAP\n- Though not discussed in RAP, other literature [2] [3] suggest increasing the number of LLM calls will substantially improve the search algorithm results. How is the overhead and accuracy of LLM+CRL planning vs. increasing LLM calls with tree-search or iterative prompting?\n- I'm concerned about the ability of the system to work on more complex visual environment. For example, if the accuracy of the casual world model on look-ahead is limited, we should still rely on executing the actions in the real simulator and extract observation there. Related to this concern, a comparison with RAP+the real simulator execution (some oracle) is missing.\n\n[1] Zhao et al., Large Language Models as Commonsense Knowledge for Large-Scale Task Planning, 2023\n\n[2] Yao et al., Tree of Thoughts: Deliberate Problem Solving with Large Language Models, 2023\n\n[3] Zhang et al., ReST-MCTS*: LLM Self-Training via Process Reward Guided Tree Search, 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": "1. How is coordinate-based action representation implemented? A 2-d vector, or simple text like “(2,3)” encoded by the encoder?\n2. What does HB action representation look like and how does it differ from TB? An example of CB, TB, and HB would help to explain.\n3. line 377, “better sample efficiency” would allow TB to perform well in extremely low-data scenarios as well, the reviewer think the experimental results doesn’t support the claim of “better sample efficiency” of TB.\n4. In table 2, the baseline language model and the causal world model both predicts the next state in natural language, how is it evaluated against the ground-truth next state?\n5. How big is the set of annotated images used to train the causal mapper m_\\theta?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper proposed a method to learn a causal world model from a sequence of image states and text action descriptions, and demonstrated superior performance in the accuracy of the learned world model.\n- Presentation of the paper is clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed to learn a causal world model from a sequence of states and actions, where the states are images and actions are described in natural language. The proposed method builds upon BISCUIT and proposes representing the actions in natural language. The paper conducted experiments in GridWorld and iTHOR environment and the R^2 scores of the learned representation are claimed to be higher in low-data regimes. Experiments also show that the learned causal world model is more accurate at predicting the next state, and improves planning performance when compared to a general LLM (Llama 3)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some components of the framework would not be available in more realistic environments, e.g. 1) a set of annotated images with ground-truth causal variables is used in training, which is likely not available as we may not know the causal variables for more realistic environments; 2) a rule-based state description generator may not be available for complex environments where we don't know what are the true causal variables.\n2. Given the simplicity of the environments, and that the proposed method is trained on these particular domains while the baseline is a general LM, the superior performance of the learned causal world model is less convincing. I would suggest comparing with a supervised fine-tuned version of the baseline LM. Since the proposed method uses a set of annotated images to train the causal mapper, supervised fine-tuning is possible with these annotated data: use the state description generator to convert the ground-truth causal variables of the states to natural language, and we can obtain a sequence of natural language action and natural language description of the states." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Can the authors elaborate on how the framework handles situations where the causal factors are not easily identifiable or where the causal relationships are highly complex?\n\nFollowing on the above point, how were the parameters chosen for the various components of the framework, and how sensitive are the results to these parameters?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The exploration of text-based action representations is particularly interesting, as it shows potential advantages in low-data regimes. I particularly liked the connection to the RL decision making problem. The results indicate that the proposed framework can potentially improve performance in causal inference and planning tasks, which is valuable for the broader ICLR community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel framework that integrates Causal Representation Learning (CRL) with Large Language Models (LLMs) to enhance reasoning and planning capabilities in interactive environments. The framework utilizes a causal world model that maps high-dimensional state representations to causal variables, which are linked to natural language expressions. This integration allows the LLM to interact with the causal world model, effectively simulating multiple possible futures before taking actions. The approach is evaluated on causal inference and planning tasks across different environments and temporal scales, demonstrating that the causally-aware method outperforms traditional LLM-based reasoners, especially for longer planning horizons." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Please see the 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": 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": "+ Do we have examples of decoded text (from CRL to LLM)? Is that human-readable? Is there any possible insights with analysis of the decoded text?\n+ (Clarification) In my understanding, the performance increase comes from (1) the text decoded from CRL makes it easier for LLM to reason and plan (2) the CRL model has a higher accuracy in predicting future states. If my understanding is correct, do we have ablation experiments on the effect of the two aspects?\n+ (Clarification) Does the Baseline LM refer to RAP + LLaMA 3 or LLaMA3 with other prompting methods?\n+ In Table 1, **TB** outperforms **HB** starting from 1.0%, outperforming **HB** much at **1.5%**, but performs weaker than **HB** in 100%. What are the trend between 1.5% and 100% subsample rate? Does it mean the necessity of **TB** decreases when the data scales up enough?\n+ The **HB** settings shall contain all information in the **TB** settings. Why would the model performs worse than **TB** with increased subsample percentage? Since **HB** performs better in 100% ($10^6$ image states), would it be better to use **HB** in subsequent experiments?\n+ Can you include SNA (success weighted by number of actions) in the planning results? It's a common metric to combine success and # of steps.\n+ (Typo) Page 4 footnote, prioi -> prior\n+ (Typo) Figure 1, missing an arrow from $\\bf C^1$ to Text Decoder." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper proposes to introduce CRL for world model to better understand the causal structure of environments. This intuitive idea provides a solid future approach to further improve the performance of world models, which is essential to build with intelligent agent with System 2 reasoning abilities.\nThe author analyzes the underlying math principles of the proposed CRL model and shows with experiments of the performance in metrics of comparison between induced state variables and ground-truth, and multi-step high-level accuracy of the predicted states. The experiments show large performance increase than baseline pure language model.\nThe author also use the result causal world model to do experiments on synthetic planning tasks. The results show consistent superior performance to baseline LM, especially on longer planning horizons." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The author proposed to introduce causal representation learning (CRL) to better understand the causal structure of a environment, and they incorporate such CRL with LLM to enable causally-aware planning by mapping causal representations to natural language." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "+ It's not clear how much the text decoded from CRL model helps the LLM reasoning. It would be clearer if the author can show some examples and/or conduct ablation experiment on this. (also see questions)\n+ The experiments compare **CB**, **TB**, and **HB** are based on the low-data scenarios, but the decision to use **TB** in subsequent experiments is for 100% data. This decision lacks a continuity of experiment settings, and it's unclear which setting would be the best for subsequent experiments. (also see questions)\n+ The paper is a good starting of causal world model studies, but I have to mention that the experiment environment is synthetic and relatively easy. It would be more fascinating to use more general environments, and there leaves doubt whether the method can be easily scaling up to more general or even real world scenarios." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Improving LLM planning capabilities using learned causal representations" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024language,\ntitle={Language Agents Meet Causality -- Bridging {LLM}s and Causal World Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y9A2TpaGsE},\nnote={under review}\n}" }, "abstract": { "value": "Large Language Models (LLMs) have recently shown great promise in planning and reasoning applications. These tasks demand robust systems, which arguably require a causal understanding of the environment. While LLMs can acquire and reflect common sense causal knowledge from their pretraining data, this information is often incomplete, incorrect, or inapplicable to a specific environment. In contrast, causal representation learning (CRL) focuses on identifying the underlying causal structure within a given environment. We propose a framework that integrates CRLs with LLMs to enable causally-aware reasoning and planning. This framework learns a causal world model, with causal variables linked to natural language expressions. This mapping provides LLMs with a flexible interface to process and generate descriptions of actions and states in text form. Effectively, the causal world model acts as a simulator that the LLM can query and interact with. We evaluate the framework on causal inference and planning tasks across temporal scales and environmental complexities. Our experiments demonstrate the effectiveness of the approach, with the causally-aware method outperforming LLM-based reasoners, especially for longer planning horizons." }, "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", "Causality", "Causal Representation Learning", "Language Agents", "Planning" ] }, "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/b3af4d8138f60e41a97ffc17d99bacdcaf62f711.pdf" }, "presentation": null, "primary_area": { "value": "applications to robotics, autonomy, planning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/4922a3e4fa9902d7ff0d5cd9d8bb3ae148ea1565.zip" }, "title": { "value": "Language Agents Meet Causality -- Bridging LLMs and Causal World 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": "See 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 targets a clear challenge in backdoor defense by eliminating the need for clean data, which is relevant to practical applications.\n\n- The experimental evaluation is thorough, encompassing multiple datasets, attack types, and model architectures. The comparison against existing defense methods provides a meaningful context for the method's effectiveness.\n\n- The authors conduct detailed ablation studies examining the impact of various hyperparameters, which helps in understanding the method's sensitivity and optimal configuration." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the limitation of existing backdoor attack defenses, which typically require an auxiliary clean dataset that may be difficult to obtain. The authors propose a clean-data-free, end-to-end method to mitigate backdoor attacks. Their approach leverages two dynamically identified pools of data: one from open set recognition-based splitting and another from altruistic model-based splitting. These pools are then utilized in the main training loop, which the authors demonstrate to be effective in producing backdoor-free models, even when trained on poisoned datasets.\n\nThe authors validate their method using three benchmark datasets and test it against seven representative backdoor attacks, including both dirty-label and clean-label attacks. They compare their approach with five existing backdoor defenses and evaluate performance across two model architectures: ResNet-18 and MobileNet-v2." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The study primarily focuses on supervised image classification tasks. This limitation should be explicitly stated in both the abstract and introduction to better set reader expectations.\n\n- The paper's motivation could be strengthened by acknowledging recent developments in clean data acquisition methods. Notable omissions include:\n\n *Zeng et al. (2023, Usenix Security)* on clean subset extraction from poisoned datasets\n\n *Pan et al. (2023, Usenix Security)* on backdoor data detection methods\n\n These omissions affect the paper's premise that clean data acquisition is inherently impossible.\n\n\n- Technical Issues:\n\n A typographical error in line 331: \"Mobileent-v2\" should be \"MobileNet-v2\"\n\n**References:**\n\n*Zeng et al. (2023)*. \"META-SIFT: How to Sift Out a Clean Data Subset in the Presence of Data Poisoning?\" Usenix Security, 2023.\n\n*Pan et al. (2023)*. \"ASSET: Robust Backdoor Data Detection Across a Multiplicity of Deep Learning Paradigms.\" Usenix Security, 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "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 introduces a novel method for defending against backdoor attacks without the need for clean data, addressing a significant limitation in previous methods.\n2. Extensive experiments across multiple datasets and attack scenarios demonstrate the robustness and effectiveness of the proposed method, outperforming several state-of-the-art defenses.\n3. The paper details multiple defensive strategies that contribute to its effectiveness, such as class completion and selective dropping, which are well-integrated into the defense strategy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the challenge of defending deep neural networks against backdoor attacks in the absence of clean data subsets. It introduces a novel defense mechanism called Bi-perspective Splitting Defense (BSD) that uses semantic and loss statistics characteristics for dataset splitting. The approach involves two innovative initial pool splitting techniques, Open Set Recognition-based Splitting (OSS) and Altruistic Model-based Splitting (ALS), and it enhances defense through subsequent updates of class completion and selective dropping strategies. The method demonstrates substantial improvements over state-of-the-art defenses across multiple benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The complexity of the proposed method involving multiple models and sophisticated data splitting strategies could be a barrier to practical deployment and computational efficiency.\n2. While the method is empirically successful, the paper could be improved by providing deeper theoretical insights into why the specific strategies employed are effective.\n3. The effectiveness of the method might depend on specific neural network architectures, and its adaptability to different or future architectures is not fully addressed.\n4. The paper could benefit from a more detailed discussion on scenarios where BSD might fail or be less effective, which would be crucial for practical applications and future improvements." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "Yes, Privacy, security and safety" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Listed in the weakness of the paper. \n\nScore can be improved if concerns listed above are resolved." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper explores a novel method to defend backdoor attacks which does not rely on clean subsets of data.\n\n- The proposed method is simple but effective and the good performance obtained by the experiments strongly supports this point.\n\n- The ablation study is organized well to clearly demonstrate the whole proposed method. And it makes the paper easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Backdoor attacks threaten deep neural networks (DNNs) by embedding hidden vulnerabilities through data poisoning. While researchers have explored training benign models from poisoned data, effective defenses often rely on additional clean datasets, which are challenging to obtain due to privacy concerns and data scarcity.\n\nTo tackle these issues, the paper proposes the Bi-perspective Splitting Defense (BSD), which splits the dataset based on semantic and loss statistics using open set recognition (OSS) and altruistic model-based data splitting (ALS). This approach enhances clean pool initialization and includes strategies to prevent class underfitting and backdoor overfitting.\n\nExtensive experiments on three benchmark datasets against seven attacks show that BSD is robust, achieving an average 16.29% improvement in Defense Effectiveness Rating (DER) compared to five state-of-the-art defenses, while maintaining minimal compromise in Clean Accuracy (CA) and Attack Success Rate (ASR)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I wonder **why the proposed BSD can effectively resist the clean-label backdoor attacks.** Clean-label backdoor attacks manipulate samples from the target class while keeping their labels unchanged. Since the BSD framework formulates backdoor defense within a semi-supervised learning context, it seems that whether the clean-label poisoned samples are part of the labeled or unlabeled subset, **the model can still associate the trigger with the target label in clean-label backdoor attacks.** Therefore, I am curious about how BSD manages to effectively resist the three clean-label attacks demonstrated in Table 2.\n\n- I recommend conducting further experiments to assess whether BSD can successfully defend against backdoor attacks **with different target labels.** This could provide valuable insights into the robustness of the defense mechanism across various scenarios.\n\n- Typos: #Line 848 --- #Line 849, 5*10e-4." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In the exp setup, it’s reported to use 7 backdoor attacks while only 3 of them are presented in Table 1. That is to say, the performance against the remaining 4 attacks on the three benchmark datasets is not shown in the paper.\n\n2. Considering the given threat model (i.e., in-training clean-data-free defense), it seems that sota defenses D-ST&D-BR [1] could also serve as baselines. These methods leverage the sensitivity of poisoned samples to transformations, which is quite different from the semantics and losses used in this paper; thus, it would be interesting to compare them and show that the proposed metrics are more accurate.\n\n3. Is the accuracy of identifying the target class reported in the paper? What if the identified class is wrong, will it affect the performance of the following steps?\n\n4. A type in Line 83: “first initialize” -> “first initializes”\n\n[1] Chen, W., Wu, B., & Wang, H. (2022). Effective backdoor defense by exploiting sensitivity of poisoned samples. Advances in Neural Information Processing Systems, 35, 9727-9737." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed defense does not require any extra clean data.\n\n2. They compare the problem of identifying clean target samples from poisoned samples with the open set recognition-based splitting problem of identifying UUCs from UKCs; inspired by which, they propose the identification mechanism." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an in-training clean-data-free backdoor defense method where the defender is required to train a clean model from scratch given a poisoned dataset without the need of any additional clean data. The key to success is to distinguish between clean samples and poisoned samples. To this end, they propose a novel identification mechanism which involves two main procedures. The first procedure is initializing a pool of clean samples and a pool of poisoned samples based on open set recognition-based splitting and altruistic model-based splitting. The second procedure is improving these pools with class completion and selective dropping strategy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited applied scenarios: The proposed defense is limited to backdoor attacks with a single target class. This is because the first step in identification involves identifying the single target class from other classes; however, in some popular attacks like BadNet’s all-to-all attacks, all classes are target classes, which disables the proposed defense.\n\n2. The proposed method seems a bit complex as it includes two main steps—pool initialization and pool updating—and each step further involves two sub-steps, respectively. These steps aim to distinguish samples from different perspectives. Only after the total four steps, the poisoned samples are filtered out from clean samples. A complex mechanism is totally fine; but, there are two related issues: 1) does the necessity of pool updating validates that the effectiveness of pool initialization is not very good? 2) it seems that the effectiveness of each step highly depends on the performance of previous steps, e.g., if the identification of the target class is wrong, then all subsequent steps are useless. That is to say, accumulative errors may exist in the proposed method and which could lead to bad defense performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The authors have already provided comprehensive analyses and experiments.\n\nHowever, an unresolved question remains: How effectively can the proposed method be applied to scenarios involving multi-target attacks?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The strengths of the proposed Bi-perspective Splitting Defense (BSD) method highlighted in the paper include:\n\n1. **Efficiency in Utilizing Available Data**: BSD makes effective use of target labels and identifies clean samples from the poisoned dataset, reducing the dependency on additional clean data, which might be scarce or raise privacy concerns.\n\n2. **Mitigation Strategies**: BSD incorporates class completion and selective dropping, helping to avoid issues like class underfitting and backdoor overfitting that could otherwise degrade model performance.\n\n3. **Balanced Performance**: While enhancing security, BSD maintains high Clean Accuracy (CA) and manages Attack Success Rate (ASR), ensuring that the model remains accurate on clean data while defending against adversarial inputs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper discusses a method to defend against backdoor attacks in deep neural networks (DNNs) by training reliable models from poisoned datasets, addressing the challenge of lacking additional clean data due to privacy concerns or scarcity. It proposes a Bi-perspective Splitting Defense (BSD) that uses open set recognition-based splitting (OSS) and altruistic model-based data splitting (ALS) to divide the dataset effectively, initializing a pool of clean samples. BSD also employs class completion and selective dropping to prevent class underfitting and backdoor overfitting. Experiments on three benchmark datasets and against seven attacks show that BSD improves Defense Effectiveness Rating (DER) by an average of 16.29% compared to five state-of-the-art defenses, while maintaining high Clean Accuracy (CA) and managing Attack Success Rate (ASR) effectively, thus providing robust security without needing extra clean data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The weaknesses identified in the paper are outlined as follows:\n\n1. **Dependency on Target Recognition**: A significant limitation of the proposed methodology lies in its dependence on accurate target recognition. This dependency introduces a vulnerability; if an advanced attack manages to circumvent the existing detection mechanisms, the entire method could become ineffective. In my opinion, the superior performance is predicated on the assumption that the target class can be accurately identified. Additionally, the framework's effectiveness diminishes when dealing with complex scenarios such as backdoor attacks that involve multiple targets, including those with dual-target labels or all-to-all attack configurations.\n\n2. **Absence of Comparative Analysis with Relevant Work**: The proposed dual-model training framework, which leverages an auxiliary model to support the primary model, shares similarities with the approach detailed in \"[1]\". However, the manuscript lacks a comparative analysis of this work. I see the proposed method has many differences from [1], but incorporating such a comparison would significantly bolster the credibility and relevance of the current submission by providing a clearer differentiation and understanding of the advantages and limitations relative to existing methodologies.\n\nReference:\n[1] The Victim and The Beneficiary: Exploiting a Poisoned Model to Train a Clean Model on Poisoned Data, ICCV 2023 Oral paper" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024biperspective,\ntitle={Bi-perspective Splitting Defense: Achieving Clean-Data-Free Backdoor Security},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y9Lbr6vFHF},\nnote={under review}\n}" }, "abstract": { "value": "Backdoor attacks have seriously threatened deep neural networks (DNNs) by embedding concealed vulnerabilities through data poisoning. To counteract these attacks, training benign models from poisoned data garnered considerable interest from researchers. High-performing defenses often rely on additional clean subsets, which is untenable due to increasing privacy concerns and data scarcity. In the absence of clean subsets, defenders resort to complex feature extraction and analysis, resulting in excessive overhead and compromised performance. In the face of these challenges, we identify the key lies in sufficient utilization of the easier-to-obtain target labels and excavation of clean hard samples. In this work, we propose a Bi-perspective Splitting Defense (BSD). BSD splits the dataset using both semantic and loss statistics characteristics through open set recognition-based splitting (OSS) and altruistic model-based data splitting (ALS) respectively, achieving good clean pool initialization. BSD further introduces class completion and selective dropping strategies in the subsequent pool updates to avoid potential class underfitting and backdoor overfitting caused by loss-guided split. Through extensive experiments on 3 benchmark datasets and against 7 representative attacks, we empirically demonstrate that our BSD is robust across various attack settings. Specifically, BSD has an average improvement in Defense Effectiveness Rating (DER) by 16.29\\% compared to 5 state-of-the-art defenses, achieving clean-data-free backdoor security with minimal compromise in both Clean Accuracy (CA) and Attack Success Rate (ASR)." }, "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": [ "Trustworthy AI", "Backdoor Defense", "Deep Neural Networks" ] }, "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/0d3930362c0349fd81800df206e42246c23cf945.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/1acc64fa528f8121ee0e865b4f1aab70836b49f8.zip" }, "title": { "value": "Bi-perspective Splitting Defense: Achieving Clean-Data-Free Backdoor Security" }, "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": "N/A" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The theoretical analysis seems correct." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a theoretical analysis of the role of rank within attention mechanisms. It challenges the prevailing practice of employing low-rank attention and discusses the implications related to the selection of the number of heads. The author establishes that low-rank attention exhibits inferior performance compared to full-rank attention, indicating that the adoption of a higher rank has the potential to enhance attention performance. Preliminary experiments are conducted utilizing toy examples with synthetic data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The rank of attention is a significant hyperparameter in the design of transformers. A common convention involves the utilization of low-rank attention, typically establishing the number of heads as ( H = d/r ). This paper, however, contests this design choice, proposing that a higher rank can enhance performance. It is crucial to note that the paper does not address the speed-accuracy trade-off associated with this adjustment. It is widely recognized that high-rank attention may yield superior performance at the expense of increased computational costs. When evaluating overall performance, particularly in terms of accuracy within a predetermined computational budget, prevailing practices may ultimately provide more favorable outcomes.\n\n2. The experiments presented in this study lack robustness, as they are primarily limited to toy experiments. I would appreciate observing performance metrics derived from real-world data applied to standard transformer sizes. It is well established that theoretical performance often diverges from practical outcomes in deep learning; thus, empirical experimentation is essential.\n\n3. This work indicates that shallow transformers may experience limitations due to low-rank attention. However, it is imperative to ascertain how these limitations manifest in deep transformers, as shallow transformers are not commonly employed in practice. If this limitation has been substantially mitigated in deep transformers, it may render further examination of this issue unnecessary." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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 simple experiments or additional references to other studies and conclusions be designed to intuitively show the impact of the low-rank problem on the performance of mainstream Transformer models?\n2. Could you further elaborate on how the proposed “majority voting” method for improving low-rank models enhances mainstream Transformer models and validate this with relevant experiments? For the experiments, model selection could refer to those in Appendix B.1 and the models used in [1][2], while the datasets could refer to those in [1][2] or other widely recognized benchmark datasets.\n[1] Bhojanapalli S, Yun C, Rawat A S, et al. Low-rank bottleneck in multi-head attention models[C]//International conference on machine learning. PMLR, 2020: 864-873.\n[2] Shazeer N, Lan Z, Cheng Y, et al. Talking-heads attention[J]. arXiv preprint arXiv:2003.02436, 2020.\n3. Also, please refer to weaknesses for other concerns." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. A deeper exploration of the low-rank problem in Transformer models.\n2. The paper is well written and easy to follow.\n3. Authors provide ample mathematical proofs to support their conclusions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This article examines the limitations and potential of low-rank attention mechanisms in transformer models, demonstrating that while low-rank attention heads require significantly more heads to match the performance of full-rank models in approximating functions like nearest neighbors, and these limitations can be mitigated by increasing the depth of the model. Through theoretical analysis and empirical validation, the study highlights that full-rank models inherently possess superior representational capabilities, especially with fewer heads, and suggests that adding more layers could partly overcome the deficiencies of low-rank models, though at the cost of increased computational complexity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors mentioned after Theorem 2 that the theoretical framework should be extendable to cases where N>2. Could you provide more specific explanations for the reasoning behind this inference? This would help further understand the applicability of your theory to specific problems.\n\n2. Although the authors have demonstrated theoretically and experimentally that low-rank attention models are insufficient for fitting certain functions in various scenarios and are significantly weaker than full-rank attention models, further clarification is needed on how these issues impact current mainstream Transformer models (such as the new models shown in Table in Appendix B.1), how the proposed methods in the paper apply to these models, and how performance improvements are achieved. I believe that related experimental results and methodological extensions would greatly help illustrate the contribution of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 Weaknesses" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Novel perspective on attention mechanisms: The paper offers a fresh perspective on the role of rank in attention mechanisms by using a simple and natural target function based on nearest neighbor search that can be represented using a single full-rank attention head for any context length, which is an interesting aspect of transformer architectures.\n\n2. Theoretical and Empirical Rigor: It combines theoretical proofs with empirical experiments, providing a robust exploration of the implications of low-rank attention on model capacity and efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper \"THE LOW-RANK BOTTLENECK IN ATTENTION\" investigates the impact of the rank of attention matrices on the representational capacity of attention-based mechanisms, particularly in transformers. It challenges the common practice of using low-rank attention and proposes that the rank can significantly influence the model's ability to approximate certain functions. Specifically, the authors present a simple and natural target function based on nearest neighbor search that can be represented using a single full-rank attention head for any context length. The paper presents theoretical analysis and empirical experiments to support its claims, suggesting that increasing the rank or the number of attention heads may lead to more expressive and parameter-efficient models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The results may rely heavily on the assumption of rotational invariance in the data distribution, which may not hold in all real-world applications.\n\n2. To make it easier for readers to understand, I kindly suggest that the authors explain in more detail the differences between this paper and previous work [1].\n[1] Low-Rank Bottleneck in Multi-head Attention Models. ICML 2020\n\n3. Can the proposed method demonstrate its effectiveness on more attention-based models?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. I am wondering the reason why authors choose to analysis nearest neighbor functions and are there any other choices of functions." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "Transformer have dominated many areas but there have been few studies on the choices of numbers of attention heads and dimensions used in attention mechanism. This paper raises doubts about this which is valuable for community to pay attention.\n\n1. The paper is well written and easy to follow. In-depth theoretical explanations are provided.\n\n2. For a simple and natural target function -- nearest neighbor function, authors show low-rank attention is fundamentally weaker than full-rank attention even when choosing very large head numbers.\n\n3. Also, this paper explores the solutions to mitigate the weakness of low-rank attention" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper questions the conventional number that community use for the rank of attention matrices and the number of attention heads. Authors provide in-depth theoretical explanations and experiments to support their arguments. In the high-accuracy regime, the required number of heads is growing exponentially to remain the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper studies only is limited to shallow transformers which are not practical to large model." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We prove that attention layers with a small query/key dimension can have weak expressive power. This weakness cannot be overcome without using a massive number of attention heads." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024the,\ntitle={The Low-Rank Bottleneck in Attention},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y9Xp9NozPR},\nnote={under review}\n}" }, "abstract": { "value": "Attention-based mechanisms are widely used in machine learning, most prominently in transformers. However, hyperparameters such as the rank of the attention matrices and the number of attention heads are scaled nearly the same way in all realizations of this architecture, without theoretical justification. In this paper, we prove that the rank can have a dramatic effect on the representational capacity of attention. This effect persists even when the number of heads and the parameter count are very large. Specifically, we present a simple and natural target function based on nearest neighbor search that can be represented using a single full-rank attention head for any context length, but that cannot be approximated by low-rank attention unless the number of heads is exponential in the embedding dimension, even for short context lengths. Moreover, we show that, for short context lengths, adding depth allows the target to be approximated by low-rank attention. For long contexts, we conjecture that full-rank attention is necessary. Finally, we present experiments with standard multilayer transformers that validate our theoretical findings." }, "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": [ "Learning theory", "Expressive capacity", "Expressive power", "Transformer", "Attention" ] }, "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/372a20c92cbd4f90e77460e8f6472635bda70f58.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/ef624b8c3cbf03475df39fefe6a8cd42d415e7b1.zip" }, "title": { "value": "The Low-Rank Bottleneck in Attention" }, "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": "It seems to me that if the update function u of definition 1 is a constant function, no information is preserved and the constructed MDP could not have the same value function as the NCMDP. Is this right? If it is, this sort of choice should be excluded in definition 1.\nLater in the same paragraph, the authors note that an h_t recording all reward history is sufficient to construct an MDP but the state space is exponentially large so there must be some tradeoff between state space and \"Markovness\". How much information can be lost from the full history before we lose the Markov condition? Is there a characterisation of the functions u that satisfy this property?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Definition 1 and the possible applications of the method are interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In traditional MDPs, the agent is tasked with the optimisation of the policy value function that is the expected sum of the rewards. In a non cumulative markov decision process, the value function is replaced by the expected value of an arbitrary function of the rewards. \nThe authors propose a method to translate an NCMDP into a regular MDP and show the expected returns of the two are equal and apply MDP solving methods to NCMDPs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The variables u, h and $\\rho$ are discussed in the paragraph following the statements of equations 3,4,5. You should introduce them before.\nAlso, their description is too vague. For example \"This can be achieved by extending the state space with ht , which preserves all necessary\ninformation about the reward history\" does not give me a good sense of what the function u should be. \nI suggest the statement of theorem 1 be rearranged to: assumptions then conclusion, instead of the current: assumption then conclusion then more assumptions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- The paper gives very specific examples of construction of $u$ and $\\rho$ that allows an efficient reduction from NCMDP to MDP of small augmented state size. Can you give examples of more general classes of non-cumulative objective function $f$ that induces $h_t$ that is of reasonably small dimensions?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The NCMDP setting considered in this paper fits many applications that does not directly fit into the MDP setting with cumulative rewards. Some examples include the weakest-link problem in network routing which maximizes minimum reward, the Sharpe ratio in finance which maximizes the mean divided by standard deviation. \n- The paper provides a straightforward solution to NCMDPs by first map an NCMDP to a standard MDP, which allows direct application of black box MDP solvers. \n- It provides comprehensive empirical results applying this method to a range of environments with non-cumulative reward objective, including classical control, portfolio optimization with Sharpe ratio, discrete optimization with lost cost objective." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the problem of finding an optimal policy for a special class of decision problem called non-cumulative Markov decision process (NCMDPs) where instead of the sum of rewards, it aims to maximize the expected value of an arbitrary function of the rewards. The paper solves NCMDPs by mapping it to standard MDPs, allowing direct application of MDP solver for NCMDPs. It also performs numerical experiments of classical control, portfolio optimization, and discrete optimization that use NCMDP objectives, and shows that their method improves both training time and final performance compared with standard MDP solvers with cumulative reward." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The mapping from NCMDP to MDP provided in equations (3)-(5) augments the state with $h_t$ that represent necessary information for reward history, and is updated as $h_{t+1}=u(h_t,\\tilde{r}_t)$ and satisfies $r_t=\\rho(h_t,\\tilde{r}_t)$. For arbitrary reward function $f(r)$, an essential factor to ensure the mapping to MDP is of reasonable size is to find an efficient functional form of $u$ and $\\rho$ that summarizes this information from reward history. However, the paper only shows a list of examples of special examples of $f$ and does not specify how to construct $u$ and $\\rho$ for arbitrary function $f$ in general, and does not provide an bound to the resulting state dimensions, which at worst case must contain the set of all historical rewards. I think this makes the overall contribution of the paper limited. \n- In the experimental section, for each environment, the paper compares their method of MDP reduction (with modified state and reward) with an RL agent with cumulative objective. This does not seem like a very fair comparison given the cumulative objective will be the wrong objective in each of these environments, and it is clear that correct specification of the objective will lead to better 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": 3 }, "primary_area": null, "questions": { "value": "* Can the proposed mapping of NCMDPs to MDPs be generalized to infinite-horizon MDPs? (when $f$ is defined as a function of set.)\n\n* It seems that if we transformed an NCMDP to a standard MDP, the resulting standard MDP is governed by the dynamics of $s_t$ and $h_t$ can be viewed as an auxiliary state. I am curious whether there are some connections with the framework called MDPs with latent dynamics (see [1], [2]).\n\n[1] Simon Du, Akshay Krishnamurthy, Nan Jiang, Alekh Agarwal, Miroslav Dudik, and John Langford. Provably\nefficient RL with rich observations via latent state decoding. In International Conference on Machine\nLearning, 2019.\n\n[2] Amortila, Philip, et al. \"Reinforcement Learning under Latent Dynamics: Toward Statistical and Algorithmic Modularity.\" arXiv preprint arXiv:2410.17904 (2024)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The authors consider the problem of control in NCMDPs, which is important and understudied in the RL community.\n\n* The proposed mapping of NCMDPs to standard MDPs is simple and intuitive, and can effectively solve many problem instances of NCMDPs.\n\n* The empirical studies are concrete, showing both the necessity of considering the problem of control in NCMDPs and the effectiveness of the proposed mappings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors consider the problem of control in finite-horizon Non-cumulative MDPs (NCMDPS). They propose a general mapping of NCMDPs to standard MDP, and hence the problem can be solved by existing standard RL methods. They use a series of numerical experiments to illustrate the NCMDP problems and the proposed transformations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The paper has technical flaws, i.e., the function $f$ is ill-defined. In the original problem statement, $f$ is defined as a function on $\\mathbb{R}^T$. But later on, the authors also use notations like $f(r_1,...,r_t)$, where $f$ should be treated as a function on $\\mathbb{R}^t$. I think it may be helpful to define $f$ as a function of a set ($f(\\\\{r_1,...,r_t\\\\})$, $\\\\{r_1,...,r_t\\\\}$ is the set of $t$ rewards, $t$ can be any integer between $1$ and $T$). This works for all problem instances in Table 1.\n\n* The authors overstate their contributions. They claim their methods work for arbitrary functions $f$. However, as indicated in the previous comment, Definition 1 is invalid for general $f$ on $\\mathbb{R}^T$. Even if $f$ is treated as a function of a set, it turns out only for certain choices of $f$ there exist fix-size $h_t$ and corresponding functions $u$ and $\\rho$. I think the authors should try to find a classification of functions f with constant-size state adaptions. Even some preliminary results such as sufficiency conditions or necessary conditions would be helpful.\n\n* I think in empirical studies the authors should compare their methods with prior works on the control of NCMDPs with a specific choice of objectives and I am curious if the proposed method (which is more general) can produce comparable performances." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1) What are the off-policy implications of this method?\n2) A regular average would appear to be a common and useful non-cumulative objective to include in the paper. Was there a reason that it was not included?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper provides a theoretically-sound and easy-to-use framework that makes it possible to optimize non-cumulative objectives, by mapping such objectives to the regular cumulative formulation, such that optimizing the cumulative objective results in the optimization of the non-cumulative objective. This framework clearly has several potential use-cases in various domains. The paper is polished, well-written, and the technical concepts are explained in a relatively clear manner." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a decision-making framework that allows for the optimization of non-cumulative objectives, by mapping such objectives to the regular cumulative formulation, such that optimizing the cumulative objective results in the optimization of the non-cumulative objective. The authors show that this MDP framework satisfies important theoretical properties that make it possible to find an optimal policy, and provide several empirical experiments in support of this." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have two concerns regarding this paper:\n\nThe first of which is that while the authors do a good job explaining how their framework can be used, they fail to properly formalize the theoretical conditions that they are operating in. For example, they do not address whether the state and action-spaces are or can be finite/infinite, if there is a restriction on the type of policy that can be used (stationary, stochastic, etc.), as well as whether there are implications of their framework being used in off-policy settings. Because of this, the extent to which their framework can be utilized is unclear.\n\nMy second concern is related to the experiments. In particular, the comparison of non-cumulative objectives to the cumulative objects seems unintuitive, given that no-where in the paper do the authors claim the non-cumulative objectives are better than the cumulative objectives. Rather, it would be more intuitive, and consistent with the rest of the text, if the experiments showed that the proposed methods indeed are able to find the optimal policy for the non-cumulative objective. As such, it would appear that the experiments in the current draft of the paper do little to support the claim that the proposed methods can be used to optimize non-cumulative objectives.\n\nAs such, I am open to increasing my score if the authors can properly formalize the theoretical conditions that they are operating in, and either 1) the authors can convince me that the experiments in the current draft are adequate, or 2) the authors improve the existing experiments in such a way that they address my concerns, or 3) the authors can provide additional experiments that address my concerns." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We extend the class of problems that can be tackled through reinforcement learning by introducing a general mapping of decision processes with non-cumulative objectives to standard Markov decision processes." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024tackling,\ntitle={Tackling Decision Processes with Non-Cumulative Objectives using Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y9e1tcWlme},\nnote={under review}\n}" }, "abstract": { "value": "Markov decision processes (MDPs) are used to model a wide variety of applications ranging from game playing over robotics to finance. Their optimal policy typically maximizes the expected sum of rewards given at each step of the decision process. However, a large class of problems does not fit straightforwardly into this framework: Non-cumulative Markov decision processes (NCMDPs), where instead of the expected sum of rewards, the expected value of an arbitrary function of the rewards is maximized. Example functions include the maximum of the rewards or their mean divided by their standard deviation. In this work, we introduce a general mapping of NCMDPs to standard MDPs. This allows all techniques developed to find optimal policies for MDPs, such as reinforcement learning or dynamic programming, to be directly applied to the larger class of NCMDPs. Focusing on reinforcement learning, we show applications in a diverse set of tasks, including classical control, portfolio optimization in finance, and discrete optimization problems. Given our approach, we can improve both final performance and training time compared to relying on standard MDPs." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "reinforcement learning", "markov decision processes", "discrete 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/418398a45b4ea19833ed50f139ad0836c17f9445.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": "Tackling Decision Processes with Non-Cumulative Objectives using Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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": "- Do the authors have any information on the identities/demographics of their human annotators? Ignoring settings where LLMs should decline to respond because of legal liability, the choice of whether an LLM should respond/decline seems entirely subjective. Different individuals will have different assessments, because of their personal politics, religious beliefs, cultural backgrounds, and more. So largely, this benchmark would seem to measure acceptance/decline accuracies grounded against the specific human pool the authors surveyed. Having more information about this pool would be helpful! \n- The underlying message of the paper is that context matters in determining whether an LLM should respond to a query. For instance, an LLM intended to train medical professionals should probably respond to medical questions. Why is it fair, then, to evaluate LLMs like GPT-4o-mini or Claude, which are intended to be used in the “layperson chatbot” context? I’m not sure I see why merely providing a “context” to these chatbots should actually alter their behavior. Given that a prevailing concern with these models is that users might employ deception to elicit responses from the models that developers do not desire (i.e., via jailbreaking or prompt-engineering)–shouldn’t these LLMs also decline to respond when the context is provided? The argument that these LLMs should respond when context is provided presumes that users will always be truthful when providing information in the prompt. \n- Could the authors share more information on the contexts?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Overall the paper is extremely well-written and clear.\n- The core idea tackled by the paper–understanding how context might (or should) affect answer/decline decisions by LLMs–is interesting.\n- The study methodology is extremely rigorous. \n- The experiments are also well-done and interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "- The paper develops a benchmark for evaluating LLMs’ refusal accuracy for problematic queries. The key point of differentiation is that the authors focus on “context.” This is the idea that auxiliary information might change whether or not an LLM should refuse to answer a query. \n- The paper evaluates a range of LLMs on this benchmark, and finds that LLMs are generally over-moderated.\n- They also conduct an extensive survey of individuals to measure how context influences perceptions regarding when LLMs should refuse to answer." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper tries to ground in the formal framework offered by contextual integrity (CI) theory. Section 2 and 3.1 spend time discussing CI and mapping it to LLMs. But its not obvious why CI is helpful here or the extent to which CI is even used in the construction of the benchmark? As best I can tell, CI is explained in the prompt used to generate conversational contexts. But it isn’t clear whether that’s necessary, or whether GPT-4o is even using that information. And the paper also mentions that all contexts were revised by the authors themselves.\n- The primary contribution (from a data perspective) would seem to be the different contexts the authors define–especially as the queries themselves come from existing work. The paper doesn’t really discuss what these contexts look like, how they were crafted, or how often they repeat across the different queries?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. How reproducible is the automatic content generation with an open source model? Have you tried this?\n2. Manual revision process appears to be expensive. Were there attempt to automate this? If so, with what level of success?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- extensive benchmark collection, with 900 distinct contexts with queries from 45 different categories.\n- the inheritance of contextual integrity theory is interesting and well motivated\n- over 47K human annotations from over 2K annotators, resulting in a high quality dataset" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper puts forth CASE-Bench, a Context-Aware Safety Evaluation Benchmark which integrates context into subsequent safety assessments of LLMs. The benchmark boasts 900 distinct contexts with queries from 45 different categories." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- the data was created with gpt-4o, bringing into question the potential for commercial use\n- explanation of limitations is lacking, would be good to delve into potential areas for improvement or 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": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See above." }, "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 identifies a key issue in current LLM safety evaluations: too much focus on rejecting single queries without considering context, which can lead to LLMs over-rejecting legitimate requests.\n2. The research methodology is solid, using Contextual Integrity theory to formalize context, power analysis to determine annotator numbers, and various statistical methods to verify result significance.\n3. The dataset creation process is thorough, with 900 query-context pairs, 21 annotators per task, and robust quality control measures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces CASE-Bench, a context-aware benchmark for evaluating LLM safety. The main innovation is incorporating context into safety assessments using Contextual Integrity theory. The study validates the significant impact of context on safety judgments through large-scale human annotation (2000+ annotators) and evaluates several open-source and commercial LLMs on this benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I wonder how the authors address this potential issue: many current jailbreaking techniques [1,2] actually work by creating \"safe-looking\" contexts to trick LLMs into generating dangerous content. Could context-based safety evaluation create new security risks?\n2. Regarding metrics, the paper mainly uses basic measures like accuracy and recall. Have the authors considered adding more detailed analysis, like measuring the impact of different types of errors?\n3. From a practical standpoint, I'm curious about the authors' thoughts on how to integrate this context-based evaluation approach into real-world LLM safety mechanisms?\n\n[1] Scalable and Transferable Black-Box Jailbreaks for Language Models via Persona Modulation\n[2] Quack: Automatic Jailbreaking Large Language Models via Role-playing" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. In which use-cases you find the differentiation by context important/useful?\n2. How can malicious actors that fake positive intent could be detected?\n3. In the results in Fig. 3, if I understand correctly, some unsafe contexts lead some categories to rated by humans as safe. That's unexpected and counterintuitive (categories between 10-20 on the x axis). Do you have an explanation for this? Can you share some examples from this category (I couldn't find any)?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "A thorough analysis of the influence of context in deciding whether a prompt/prompt response is problematic\nA dataset that may be useful for other research" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the problem of understanding whether a prompt/prompt reply is problematic or not given a particular context. It proposes that not all responses are problematic and some responses are problematic only if they are in a particular context. As such, it creates a dataset on 45 different categories, all with positive and problematic context. The dataset is used to evaluate state of the art models and also alignment/agreement with human annotators." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I'm personally not convinced that the context should dictate the perceived harm in a prompt/prompt response. A malicious actor could always fake a positive intent/context to get the information they want and create harm. As such, I find the premise/motivation of the work weak." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A context-aware safety evaluation benchmark for safety evaluation that considers context." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024casebench,\ntitle={{CASE}-Bench: Context-Aware Safety Evaluation Benchmark for Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=y9tQNJ2n1y},\nnote={under review}\n}" }, "abstract": { "value": "Aligning large language models (LLMs) with human values is essential for their safe deployment and widespread adoption. Current LLM safety benchmarks often focus solely on the refusal of individual problematic queries, which overlooks the importance of the context where the query occurs and may cause undesired refusal of queries under safe contexts that diminish user experience. Addressing this gap, we introduce CASE-Bench, a Context-Aware Safety Evaluation Benchmark that integrates context into safety assessments of LLMs. CASE-Bench assigns distinct, formally described contexts to categorized queries based on Contextual Integrity theory. Additionally, in contrast to previous studies which mainly rely on majority voting from just a few annotators, we recruited a sufficient number of annotators necessary to ensure the detection of statistically significant differences among the experimental conditions based on power analysis. Our extensive analysis using CASE-Bench on various open-source and commercial LLMs reveals a substantial and significant influence of context on human judgments ($p<$0.0001 from a z-test), underscoring the necessity of context in safety evaluations. We also identify notable mismatches between human judgments and LLM responses, particularly in commercial models within safe contexts. Code and data used in the paper are available at https://anonymous.4open.science/r/CASEBench-D5DB." }, "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": [ "safety", "benchmark", "context", "large language model", "contextual integrity" ] }, "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/3121a47f2b3b9fe2b5f7a2502c7a347404ea5730.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": "CASE-Bench: Context-Aware Safety Evaluation Benchmark 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": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": { "value": "The paper is desk rejected for significant textual overlap with [1]. This decision was confirmed by multiple members of the program committee. Table 1 in both papers is identical. The related work sections also have significant overlap. The amount of textual overlap without proper attribution makes this a case of plagiarism. \n\n[1] Towards Large Language Models as Copilots for Theorem Proving in Lean by Peiyang Song et al. https://arxiv.org/pdf/2404.12534" }, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "Submission Desk Rejected by Program Chairs" }, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": { "value": "agent-based system for math reasoning" }, "_bibtex": { "value": "@misc{\nzhang2024collaborative,\ntitle={Collaborative Theorem Proving with Large Language Models: Enhancing Formal Proofs with ProofRefiner},\nauthor={Haoyi Zhang and Andrew Liu and Zixuan Wang and Feiyang Wang},\nyear={2024},\nurl={https://openreview.net/forum?id=y9xNQZjUJM}\n}" }, "abstract": { "value": "Abstract： Theorem proving presents a significant challenge for large language models (LLMs) because formal proofs can be rigorously verified by proof assistants like Lean, leaving no room for errors. Existing LLM-based provers typically operate autonomously, but they often struggle with complex and novel theorems where human insights are crucial. We propose a new framework that positions LLMs as collaborative assistants in theorem proving to address this. This framework enables the seamless integration of LLM inference into the Lean environment, allowing developers to build various proof automation tools. These tools offer features such as suggesting proof steps, completing intermediate goals, and selecting relevant premises, thereby enhancing the theorem-proving process. Users can leverage our pretrained models or integrate their own, supporting local and cloud-based execution. Experimental results demonstrate that our approach is more effective in aiding humans and automating the theorem-proving process than existing rule-based systems. Additionally, we introduce a system called ProofRefiner, which refines questions and answers through dynamic dialogue adjustments to ensure relevance and precision." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Haoyi_Zhang1", "~Andrew_Liu8", "~Zixuan_Wang17", "~Feiyang_Wang4" ] }, "authors": { "value": [ "Haoyi Zhang", "Andrew Liu", "Zixuan Wang", "Feiyang Wang" ] }, "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": [ "Agent-base System", "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": { "value": "zhang|collaborative_theorem_proving_with_large_language_models_enhancing_formal_proofs_with_proofrefiner" }, "pdf": { "value": "/pdf/39e133160ee46970609314d991c12a9d9f626400.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Collaborative Theorem Proving with Large Language Models: Enhancing Formal Proofs with ProofRefiner" }, "venue": { "value": "ICLR 2025 Conference Desk Rejected Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Desk_Rejected_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": "- Feel free to discuss any of the main weaknesses I listed above.\n\n- line 204: \"In the end, we seek to obtain strict logical rules for use in a rule list.\" If I understand correctly, once the temperature is low enough, the learned predicates will already be \"almost\" discrete. At this stage, couldn't you just replace them with their hard counterparts, and obtain a hard rule list classifier? The lower the temperature, the closer the original soft RL classifier and the corresponding hard RL classifier will behave in terms of predictions and loss, meaning there's little to be gained, performance-wise, from using the soft version, and much to be gained (for the few inputs that fall in the affected region of input space), interpretability-wise, by using the hard version. So why not do that? Have you evaluated empirically what happens if you follow this route?\n\n- Appendix C: what split did you use for evaluating performance during grid search?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "**Originality**: NyRules combines existing ideas and improves on them. The level of novelty is acceptable.\n\n**Quality**: All technical contributions appear to be good. The experiments are carried out on several \"real-world\" data sets as well as on synthetic data, and the results are 5-fold cross-validated (which is refreshing; however: why 5 and not 10+?). The choice of competitors is also sensible. Performance improvement on \"real-world\" data is rather robust, although not dramatic in most cases (see also Weaknesses).\n\n**Clarity**: The text is generally well written, and figures/plots are altogether clean and useful. However, parts of the manuscript feel a bit rushed: I was puzzled by a few equations (which are not always entirely formal) and statements, see below.\n\n**Significance**: NyRules contributes to research on differentiable rule learning, well motivated by explainability requirements. The contribution is neither niche nor exceedingly significant." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Neuro-Symbolic Rule Lists (NyRules), an upgrade to standard rule\nlists -- a well known class of white-box classifiers -- that in addition to learning\nrules also learn how to discretize continuous input features automatically. (This\nis where the \"neuro-symbolic\" moniker comes form.) This feat is achieved by\nintroducing a continuous relaxation of rule list learning, which is combinatorial\nin nature. NyRules is similar in spirit to existing differentiable rule learning\napproaches, except it focuses specifically on rule lists and reuses/improves upon some\nkey techniques: 1) differentiable thresholding functions from Yang et al.,\nused as-is; 2) the differentiable conjunction-of-a-subset operation from Xu et al., improved\nto avoid vanishing gradients; 3) linear relaxation + annealing to enable end-to-end learning of\nthe rule list. NyRules are evaluated on several data sets against several\ncompetitors." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Quality**: \n\n- There is a pretty big issue with Table 1, which I wish sorted out: in many situations, XGBoost performs *much* better than any of the competitors, yet it is never marked in bold in Table 1. This is the case also for the rings dataset. This choice is not explained in the text (I may have missed it!) and risks conveying a twisted perception of performance improvement.\n\n I understand that boosted trees can be less interpretable than rule lists, and that -- since the paper does not attempt to evaluate interpretability of learned models at all -- this cannot be demonstrated. But as a result, the results hyperfocus on prediction accuracy, where NyRules does *not* have an edge over XGBoost (to the best of my understanding).\n\n **Suggestion**: the authors should explicitly explain in the paper why XGBoost results are not bolded, even when they outperform other methods, and clarify that the focus is on comparing interpretable rule-based methods, with XGBoost included as a non-interpretable benchmark. Adding this explanation would provide important context and avoid potential misinterpretation of the results.\n\n- A key issue is that there is not specific study of the impact of the threshold learning step, which is a big selling point of the proposed method. I am left wondering whether it is indeed important for performance and interpretability.\n\n **Suggestion**: the authors could conduct an ablation specifically focused on the threshold learning component by, e.g., comparing NyRules with a version that uses pre-discretized features, to directly demonstrate the impact of learned thresholds on performance and potentially on interpretability.\n\n- The authors equate rule length with interpretability, but this is not exactly exact.\n\n **Suggestion**: the authors should acknowledge this limitation in the Limitations section discussing that while rule length is used as a proxy for interpretability, a more comprehensive evaluation of interpretability, ideally including user studies, would provide a more accurate assessment.\n\n**Clarity**:\n\n- NeSy is **neuro**-symbolic because the perceptual component -- responsible for encoding low-level inputs such as images into high-level symbolic variables -- is implemented using neural networks. In this paper the perception component is minimal: it corresponds to the thresholding step. I honestly feel calling the proposed method \"neuro-symbolic\" is a stretch, and could be interpreted as disrespectful of what people in NeSy AI aim to do. I am not going to penalize the paper based on this, but I wanted to make it clear that I did not appreciate this choice.\n\n- Some equations are repeated unnecessarily (the definition of r(.) and eq. 2).\n\n- p2: Last equation: $\\alpha \\le x_i \\le \\beta$ is not a function, it is a condition; please add an indicator function. This also applies to line 184.\n\n- Eq 1: rl(.) is undefined.\n\n- Eq 1: the \"s.t.\" does not make much sense, because this is neither a satisfiability nor an optimization problem. It should be replaced with an if-and-only-if. In short, I'd write:\n$$\n rl(x) = c_j \\qquad \\Leftrightarrow \\qquad \\exists j s.t. (a_j(x) = 1) \\land \\forall i . p_i > p_j . (a_i(x) = 0).\n$$\nSince priorities are unique, the condition $i \\ne j$ is unnecessary. Actually, in its current form the equation seems incorrect, as it requires lower-priority conditions not to fire (i \\ne j, p_i < p_j), which to the best of my understanding is not needed. Or did I get this wrong?\n\n- line 188: mangled sentence.\n\n- line 262: what does the argmax_l act on? The argument is c.\n\n- line 278: same.\n\n- I cannot follow the comment at line 280: what is the point of talking about boosted rule sets at this point? This is the only place in the paper where they are mentioned.\n\nI am willing to increase my score provided the authors address the main issues I raised." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Answers / changes following the major comments listed above will address most of the limitations of this current version of the paper.\n\nAdditional questions:\n- The relevance of the predicate layer is highlighted for the Ring dataset that has exclusively continuous features. Did you observe a drop in performance compared to other methods for datasets with no continuous features ? \n- Have you tried other differentiable logical conjunction layers ?" }, "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 clearly written and of great quality.\nThe solution proposed to solve the vanishing gradients for the chosen computation of the conjunction is well formulated and the efficiency of the relaxation proven with the ablation study.\nSupplementary material (with code) is provided for reproducibility." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper paper proposes a novel end-to-end differentiable model to learn rule lists called NyRules and addresses limitations from existing work.\nThe method takes into account the discretisation of continuous features into predicates, the learning of conjunctions & rules as well as the ordering of the rules into rule lists.\nNyRules performances were validated on both synthetic and real world datasets and compared to existing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major comments:\n- Discretization on the fly has already be done in existing works (for example: _Kusters, Remy, Yusik Kim, Marine Collery, Christian de Sainte Marie, and Shubham Gupta. \"Differentiable Rule Induction with Learned Relational Features.\" In International Workshop on Neural-Symbolic Learning and Reasoning. 2022_). However, the thresholding layer presented in this paper is still very relevant and different from _Kusters et al_ approach for instance.\nThe paper's contributions need to be put into perspective. Also a comparison between the two different thresholding layers would strengthen the quality of the paper.\n- The improvements provided by the rule ordering are not proven. An additional ablation study would be required in order to differentiate the impact of the predicate layer and the active priority. Especially for comparison with RLNet which is also a neural-based approach to learn rule lists.\n- The final else (with no predicates) in the rule list is not described nor explained.\n- The fact that $l=2$ (binary classification) in the entire paper should be mentioned from start explicitly. It is indirectly mentioned only in 3.4 where the BCE loss is specified, in 5.1 where only datasets for binary classification are used and in future works (6.2). The reader is told this method is applicable for multi-classification with parameter $l$.\n- The interpretation of $c_j$ in the rules is not explicit. \"if $a_j$ then $c_j$\" is then converted in Figure 1 into \"if [...] then P(Disease) = 94%\" for example. Where is that value coming from ? softmax of $c_j$ ? I believe it needs to be explicit.\n\nOther comments:\n- Related work is lacking contributions from different fields. For instance, fuzzy logics with studies like _van Krieken, Emile, Erman Acar, and Frank van Harmelen. \"Analyzing differentiable fuzzy logic operators.\" Artificial Intelligence 302 (2022): 103602._ which analysed different differentiable logical conjunctions.\n- The description of the temperature annealing schedule seems to be missing.\n- In 3.3 the number of rules $k$ and the number of classes $l$ _are_ (typo?) fixed beforehand. \n- Figure 3.c, X1=1 and X2=1 = 0 ?\n- References to datasets sources are missing\n- Pre-processing of the datasets is not described. How are the datasets continuous feature discretised for other methods ?\n- The distribution of the features types in the datasets could be provided to highlight the relevance of the predicate layer of NyRules." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Is there a reference for GREEDY? What implementations of the algorithms did you use?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- nice formal presentation of the work" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a rule learning system for numerical data, that integrates discretization and learning. This is, by itself, not a novel contribution. Classic rule learning algorithms such as Ripper can do that as well. It also optimizes the rule order of the resulting list, which is a somewhat debatable contribution, because rules in a decision list are not independent of the order. This is also a problem for interpretability, BTW. Results are presented in F1-scores, which is somewhat unconventional for work in this area, and makes a comparison harder. I also miss a comparison to the (still) standard benchmark Ripper, in the Java implementation available, e.g., in Weka (this is important, the Python implementation Wittgenstein is considerably worse, and unfortunately many of the published claims of superiority over Ripper are based on that implementation). As a result, I find it hard to assess the value of this work. I do think that the paper (although generally well written) could be much stronger and more credible with respect to the experimental evaluation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- experimental results are somewhat non-standard, in particular the use of F1. Why do you think F1 is a good measure for evaluating classification problems? It is quite adequate for information retrieval, because there is an asymmetry between the positive and negative class, so that it is o.k. that unretrieved irrelevant documents do not influence the result, but I don't see why ignoring true negatives should be a good idea in any of the problems studied here.\n\n- The results also seem to be flawed at times. For example, results of CORELS at 0.33 or 0.27 when other algorithms are above 0.9 (similar for SBRL) indicate for me that the algorithms have not been adequately used. Possible explanations could be that they did not learn any rules, or learned for the wrong class, or similar. In my view, no algorithm can be as bad as 0.16 for Titanic. I guess you could invert its predictions and would get a better result (this would be clearer if accuracy had been used as an evaluation metric).\n\n- The authors seem to assume that decision lists are just an ordering of rules. Of course, an ordered rule set is a decision list, but the opposite is not true. This affects optimality and interpretability. Consider, for example the decision list\nIF A then +\nELSIF B then -\nELSIF C then +\nELSE -\nNote that each rule is only valid in context. It is, e.g., not true that all C's are positive, only if they are not B.\nIf you simply reorder the rules in the decision list (e.g., by swapping the two positive rules) this may drastically change the semantics. This is why classical rule learning algorithms like Ripper always learn a decision list in Sequence (Ripper does optimize it later on). \n\nOn the other hand, you could also rewrite the decision list as a rule set as follows:\nIF A then +\nIF B and NOT A then -\nIF C and NOT B and NOT A then +\nIn that case, you could reorder the rules as you wish. If rules like this are learned, they would be longer (which would be consistent with what the authors observed on their algorithm), but, on the other hand, the ordering is practically irrelevant.\n\nSo probably, the authors learn a mixture of both (optimizing both the order and the rule bodies at the same time), but this should be discussed deeper.\n\n- The ordering also has an effect on interpretability, of course. A long decision list cannot be easily interpreted, because the rules cannot be interpreted independently, you always have to interpret them in context of all previous rules. So the claim already posed in the introduction that decision lists are interpretable is debatable, in particular for long lists. It does hold if rules can be interpreted independently, but in that case the order would be irrelevant (see above). The authors should more clearly discuss the trade-off between ordered rules and interpretability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Q1. Could you show the experimental results on the average running times of the proposed method to the baselines? \n- Q2 (optional). From the objective function in Section 3.4, the proposed method does not explicitly impose constraints on the sparsity of predicates in each rule. However, from Figure 5(b), the experimental results suggest that the proposed method tends to obtain sparse rules, which was surprising to me. Did you carry out any post-processing for it? In addition, I guess that the required length of each rule to maintain sufficient accuracy may vary depending on the dataset. Does the trend in Figure 5(b) differ depending on the dataset?\n- Q3 (optional). In the learning problem of the proposed method, the constraints on minimum and maximum supports are relaxed as soft constraints (i.e., regularizer). In the experiments, how often did the learned rule lists violate the constraints on minimum and maximum supports? Also, while the learned rule lists may not violate this constraint if we set $\\lambda$ to be a sufficiently large value, how does it affect the accuracy or computational cost of the proposed method?\n- Q4 (optional). If I understand correctly, the proposed learning algorithm gradually decreases the annealing temperatures $t_{\\pi}$ and $t_{rl}$. Could you show the pseudo-code of a procedure for determining the values of $t_{\\pi}$ and $t_{rl}$ in each step during training?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Overall, I like this paper, and I think the authors make solid contributions to the community on interpretable machine learning. \n- S1. This paper is well-written, well-organized, and easy to follow. \n- S2. The authors introduce a differentiable representation of rule lists, which is a novel approach to the best of my knowledge. While the proposed representation mainly consists of existing methods, including [Yang et al. 2018], [Xu et al. 2024], and [Jang et al. 2017], it also includes some unique techniques, such as relaxed formulation for alleviating the vanishing gradients issues. In addition, I think the illustrative demonstration of each continuous relaxation (Figures 3 and 4) helps readers understand the effectiveness of the proposed approach well. \n- S3. The experimental results demonstrated that the proposed method often achieved better accuracy than the existing methods. The proposed method also attained comparable accuracy to XGBoost, which was surprising to me." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new framework for learning rule lists by gradient descent, named NyRules. First, the authors introduce a differentiable representation of a rule list using some existing techniques, such as the Gumbel-Softmax. Then, they formulate the task of learning a rule list that minimizes the empirical risk and the constraints on the support as a continuous optimization problem. One of the advantages of the proposed method is that it does not require pre-processing steps such as discretization and rule mining, which are required by most existing methods and often incur additional computational costs. Experimental results demonstrated that the proposed method achieved good accuracy compared to the baselines across all the datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I think this is a good paper, I believe the following concerns should be addressed to improve the quality of this paper: \n- W1. One of my major concerns is the computational cost of the proposed method. I could not find the experimental results on the running times of the proposed method and baselines, even in the appendix. To claim that the proposed method overcomes the scalability issue of the existing methods, I think the running time of the proposed method should be compared with the existing methods. Although I agree that it is a desirable advantage that the proposed method does not require discretization and rule mining as pre-processing, I am worried about the scalability of the proposed method because its average running time was not reported in the paper. \n- W2. This paper looks to be missing descriptions of some important information. For example, this paper seems to assume a binary classification task implicitly. I think it should be explicitly mentioned at, for example, the beginning of Section 2. In addition, as mentioned above, this paper seems to lack the experimental results on the computational times of the proposed method and baselines. I think such information should be reported because scalability is one of the important factors for practitioners when deciding which algorithm they should use for their task." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce NyRules, an end-to-end trainable model that unifies discretization, rule learning, and rule order into a single differentiable framework." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024neurosymbolic,\ntitle={Neuro-Symbolic Rule Lists},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yAN2oPHs7y},\nnote={under review}\n}" }, "abstract": { "value": "Machine learning models deployed in sensitive areas such as healthcare must be interpretable to ensure accountability and fairness.\nRule lists (_**If**_ $\\texttt{Age} < 35 \\wedge \\texttt{Priors} > 0$ _**then**_ $\\texttt{Recidivism} = $True, _**else if**_ \"Next Condition\" ...)\noffer full transparency, making them well-suited for high-stakes decisions.\nHowever, learning such rule lists presents significant challenges. Existing methods based on combinatorial optimization require feature pre-discretization and impose restrictions on rule size. Neuro-symbolic methods use more scalable continuous optimization yet place similar pre-discretization constraints and suffer from unstable optimization. To address the existing limitations, we introduce NyRules, an end-to-end trainable model that unifies discretization, rule learning, and rule order into a single differentiable framework. \nWe formulate a continuous relaxation of the rule list learning problem that converges to a strict rule list through temperature annealing.\nNyRules learns both the discretizations of individual features, as well as their combination into conjunctive rules without any pre-processing or restrictions.\nExtensive experiments demonstrate that NyRules consistently outperforms both combinatorial and neuro-symbolic methods,\neffectively learning simple and complex rules, as well as their order, across a wide range of 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": [ "Neuro-Symbolic;Rule Induction; Intepretability" ] }, "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/142754718917cfd6b93513ef46d12dc79268e987.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/f889ceb6510d333a99b87ca86b3e449c309628f7.zip" }, "title": { "value": "Neuro-Symbolic Rule Lists" }, "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": "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": 2 }, "primary_area": null, "questions": { "value": "Please see the discussed weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Strengths:\n1. As said by the authors, it is sensible that adaptivity is important (in NNs and also in GNNs).\n\n2. The adaptivity mechanism itself (not the method as a whole) seems simple (which is good)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an adaptive message-passing scheme that allows the updating of nodes in an independent way with respect to the number of message-passing steps (layers). The authors propose several ways to train the network using this mechanism, and several test time usages with different encodings. The paper concludes with several experiments to demonstrate the performance of ADMP-GNN." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses:\n\n1. The paper lacks discussions on highly relevant papers that suggest adaptivity in GNNs, from the message-passing (most important for this paper), and other mechanisms like the activation function and the normalization layer. Please see references R1-R5 below. \n\n2. Regarding W1, the papers R1 and R2 seem to do the same basic idea as here (even if the implementation is a bit different), so it does not seem novel, and it is also not discussed in this paper.\n\n3. Adaptivity in message passing can also be obtained with graph transformers/graph attention layers because they can adaptively choose which nodes should be discussed with which, but the authors did not discuss them in the paper or within the experiments. Similarly, the method should be compared with rewiring methods.\n\n4. It is not clear what the main benefit of this method is beyond adaptivity (which is important). For example, there is no discussion on which inherent problems in GNNs it can solve, like over-smoothing or over-squashing, and there is also no discussion on the expressiveness of this method.\n\n5. Regarding the design of the adaptive mechanism (section 3.2), it is interesting to know what the network actually learns to adapt to and how. However, this was not shown in the paper, so it lacks insights.\n\n6. The time complexity of the method is quite high. Although it is discussed by the authors, they do not properly compare the required runtimes with other methods, and they should also compare their performance with other methods that require more time, such as graph transformers and subgraph methods.\n\n7. I am a bit concerned about section 3.5; it shows heuristics to use the method on test nodes, however these proposed heuristics and encodings were also used in other methods and were shown to be quite strong, for example, the 'Walk Count' was used in the well-known RWSE encoding (see R6 for reference), and I think that it is hard to disentangle the results obtained by these approaches from the actual method proposed here, which is the adaptive message passing mechanism.\n\n**Regarding experiments, I have several concerns:**\n\n8. First, it is only conducted on up to 5 layers. However, most GNNs perform quite well in this regime (e.g., see R6). It is more interesting to know how this method performs when there are many layers, like 64,128,256, etc.\n\n9. It is not clear which splits were used. While the authors do say they use 10 seeds, the splitting of the data is not clear and does not seem to be in line with known splits of the datasets used here (e.g., as in R7). \n\n10. The results are poorly compared with relevant methods, as previously discussed, and this makes it hard to quantify the contribution of the proposed method properly. Also, the uncertainty on which splits were used in the experiments makes it harder to compare with known results from previous papers.\n\n11. The datasets used here were shown to suffer from multiple problems (see R8, R9), and therefore, showing results on additional and more up-to-date datasets is required. I would strongly suggest trying to benchmark your method on those shown in R9 and additional tasks like graph classification or regression. \n\n12. The ablation study is discussed in the main paper, but the results are missing.\n\n[R1] Adaptive Message Passing: A General Framework to Mitigate Oversmoothing, Oversquashing, and Underreaching\n\n[R2] Cooperative Graph Neural Networks\n\n[R3] Improving Expressivity of GNNs with Subgraph-specific Factor Embedded Normalization\n\n[R4] GRANOLA: Adaptive Normalization for Graph Neural Networks\n\n[R5] DiGRAF: Diffeomorphic Graph-Adaptive Activation Function\n\n[R6] A Survey on Oversmoothing in Graph Neural Networks\n\n[R7] Geom-GCN: Geometric Graph Convolutional Networks\n\n[R8] Pitfalls of Graph Neural Network Evaluation\n\n[R9] A critical look at the evaluation of GNNs under heterophily: Are we really making progress?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 well does the centrality-based layer selection policy algin with the one given in the oracle accuracy?" }, "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 the adaptive depth aggregation is well-motivated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores using node-wise adaptive depth aggregation for GNN. To this end, a progressive training strategy is proposed, and the proposed method is evaluated on several benchmark datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It is understandable that different nodes may benefit from different depths of aggregation, and the results in Table 2 show that the maximum test accuracy by using the aggregation results from different layers is much better than what is achieved by using the aggregation results from a single layer. However, the main challenge is determining the optimal depth for each node. The heuristic method proposed in this paper utilizes clustering and then uses the validation accuracy for each cluster to determine the optimal depth for the whole cluster that can easily not be optimal, e.g., maybe, in some clusters, the validation nodes are not too small and not representative enough. Indeed, the accuracy of the proposed method is not much better than the GCN baseline." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "N/A" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The idea is novel to me. However, since I haven’t worked on GNNs since 2022, I’m unable to provide an accurate assessment of its novelty.\n\n2. The paper proposes a sequential training approach, which helps reduce computational complexity.\n\n3. To improve generalization on test nodes, the paper introduces several heuristic metrics—k-core, PageRank, and Walk Count—to determine the optimal number of aggregation layers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The key idea of this work is to use varying numbers of message-passing steps to aggregate information uniquely for each node, allowing different nodes to utilize different aggregation and propagation layers. After a certain number of layers, the model first calculates the loss for these nodes, performs backpropagation, and then freezes their gradients to avoid conflicts in subsequent aggregation layers. The paper proposes an adaptive framework to determine the optimal number of layers for each node." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The improvement is somewhat limited on certain datasets, such as Cora and Citeseer.\n\n2. The paper is under 9 pages, and, in my view, it feels somewhat incomplete." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- When executing the experiments for figure 1, are the L different GNNs trained with the same dynamics? (batch size, epochs, learning rate ...) Because deeper models might be harder to train. See also [5], these authors questioned the oversmoothing issue, and they claim it's just trainability issue for deep networks. \n- In figure 1, I can understand the green curves have an upside-down V shape, too few layers lead to under-reaching, and too many layers lead to over-smoothing. But why are the blue curves V shape? It does not make sense to me, why it drops first and adding more layers is better again. \n- In line 193-194, is there a motivation why no hidden representations h is included in the exit function?\n- In line 236, the authors mentioned about the conflicts in gradients. Is there any proof or related work mentioning this?\n- I wonder if stochastic depth [6] training also helps, which can be a candidate to your sequential training trick. \n- Is this work directly applicable to graph property prediction, like, just by adding a pooling layer? If yes, why no experiments on graph prediction tasks? If not, why?\n- In line 313, 316, what is ADMP-GNN PT? \n- Some experiment results seem pretty bad, for example, in table 1, CS column, 2 layers, ADMP-GNN ALM is much worse than GCN(*). Why? Because of the gradient conflicts?\n\n[5] Peng, Jie, Runlin Lei, and Zhewei Wei. \"Beyond Over-smoothing: Uncovering the Trainability Challenges in Deep Graph Neural Networks.\" arXiv preprint arXiv:2408.03669 (2024).\n[6] Huang, Gao, et al. \"Deep networks with stochastic depth.\" Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14. Springer International Publishing, 2016." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Overall, the writing is pretty good and clear.\n- The motivation about adaptive layer for each node is convincing. \n- The synthetic dataset experiment is pretty well designed. \n- The architecture is efficient and has the same complexity as a normal GNN.\n- The empirical experiments are promising and show the model-agnostic property of the proposed method. The ablations are extensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed Adaptive Depth Message Passing GNN (ADMP-GNN), which learns a preset-max-depth GNN, and outputs predictions per layer. The authors also proposed training tricks and layer selection heuristics during test time. The network is efficient and shown to be effective on some real-world datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some related works are missing. The authors claim none work has been done for adaptive layer GNNs, which is not true. There are [1] [2] [3] [4], which may not have exactly the same method, but they also share the idea of adaptive message passing for nodes. \n- Similarly, these works should also be compared in the experiments. \n\n[1] Spinelli, Indro, Simone Scardapane, and Aurelio Uncini. \"Adaptive propagation graph convolutional network.\" IEEE Transactions on Neural Networks and Learning Systems 32.10 (2020): 4755-4760.\n[2] Finkelshtein, Ben, et al. \"Cooperative graph neural networks.\" arXiv preprint arXiv:2310.01267 (2023).\n[3] Faber, Lukas, and Roger Wattenhofer. \"GwAC: GNNs with Asynchronous Communication.\" Learning on Graphs Conference. PMLR, 2024.\n[4] Errica, Federico, et al. \"Adaptive Message Passing: A General Framework to Mitigate Oversmoothing, Oversquashing, and Underreaching.\" arXiv preprint arXiv:2312.16560 (2023)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "we propose Adaptive Depth Message Passing GNN (ADMP-GNN), a novel framework that dynamically adjusts the number of message-passing layers for each node, leading to enhanced performance." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024admpgnn,\ntitle={{ADMP}-{GNN}: Adaptive Depth Message Passing {GNN}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yAU5X77S06},\nnote={under review}\n}" }, "abstract": { "value": "Graph Neural Networks (GNNs) have proven to be highly effective in various graph representation learning tasks. A key characteristic is that GNNs apply a fixed number of message-passing steps to all nodes in the graph, regardless of the varying computational needs and characteristics of each node. Through empirical analysis of real-world data, we show that the optimal number of message-passing layers differs for nodes with different characteristics. This insight is further validated with experiments on synthetic datasets. To address this, we propose Adaptive Depth Message Passing GNN (ADMP-GNN), a novel framework that dynamically adjusts the number of message-passing layers for each node, leading to enhanced performance. This approach is applicable to any model that follows the message-passing scheme. We evaluate ADMP-GNN on the node classification task and observe performance improvements over a wide range of GNNs." }, "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": [ "Graph Neural Networks" ] }, "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/ba914398afee412638d7f3f260cef3b68fe74592.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/7a13519f8b8bc2912dac822542756d5082ea4b7c.zip" }, "title": { "value": "ADMP-GNN: Adaptive Depth Message Passing GNN" }, "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": 3 }, "primary_area": null, "questions": { "value": "- The model uses frozen vision encoders. Does it mean existing pretrained visual representation is sufficient for robot manipulation? I wonder if the authors spot any cases where pretrained visual encoder is insufficient and lead to unsatisfactory performance.\n- In pretraining on heterogeneous data with varied control frequency, the model takes control frequency as conditioning. I wonder how this strategy works in practice - does the model learns policy prior corresponding to different control frequency?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper presents a complete and remarkable research work that pushes forward the boundary of large-scale robot learning. \n- The model is developed on top of the diffusion transformer with a unified action space, which allows large-scale pretraining on heterogeneous robot data to boost the performance\n- The authors collect the largest robot dataset for bimanual manipulation with comprehensive task coverage for fine-tuning the model\n- The experiments show that the advantage of the model from a foundation model aspect: generalization, few-shot learning, and scaling behavior" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper develops a 1.2B-parameter robotics foundation model that is trained and evaluated on real robot data for bimanual manipulation. The model is trained with imitation learning: (i) pretrained on 1M trajectories combining available datatsets collected for different robots, and (ii) fine-tuned on a self-collected dataset with 6k demonstrations for a Mobile Aloha robot. The model adopts a diffusion transformer (DiT) architecture that takes multi-modal inputs (images, language, etc.) and generate action chunks with multimodal distribution. \n\nThe model is evaluated on 7 real robot tasks against mainstream baselines. The comparison shows that the model can: (i) generalize zero-shot to novel objects, scenes and language, (ii) learn new skills with few data, (iii) accomplish dexterous tasks. Ablation studies show that larger model and pretraining with large data significantly boost the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the paper demonstrates that the foundation model is allows zero-shot and few-shot generalization, and can achieve dexterous manipulation, each of these characteristics is only validated on ~one task and may be insufficient. Evaluations on more tasks and existing benchmark tasks will complete the results.\n- It seems that the baselines are not trained on the complete fine-tuning dataset. This doesn't form an apple-to-apple comparison.\n- The writing of the paper has room for improvement. Some of the sentences are too long, which prevent reading of the paper smoothly.\n- Mistakes in citing papers. For example, it seems that \"Xie et al. 2020\" is wrongly used as the reference of DiT in line 83 and 319." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "My questions are mainly also discussed in the weakness. Here is an summary:\n1. Given that the diffusion transformer already contains multiple nonlinear layers, why does adding an additional layer improve performance? Have you conducted ablation studies to support this choice?\n2. In the original UNet diffusion policy, the last layer is a Conv1dBlock. Have you compared the performance of your MLP block with this alternative?\n3. What is the the MLP block performance gains on other tasks?\n4. Can you provide more insight into how the model design specifically addresses high frequency change?\n5. What are the performance gains from using all data through this unified action space compared to previous methods that only use robots with similar action spaces or retain subsets of inputs with a common structure?\n6. How does the proposed padding strategy improve performance compared to zero padding? Could you provide comparative results?\n7. How do these normalization layers integrate into the transformer architecture, and how do they address instability? Have you tested different configurations of these layers?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper demonstrates strong performance in scaling up robotics models. It presents several interesting components that improve training stability and performance.\n\nThe unified action space, and especially the padding technique, is interesting.\n\nThe paper shows capabilities on several challenging real-world bimanual manipulation tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an effort toward building a foundation model for bimanual manipulation. It proposes techniques to unify the action space, enabling training on a very large robot dataset. The authors also scale up the model to 1.2B parameters, making it the largest diffusion transformer model for robotics. In this process, the authors identify several key elements to improve training stability and performance. The resulting model achieves good zero-shot generalization performance on unseen and complex tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Several claims are not very precise and not very clear. For example, the authors mention the nonlinearity and high frequency of robotic data. While it is true that the data is nonlinear, how does the proposed method tackle this challenge? The authors argue that changing the last linear layer to an MLP block solves this problem and brings significant performance improvements. While the performance is impressive, I think this requires more careful ablation experiments. Firstly, the entire diffusion transformer is already highly nonlinear due to stacking multiple layers, so why does adding more layers help in this case? Secondly, in the original UNet diffusion policy, the last layer is a Conv1dBlock; have the authors compared with that? Lastly, why is it only evaluated on the dexterity task?\n\nIt is also unclear how the model design choices address the “high frequency of robotic data.” Given that these two claims are highlighted in the abstract as the main challenges, I believe they require more careful analysis and discussion.\n\nAs the unified action representation is a major contribution of this paper, there should also be more analysis of this aspect. For example, what are the performance gains from using all the data because of the unified action space, compared to previous methods that use “robots with similar action spaces (Yang et al., 2023; Ghosh et al., 2023; Kim et al., 2024) or retain only a subset of inputs sharing the same structure (Collaboration et al., 2023; Yang et al., 2024)”? Additionally, what is the performance gain of the proposed padding strategy compared with padding with all zeros?\n\nThe authors argue the necessity of using RMSNorm and QKNorm, but they only show the loss without them (Figure 4(a)), which provides very little information on how effective the proposed approach is and whether it addresses the instability issue. It also does not mention how to integrate these normalization layers within the transformer.\n\nScaling up the models and data is certainly attractive, and the paper shows impressive results. However, most of the analysis are “binary”, which means the results are either with or without. Showing more datapoints (model with different size, using different percentage of the dataset) will present more insights to the reader." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 4 }, "primary_area": null, "questions": { "value": "- Could the authors elaborate on why diffusion models were specifically chosen over other generative methods like VAEs or GANs for this task? While diffusion models show high expressiveness, a comparison or rationale would clarify their unique benefits in bimanual manipulation.\n\n- The Physically Interpretable Unified Action Space is innovative, but how does it handle robots with vastly different kinematics or action constraints?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The paper introduces a novel application of diffusion models to bimanual manipulation, addressing the high-dimensional, multi-modal action space through a Physically Interpretable Unified Action Space. This approach is a creative extension of diffusion models in robotics, particularly for dual-arm coordination, a challenging domain with limited prior work.\n\n- The model is rigorously tested, with comprehensive experiments demonstrating superior performance over existing baselines. The use of the largest multi-robot dataset and a specialized bimanual dataset for fine-tuning enhances the validity of results and supports the model's effectiveness.\n\n- This work contributes substantially to the field by advancing foundation models for robotic manipulation. RDT’s capabilities for zero-shot generalization, few-shot learning, and instruction following mark a significant step towards adaptable and scalable robotic models, with promising implications for real-world applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the Robotics Diffusion Transformer (RDT), a large-scale diffusion-based model for bimanual robotics. RDT introduces a Physically Interpretable Unified Action Space to standardize actions across robots, enhancing generalization, and uses diffusion models to handle complex, continuous action distributions. Pre-trained on the largest multi-robot dataset and fine-tuned on a custom bimanual dataset, RDT demonstrates strong zero-shot and few-shot generalization, outperforming existing methods in dexterous, real-world tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper introduces a Physically Interpretable Unified Action Space for handling data heterogeneity, but additional details on potential limitations or failure cases during training with highly diverse data would be beneficial. This could include examples where action standardization might lead to loss of unique features across robots.\n\n- Although the experiments show impressive results, expanding the evaluation to more varied and complex real-world tasks (beyond the 6,000-episode dataset) and more hardwares(beyond ALOHA) could further validate RDT's robustness.\n\n- The paper proposes several innovative multi-modal encodings (e.g., masking, cross-attention) but lacks ablation studies on these design choices. Showing how each component contributes to performance could clarify their impact on handling visual and language-conditioned tasks effectively.\n\n- There's a typo at L76, it should be \"data\" instead of \"date\" ." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. \"Does the RDT-1B fine-tuning use the entire self-collected dataset, and is it not fine-tuned separately for each task in the evaluation?\"\n2. \"Regarding the computation of success rate for the 'wash cup' task with 'seen cup1': the success rate (SR) for 'get water' is 50, for 'pour water' is 87.5, and for 'place back cup' is also 87.5, yet the overall SR is listed as 50. Since the 'get water' subtask has an SR of 50, and the following subtasks have SRs below 100, how is the total SR calculated as 50?\"" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "A foundational model for bimanual manipulation is absent in the current community, which is an important direction; aligning different robot embodiments is also a crucial question for pre-training on large-scale datasets. The proposed action representation is simple yet effective for pre-training on diverse robot datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a unified action representation to align different robots, facilitating pre-training on diverse robot datasets for bimanual manipulation. Additionally, it introduces a diffusion transformer-based architecture with several modifications for enhancing policy learning, and scaling up with large datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Compared to the robot datasets used for pre-training the baseline, such as OpenVLA, this paper appears to use a more diverse set of datasets, including additional bimanual manipulation datasets like ALOHA and Mobile ALOHA, contributing nearly 10% of the total datasets. Fine-tuning a baseline pre-trained on single-arm datasets for a bimanual manipulation setting may result in poor performance on bimanual tasks, making it difficult to demonstrate that using the diffusion transformer architecture is superior to using a large language model as the pre-training backbone." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024rdtb,\ntitle={{RDT}-1B: a Diffusion Foundation Model for Bimanual Manipulation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yAzN4tz7oI},\nnote={under review}\n}" }, "abstract": { "value": "Bimanual manipulation is essential in robotics, yet developing foundation models is extremely challenging due to the inherent complexity of coordinating two robot arms (leading to multi-modal action distributions) and the scarcity of training data. In this paper, we present the Robotics Diffusion Transformer (RDT), a pioneering diffusion foundation model for bimanual manipulation. It is built on scalable Diffusion Transformers (DiTs), which can effectively represent multi-modality, with innovative designs to deal with the heterogeneity of multi-modal inputs and to capture the nonlinearity and high frequency of robotic data. To address data scarcity, we first introduce a Physically Interpretable Unified Action Space, which can unify the action representations of various robots while preserving the physical meanings of original actions, facilitating learning transferrable physical knowledge. With the above designs, we managed to pre-train RDT on the largest collection of multi-robot datasets to date and scaled it up to $1.2$B parameters, which is the largest diffusion-based foundation model for robotic manipulation. We further fine-tuned RDT on a self-created multi-task bimanual dataset with over $6$K+ episodes to refine its manipulation capabilities. Experiments on real robots demonstrate that RDT significantly outperforms existing methods. It exhibits zero-shot generalization to unseen objects and scenes, understands and follows language instructions, learns new skills with just 1$\\sim$5 demonstrations, and effectively handles complex, dexterous tasks. Code and a Demo video are provided in the supplementary materials." }, "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": [ "robot learning", "diffusion models", "foundation models", "bimanual manipulation" ] }, "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/e78850b39075ba973908cf24578b26a9b4c85065.pdf" }, "presentation": null, "primary_area": { "value": "applications to robotics, autonomy, planning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/790d1f8285e80d0c411bcec3d39cc31ca25b7072.zip" }, "title": { "value": "RDT-1B: a Diffusion Foundation Model for Bimanual Manipulation" }, "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": 1 }, "primary_area": null, "questions": { "value": "1. what is the meaning of \"root-free\" scheme?\n2. Can you please elaborate on how you define the sequence of the curvature matrices in \"Iterate matching\" scheme to measure the performance of your method? \n3. Can authors reference which work shows superior generalization of the FNGD for p's either than 2, 4, 1?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors introduce a unifying framework for a class of preconditioned gradient agorithms as special cases of fractional natural gradient descent, and gives a general framework to implement it more efficiently in a spectral parameterization that is motivated by a Riemannian optimization algorithm in a reparameterization of the space to the normal coordinates." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduce a unifying framework way to calculate Fractional natural gradient descent estimates such as RMSprops or Shampoo by considering a diagonalization of the curvature matrix $S$ as $B diag(d) B^\\top$ for orthogonal matrix $B$, and updating $B$ and $d$ instead. This parameterization allows them to calculate the fractional power easily as $B \\diag(d^{-1/p})B^\\top$. To update the parameters $B$ and $d$, the challenge is that there are constraints on the space of the middle matrix that has to be diagonal, and $B$ has to be orthogonal. The authors then seems to use a reparameterization of the space $(B,d)$ in what they call a \"root-free\" approach, so that they can take the Riemannian gradient steps in that reparameterization with any constriants, in a specific manifold metric that is related to a concept called \"Gaussian variational distrbution.\" To do so, the they use previous ideas going back to work of Amari et al." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. It is nots clear what is their contribution and what has been don ein previous work.\n2. Description of the use of the gaussian variational approach in line 286 is very vague, it is not even clear why this approach is well-motivated or what is $\\ell$.\n3. Authors keep taking about suing local coordinates but it's not exactly what coordinate system for the manifold of matrices they are referring to, do you mean the normal coordinate system using the xponential map?\n\n4. The definition of function $Tril$ is unclear, and authors define what they mean by $Cayley$ in the middle of the paper.\n5. The high level ideas are vague and include terms that is not clear what they exactly refer to.\n6. It would be nice if authors can provide further motivation or example on which practical case the kronecker-structured spectral parameterization is useful. \n7. It is not clear what is the manifold metric the authors consider in their Riemannian appraoch. they connection of the gaussian variational framework, which seems to go back to work of Amari, to obtaining a manifold metric seems unclear. \n\nOther issues:\nH is not defined in equation 3.\nU and Tril are not defined in algorithm box on line 162\nLine 107 the notation is super unclear, is Mat(g) the diagonal matrix with entries of g? If so what is the difference between $G^\\top G$ and $G G^\\top$, and $S_C, S_K$?\nAuthors refer to “root-free” method but they don’t define what they mean\nEquation 10 is not clear at all\nAuthors don't provide any intuition of their approach/contribution e.g. for equation 10 or for the algorithm box in line 162\n\nMinor issue:\nClaim 1 is named Lemma 1 in 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": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "NA" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Consider a unified framework for adaptive methods through the fractional power of the Fisher matrix." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method for fractional natural gradient descent (FNGD) using learnable spectral factorizations. It introduces a way to factorize the Fisher matrix and apply fractional powers, aimed at enhancing neural network training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Clarity Issues: Notation and concepts are unclear, making the method difficult to follow. For example:\n- What is $B$ matrix?\n- The purpose of learning the Kronecker factorization is under-explained \n- What is matrix H? (In line 101)\n- Meaning of the “mat” notation (line 107)\n- What are matrices $S^K$ and $S^C$?\n- What is $\\otimes$ product?\n- Description of the “Gaussian problem”?\nFor that concept or notations, the authors should give an appropriate introduction and explanation and references\n\nWeak Motivation: \nThe paper provides an insufficient explanation as to why this complex approach is necessary or how it significantly advances existing methods. It does not provide new insight or guidelines (or at least does not explain clearly) that can potentially be utilized to result in efficient optimization or training.\n\nAmbiguity in Implementation: \nThe handling of factorization ambiguities in high-dimensional settings is unclear, particularly in terms of computational impact and scalability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 2 }, "primary_area": null, "questions": { "value": "Could you provide a more intuitive explanation of the advantages of using spectral parametrization over traditional matrix decomposition methods, especially regarding computational efficiency?\n\nIt would also be valuable to include more comparisons with recent advancements in adaptive gradient methods, particularly as applied to large-scale language models.\n\nAdditionally, a clearer summary of the implications of using arbitrary fractional powers in gradient descent in both the introduction and conclusion would enhance accessibility, helping readers better grasp the broader impact of this approach.\n\nWhy was the KFAC method excluded from the numerical comparison?\n\nAs an early suggestion, moving the Kronecker extension to the appendix could allow more space for a structured development of the core method in the main text, which may improve readability." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The introduction of learnable spectral factorization to FNGD is an innovation that addresses computational bottlenecks in second-order methods, such as those found in Shampoo or KFAC.\n\nThe authors effectively address numerical instability, making the proposed method suitable for low-precision arithmetic, which may improve the efficiency in large-scale machine learning tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel approach to fractional natural gradient descent (FNGD) by a learning method, which addresses the challenges of computing the matrix fractional powers. The authors present a Riemannian framework to learn eigen-factorized Fisher estimations on-the-fly, which allows the efficient application of arbitrary fractional powers without requiring matrix decompositions. Numerical experiments show the effectiveness in positive-definite matrices optimization problems and neural networks training tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "To strengthen the argument for using this method over existing alternatives, an ablation study examining the effects of different spectral factorization techniques on convergence speed and generalization performance would be beneficial.\n\nWhile the paper demonstrates scalability in terms of Kronecker-structured preconditioners, additional details on computational complexity and memory requirements—particularly in comparison to popular methods like Adam or Shampoo—would offer a clearer understanding of the associated trade-offs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1) Is the method implemented as a PyTorch optimizer?\n\n2) How many parameters are required for the method's performance?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The proposed method is focused on interesting and valuable problems to solve. The paper has a good and easy-to-follow introduction which motivates the proposed method. The method has a big potential from the practical perspective as it exploits curvature information." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In the paper, the author propose a novel optimization method for deep learning. It aims to learn spectral factorization of Fisher preconditioning without matrix decomposition. By knowing the spectral factorization it is possible to apply arbitrary power of the preconditioning. Finally, the authors provide evidence of practical effectiveness and performance of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The style of presentation is confusing in some parts of the paper. Next, I list some examples: a) Figure 1 shouldn’t be presented as is. It has two methods inside and half of the paper description. It ruins the style and flow of the paper. b) The authors present theoretical results as “claims”, however, next they prove it as lemmas. So, it would be reasonable to formulate them as lemmas in the main text as well. c) In lines 151-161 and 242-248, there is an extra spacing which shouldn’t be there. d) Some colorings are confusing. For example, line 353, 745, 833, 932. It might be forgotten highlights of the changes. \n\n2. The experimental section. As the paper doesn’t provide convergence proofs, we can say it is mostly an empirical paper backed by theoretical intuition. a) Hence, the extensive experimental results are expected. In the paper, only one figure presents the performance of the methods for 3 vision transformers on one dataset. It does not seem enough for an experimental paper. So, I would recommend adding experiments for common DL benchmarks to understand the methods' performance in comparison to state-of-the-art methods. Some examples can be found here https://mlcommons.org/benchmarks/algorithms/ b) Another issue is that the experiments’ description is lacking proper details and descriptions. It is limiting the reproducibility of the results. There are no methods parameters, learning rates, and so on. The attached code may help as well. \n\nSome minor misprints: a) Please, unify the use of “RMSprop”, as currently there are multiple variants of it (RMSprop, RmsProp, RMRprop) b) Line 709: misprint in “Given”" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "**Q1:** In Line 318, the authors manage a phrase: ''When changing coordinates, the metric representation has to be changed accordingly Lee, (2018) to make RGD invariant to coordinate transformation''. Please, could the authors clarify a steps like that with more formal presentation?\n\n**Q2:** Can the authors clarify a conceptual difference between the *Procedure for Full-matrix Spectral Parametrizations* and procedure in *Riemannian Approach for Obtaining Root-free Update Schemes*? As, naturally, the *Step 1* and *Step 2* of the latter approach describes the whole pipeline in general.\n\n**Q3:** Why does the *Claim 5* not considered as a definition?\n\n**Q4:** Could the authors clarify, please, how does it follows from Figure 4, that fraction roots are better than p=2?\n\n**Possible suggestions:**\n\n1) Rename ''default update scheme'' and ''our scheme'' with more sophisticated names.\n2) Explicitly write Lemmas 1-5.\n3) Formally define Fisher-Rao metric in the gentle introduction to the approach.\n\n\n\nJohn M Lee. Introduction to Riemannian manifolds, volume 2. Springer, 2018." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "**S1:** The authors study a new form of the preconditioneer – spectral factorization of the curvature approximation. They show, that for such a preconditioneers can b eapplied an update rule allowing to learn the spectral factorization at each step of the optimization problem.\n\n**S2:** The approach of the paper extends to the Kronecker factorization and analysis of the spectral decomposition for the Fisher-Rao metric." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an adaptive update scheme for full-matrix spectral factorization. This method serves as the efficient approximation of the preconditioning matrix, thus, allowing to perform a step of natural gradient descent faster." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**W1:** The writing of the paper is hard to follow.\n\n**W2:** The are no explicit formulation of lemmas in the paper.\n\n**W3:** The key points in the theoretical analysis is 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": 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": 3 }, "primary_area": null, "questions": { "value": "1. How did you generate matrix $Q$ in the positive-definite matrix optimization problem of Section 4?\n2. For the low-precision NN training in Section 4, you claimed that $p=1$ is better than $p=2$, but the difference is not significant from the plots presented. Do you have any other numerical result showing a more significant discrepancy?\n3. Did you have a complete derivation of the procedure for Kronecker-structured Spectral Parametrizations?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. This approach allows for the cheap application of arbitrary fractional powers, which distinguish it from other similar RGD approaches (e.g. Cholesky-based RGD).\n2. This approach circumvents the numerical instabilities of matrix decompositions, making it easy to implement and more practical.\n3. This paper is an extension based on previous literatures which viewed learning the curvature approximation (preconditioner) as learning the covariance of a Gaussian variational distribution by performing RGD on the manifold of dense or Kronecker-factorized positive-definite matrices. This paper further separates the diagonal part and orthogonal part by spectral decomposition, and incorporates the new constraints arising from the spectral decomposition for the Fisher-Rao metric in derivation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a stable and efficient Riemannian framework to learn eigen-factorized Fisher estimations, allowing for the cheap application of any fractional powers. The efficacy of this approach is demonstrated in numerical tasks including positive-definite matrix optimization and low-precision NN training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. No numerical performance in large-scale setting is presented." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "My questions mostly overlap with the confusion I have in the Weaknesses section above. It would be great if authors could address those concerns, in which case I would happily increase my score." }, "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 addresses a relevant and interesting problem in optimizing deep neural networks, which may inspire future research ideas. \n\n2. In addition to introducing the counterpart for the classical NGD methods, the paper also introduces the update rule that works for methods like Shampoo, which preconditions on both sides of the matrix iterate. \n\n3. The paper clearly describes its own contribution and credits the contribution of the previous works." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new update rule for the approximation of Fisher Information matrix in the natural gradient descent and its variants like RMSprop and Adam(W). The intuition is based on the observation that the preconditioner can be interpreted as the inverse covariance matrix of a Gaussian variational distribution (Lin et al. 2024). The authors propose to learn a decomposition of the preconditioner so that the fractional inversion can be computed efficiently. Numerical experiments show that the iterates generated by such a method have comparable efficiency in approximating the Fisher information and good performance when used to train neural networks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main motivation of the method, which is the flexibility of choosing p, is mentioned but not very well established. I think the authors should at least make an effort on the numerical side. They mention that \"This shows that the potential of using other fractional roots.\" by the end of the section 4, which seems weak as this is the main reason for potential practitioners to use this method over its simpler counterpart. Some results showing that the best training result can only be obtained by some p other than 1 or 2 can be very persuasive in this case. \n\n2. In terms of the computational cost, the comparison between this method and previous methods is not clear. The authors calibrated the running time for each method in section 4 by adjusting the number of iterations to update the preconditioner, but it is still vaguely defined and not quantified. One good comparison is to those methods which directly compute the matrix decomposition using high-precision arithmetic. I think this will establish the other important motivation of the method.\n\n3. The method suffers from a per-iteration error of $O(\\beta_2^2)$, which is non-negligible for constant $\\beta_2$. Also, the inverse approximation when calculating the Cayley map introduces another source of inaccuracy. \n\n4. There are some notation ambiguity in the paper. For example, Kronecker product and functions like Mat($\\cdot$) are not defined clearly. Diag($\\cdot$) and diag($\\cdot$) are used interchangeably. Eq 3 seems to come from nowhere when it first appears in the paper (no introduction before and after, no definition of $\\mathcal{H}$). Claim 2 eventually becomes Lemma 2 in the appendix." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024fast,\ntitle={Fast Fractional Natural Gradient Descent using Learnable Spectral Factorizations},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yBLBls6ryd},\nnote={under review}\n}" }, "abstract": { "value": "Many popular optimization methods can be united through fractional natural gradient descent (FNGD), which pre-conditions the gradient with a fractional power of the inverse Fisher:\n RMSprop and Adam(W) estimate a diagonal Fisher matrix and apply a square root before inversion; other methods like K-FAC and Shampoo employ matrix-valued Fisher estimates and apply the inverse and inverse square root, respectively.\n Recently, the question of how fractional power affects optimization has moved into focus, e.g. offering trade-offs between convergence and generalization.\n Gaining deeper insights into this phenomenon would require going beyond diagonal estimations and using cheap and flexible matrix-valued Fisher estimators capable of applying any fractional power; however, existing methods are limited by their expensive matrix fraction computation.\n To address this, we propose a Riemannian framework to learn eigen-factorized Fisher estimations on the fly, allowing for the cheap application of *arbitrary* fractional powers.\n Our approach does not require matrix decompositions and, therefore, is stable in half precision.\n We show our framework's efficacy on positive-definite matrix optimization problems and demonstrate its efficiency and flexibility for training neural nets." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "natural gradient", "Riemannian optimization", "positive-definite manifold", "Kronecker-facotrized", "Shampoo" ] }, "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/b56f0b045f9cfb07ed219cb432853ec91bf0af6d.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": "Fast Fractional Natural Gradient Descent using Learnable Spectral Factorizations" }, "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": "- Are all the experiments conducted with constant head Transformers? I couldn't find this detail in the Colab notebook since the checkpoints are directly imported for analysis.\n\n- According to [3], networks can learn either \"clock\" or \"pizza\" algorithm, depending on the details of training. How do the authors ensure that the learned algorithm is always \"pizza\"?\n\n- Are Figures 2,5 empirical results or just theoretical visualizations? \n\n- Lines 299-303: The omission of x/2 term seems rather hand-wavy. How does one see that computing the bounds after ignoring this term is a reasonable thing to do? It is unclear to me even after referring to Appendix K.\n\n- Does Table 1 imply that the contribution from $\\\\varepsilon_\\\\phi$ is very large? This seems unlikely since $\\\\psi_i - 2\\\\phi_i$ are quite small in Figure 3.\n\n- Line 349: Does \"brute force\" simply mean empirical?\n\n- Line 373-377: Zhong et al. [3] do not claim that logits are always of the \"pizza\" form. They claimed that either \"clock\" or \"pizza\" algorithm can be found, depending on the hyperparameters.\n\n- Lines 395-397: I am puzzled by the $R^2$ results. The paper sets out to explain the \"pizza\" algorithm. However, the logits seem to align better with the \"clock\" algorithm. \n\n- Lines 401-403: This paragraph claims that the secondary frequencies are important. However, Figure 4 shows that the top Fourier component explains more than 99% of the variance in all of the neurons. How does one reconcile these two results?\n\n[1] Gromov, Grokking modular arithmetic (2023)\n\n[2] Doshi et al., To grok or not to grok: Disentangling generalization and memorization on corrupted corrupted algorithmic datasets (2024)\n\n[3] Zhong et al., The Clock and the Pizza: Two Stories in Mechanistic Explanation of Neural Networks (2023)" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The analysis of secondary frequencies is new.\n\n- The results are presented in a quite candid fashion, with the limitations/discrepancies of the analysis freely discussed, along with potential explanations.\n\n- The code used to produce the results is provided in the form of a link to a Colab Notebook." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the role of MLP layers in a constant-attention Transformer (i.e. ReLU-MLP network) trained on modular addition. They write the overall operation of the network as $embedding \\\\rightarrow ReLU \\\\rightarrow fully-connected$ and write the trigonometric functions corresponding to these components. They show that the sum over same-frequency neurons implements a sum over random phases, which can thought of as approximating an integral. They provide empirical evidence for their claims and show that their method enables the computation of non-trivial error bounds on the outputs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concerns lie in regards to the novelty and soundness of this work.\n\n- The authors only consider Transformers with constant attention. This makes the network a ReLU-MLP and not a Transformer. This should be made clear in the paper. If the objective is to study Transformers, then non-constant attention should also be examined.\n\n- Interpretability for MLPs trained on modular addition has already been extensively studied in literature [1,2]. Although these prior works focused on MLPs quadratic activations, many of the qualitative conclusions are similar to the ReLU case studied in this work. The authors should present a detailed comparison to these works and highlight the novelty. (Specifically, what are the insights from the current work that are lacking upon merging the findings of [1,2,3].)\n\n- As stated, it is difficult to draw conclusions from the error bounds -- especially the part with splitting ReLU into identity and absolute value components. This part would benefit from further justification as well as clarity of notation (e.g. lines 309-310)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* You suggest that this approach could be useful for practitioners in other domains. What properties of the model/problem would make it more likely that this approach would yield a good explanation?\n* You leave deriving tighter error bounds to future work. What avenues for proving these bounds are you considering?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "* **Addressing open problems in mechanistic interpretability.** While modular addition is a well-studied problem, the MLP sublayer's behavior has mostly been treated as a black box. By explaining what the MLP is doing, the authors help move the study of this model beyond this black box treatment. While understanding one piece of a toy model may seem like a modest contribution, this is a major missing piece of an active area of research and a major step towards producing and verifying mathematical descriptions of MLP behavior. \n* **Resolving the discrepancy between \"clock\" and \"pizza\" algorithms.** In section 5, the authors solve the mystery of why \"pizza\" algorithm logits resemble \"clock\" logits by explaining the role that secondary frequencies play. This is also an unresolved question in modular addition.\n* **Balance of perspectives.** The authors are thorough about including empirical/computational results, mechanistic descriptions of the modular addition task, and proofs. They also extend results to 150 other transformer models (which differ in their random seed)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "MLP layers in transformer models remain poorly understood in the mechanistic interpretability literature; even detailed analyses of toy models like modular addition transformers tend to treat them as black boxes that compute certain functions and verify their input-output behavior by exhaustive search over the input space. The authors propose to fully explain—by which they mean find a description that is linear in the number of parameters—the MLP's behavior in the modular addition task. They determine that the MLP is responsible for numerically computing an integral of a trigonometric identity that is useful for the modular addition task, and even prove non-vacuous bounds on the error of their description." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* **Limited scope.** Although the paper is thorough and addresses an open question in mechanistic interpretability, its concrete contribution is simply to explain one component of one model (albeit under multiple random seeds) on one toy problem. While this is a step forward, and the authors briefly describe the broader implications for non-algorithmic interpretability practitioners in Section 7, the paper's contribution is circumscribed (arguably modest) as written. \n* **Bounds are crude.** As the authors acknowledge (354), the bounds they prove are quite loose, owing to an incomplete understanding of the model. Proving tighter bounds or arguing more convincingly for why tighter bounds are not possible would improve the paper. \n* **Clarity.** The argumentation in some of the mathematically dense sections (especially Section 5) can be hard to follow." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Have the authors investigated the effect of weight decay tuning on the clarity of their experimental results?\n2. What factors influence the emergence of secondary frequencies in the model?\n\nLine 023: \"trignometric\" typo in abstract\nLine 992: Broken equation ref" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Simple setting with a clear explanation of their setting\n2. Offered empirical check of their theoretical results" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper examines how 2-layer ReLU MLPs implement the trigonometric identities required by the 'Pizza' algorithm (previously described in earlier work) within a one-layer modified transformer architecture. The authors found that the MLP approximate a trigonometric integration by summing over its hidden neurons, they also provide non-vacuous bounds on the output of the MLP on all inputs in time linear in the parameters." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The work only focuses on explaining details of a component from the previously proposed Pizza Algorithm (Zhong et al., 2023), specifically how MLPs sum periodic features generated before activation functions. Furthermore, the simplified model (lines 125-126) is nearly identical to the fully solved model presented in [1], with ReLU activation replacing the quadratic activation. Since exact finite summation for ReLU remains an open problem, the authors instead approximate the sum using integral methods and argue that neurons effectively approximate this integral. \n\nWhile I appreciate the authors work on explaining the role of secondary frequencies in Section 5, it represents only a minor contribution. Overall, this paper's contribution falls below the standard expected at ICLR-level conferences.\n\n[1] A. Gromov, Grokking modular arithmetic, arXiv:2301.02679" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Is it possible to extend the proposed framework to MLP layers in a deeper transformers?\n\n- The study seems only focuses on understanding MLPs within a constrained setting where the MLP approximates trigonometric functions, is it possible to extend to non-trigonometric functions? \n\n- I am still confused on the definition of “complete interpretation”, are there any evaluation metrics?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The authors provides a depth analysis, with a formal definition of “complete interpretation”, a clear problem setting, and an evaluation on a simplified toy model, which enables a focused and mathematically rigorous analysis.\n\n- To evaluate the practical usage of the numerical integration interpretation, the authors validate the methods by establishing non-vacuous bounds for the MLP outputs in linear time." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors address the challenge of interpretibility of MLP layers in transformer models. Focusing on toy models trained for modular addition, the authors go beyond “approximate explanation” and propose to form a “complete interpretation” through a quadrature scheme, where neurons can be understood as approximating areas under curves in trigonometric functions. The method allows them to bound the MLP in total time linear in the neuron numbers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper is in general hard to follow for a broader audience, especially for those unfamiliar with topics such as mechanistic interpretability, quadrature themes etc. \n\n- A discussion on the scalability beyond the toy models is missing. While a simple toy model allows for a focus analysis, a discussion on scaling to more complex real-word tasks is beneficial.\n\n- As the author mentioned that there are couple of previous studies on the interpretability of MLPs, a comparative study with existing methods would be beneficial to fully understand the advantage of the proposed method." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We provide mathematical and anecdotal evidence that an MLP layer in a neural network implements numerical integration." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024modular,\ntitle={Modular addition without black-boxes: Compressing explanations of {MLP}s that compute numerical integration},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yBhSORdXqq},\nnote={under review}\n}" }, "abstract": { "value": "Interpreting the behaviour of MLP layers in transformer models remains an open problem.\nThe goal of mechanistic interpretability is identifying the specific function learnt by the model, which is challenging to do in nonlinear domains---like MLPs---where low precision approximations of the learnt function can leave many degrees of freedom.\nWhile MLP layers that approximately implement linear functions can be well understood with low-precision approximations, we take on the challenge of approximating densely-connected MLPs.\nWe propose a formal definition for “full interpretation” of a circuit: a compression of the explanation to a size that is linear in the parameter count of the circuit.\nWe investigate in the classic setting of the toy models trained on modular addition Nanda et al. (2023) and Zhong et al. (2023), where the MLP layer is treated as a black box. We extend the analysis to describe exactly *how* the MLP layer computes the required trignometric identities. \nWe find that the MLP layer in one-layer transformers implementing the “pizza” algorithm can be understood as evaluating a quadrature scheme, where each neuron computes the area of a rectangle under the curve of a trigonometric integral identity.\nWe confirm this interpretation by using it to compress the MLP layer of a collection of modular addition transformers and prove non-vacuous bounds on the outputs of the MLP layer on *all* inputs in total time *linear in the number of neurons*.\nOur code is available at [https://tinyurl.com/mod-add-integration](https://tinyurl.com/mod-add-integration)." }, "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": [ "mechanistic interpretability", "proof", "guarantees", "interpretability", "numerical integration" ] }, "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/c2fb49ccd27ef2d6ce0fa9b8241caf1ae486b2bf.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": null, "title": { "value": "Modular addition without black-boxes: Compressing explanations of MLPs that compute numerical integration" }, "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": "refer to weakness" }, "rating": { "value": 10 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Achieves state-of-the-art performance using only a single codebook with 40 or 75 tokens, demonstrating remarkable efficiency.\n- Covers multiple domains, including audio, speech, and music.\n- Its single codebook capability supports efficient training of large language models (LLMs).\n- Provides a comprehensive analysis of different settings, along with detailed ablation studies that validate the impact of each component.\n- The paper is well-written and easy to follow, with a comprehensive analysis included." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces WavTokenizer, a codec tokenizer for audio that achieves extreme compression and superior reconstruction quality compared to previous state-of-the-art models. WavTokenizer requires only a single quantizer with 40-75 tokens per second for 24kHz audio, while still preserving high subjective quality and rich semantic content. Extensive experiments demonstrate WavTokenizer's effectiveness across speech, audio, and music, with competitive results in both objective and subjective metrics, and ablation studies confirm the impact of each component." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Here’s a refined version of the weaknesses:\n\n- The evaluation could be more thorough by incorporating existing benchmarks such as Codec-Superb and DASB to enable a more comprehensive comparison of the proposed method against existing models under standardized settings.\n- The model currently supports only a 24kHz sampling rate; I wonder if you anticipate any challenges in adapting WavTokenizer to different sampling rates. It would be valuable to study its performance across different sampling rates, such as lower (16kHz) and higher (44kHz or 48kHz) rates." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 3 }, "primary_area": null, "questions": { "value": "1, Missing baselines: SemantiCodec[1], Single-Codec[2], Mimi (Moshi) [3].\n\n2, Weak Semantic Performance: Compared to the ARCH benchmark (https://huggingface.co/spaces/ALM/ARCH), the semantic performance of WavTokenizer is far worse than models like HuBERT base. Therefore, the claim of \"rich semantic information\" does not hold up well. Moreover, it is unclear why WavTokenizer's semantic capabilities were compared only against acoustic codec models, rather than semantic codecs. Why not compare against semantic codecs like SpeechTokenizer[4], SemantiCodec[1], or Mini[3]?\n\n3, Single-VQ Assumption: The entire premise of the paper is based on the assumption that using a single-VQ codec is better for audio language modeling compared to RVQ. But is this correct? The paper mentions that RVQ-based codecs like DAC require 900 tokens per second, which is true, but it fails to acknowledge that RVQ codecs typically have a temporal resolution of only 50Hz, and the most advanced models like Mini have even lower resolution at 12.5Hz. In practice, language models typically need to model at frequencies of 50Hz or less. So, what is the real advantage of a single-VQ codec with 40 or 75 tokens per second? Moshi and Mini have already demonstrated success in low-latency speech dialogue applications, but what about single-VQ? Is reconstructing all acoustic details truly beneficial for language modeling? Does an LLM truly need to model timbre and detailed acoustic features, or is focusing solely on semantic content sufficient? Therefore, you must at least demonstrate the effectiveness of single-VQ in practical applications, as stated in the weakness section, rather than constructing the paper solely based on this assumption.\n\n[1]: Liu H, Xu X, Yuan Y, et al. SemantiCodec: An Ultra Low Bitrate Semantic Audio Codec for General Sound[J]. arXiv preprint arXiv:2405.00233, 2024.\n\n[2]: Li H, Xue L, Guo H, et al. Single-Codec: Single-Codebook Speech Codec towards High-Performance Speech Generation[J]. arXiv preprint arXiv:2406.07422, 2024.\n\n[3]:Défossez A, Mazaré L, Orsini M, et al. Moshi: a speech-text foundation model for real-time dialogue[J].\n\n[4]:Zhang X, Zhang D, Li S, et al. Speechtokenizer: Unified speech tokenizer for speech large language models[J]. arXiv preprint arXiv:2308.16692, 2023." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The overall approach is sound. The paper is straightforward and easy to understand, with a clear structure that makes it accessible to readers. Additionally, the introduction provides a good overview of the context and background, helping readers to understand the problem." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces WavTokenizer, a single-vq codec aimed at simplifying and improving the current audio language modeling approaches by replacing codec with multiple-vq. It claims to achieve competitive reconstruction quality while enhancing integration with language models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper introduces WavTokenizer, a single-VQ codec designed to replace the current codec + LLM audio generation systems. However, there is no experimental evidence to support that WavTokenizer can effectively achieve this. Existing codec + LLM systems can generally be categorized into RVQ-style generation (e.g., VALL-E, MusicGen, Moshi) or semantic-to-acoustic approaches (e.g., SpearTTS, SeedTTS, MusicLM), all of which demonstrate strong performance.\n\nFirstly, there are no experiments showing that WavTokenizer combined with an LLM outperforms any of the existing systems in practical applications such as TTS or music generation. Secondly, when considering the codec itself, WavTokenizer's reconstruction quality is significantly worse than RVQ-based codecs, and its semantic performance falls short compared to semantic token like HuBERT or WavLM.\n\nGiven these shortcomings, it is difficult to see any tangible improvements or significant contributions that WavTokenizer offers to the current field of audio generation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 3 }, "primary_area": null, "questions": { "value": "- How was VQ utilization calculated in Figure 2(b) of the paper?\n- It seems that WavTokenizer is not inherently streamable. Can WavTokenizer be extended to support streaming encoding and decoding?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The acoustic codec representation model is a crucial technology in the current speech domain. WavTokenizer addresses one of the core challenges in the field by achieving high-quality audio reconstruction with only 40 or 75 tokens.\n- WavTokenizer introduces a novel single-layer quantizer concept, demonstrating its potential in TTS tasks and offering a promising single-layer solution for codec-LLM architectures.\n- From a methodological standpoint, WavTokenizer revisits vector quantization (VQ) in the speech domain and proposes a larger codebook space, a more powerful decoder (with attention mechanisms), and an extended context modeling window. These innovations appear to be effective.\n- The model achieves strong experimental results across reconstruction tasks, semantic understanding tasks, and downstream TTS tasks.\n- The open-sourcing of the complete training and inference code, along with model weights, will contribute to the development of the research community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces WavTokenizer, while preserving the classical acoustic codec model paradigm, achieves high-quality audio reconstruction using only 40 or 75 tokens per second. By proposing a larger codebook space, integrating attention mechanisms, and extending the context window, WavTokenizer demonstrates impressive results in audio reconstruction, semantic understanding, and downstream TTS tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall, this work does not present significant weaknesses. However, the design of a highly powerful decoder in WavTokenizer raises some concerns. Specifically, I am concerned that the increased model parameters and the introduction of attention mechanisms may potentially slow down the codec's reconstruction speed?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Did the authors try other techniques to improve VQ codebook utilization? The current approach closely mirrors EnCodec, but DAC demonstrates that low-dimensional code lookups and L2-normalization can significantly improve RVQ scalability. A useful reference to consider is [1], which shows effective scaling strategies in image reconstruction that could be also valuable for audio codecs. I recommend continuing work on the paper, as a single-quantizer audio codec has potential, but the paper’s current scientific contribution is limited.\n\n\n2. What is the motivation for evaluating semantic representation? Neural audio codecs are expected to encode low-level acoustic features rather than abstract semantic concepts. Comparing audio codecs on semantic representation could be misleading, especially given the results e.g. all codecs score below 10% on the SLURP dataset, while self-supervised models like HuBERT achieve nearly 50% (not shown in this paper). This gap calls into question the statement on line 445 that \"WavTokenizer effectively captures rich semantic information\" which may be an overclaim. That said, audio codecs might have a significant impact on representation learning, as shown by EnCodecMAE [2], so it may be more appropriate to treat semantic representation as a downstream task. WavTokenizer's single codebook could be particularly useful for discrete targets in BERT-like setups.\n\n[1] Zhu, Lei, et al. \"Scaling the Codebook Size of VQGAN to 100,000 with a Utilization Rate of 99%.\"\n\n[2] Pepino, Leonardo, Pablo Riera, and Luciana Ferrer. \"EnCodecMAE: Leveraging neural codecs for universal audio representation learning.\"" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper effectively motivates and addresses an important problem in neural audio codecs - quantizing audio into a single sequence of tokens rather than multiple sequences, which complicates modeling for downstream tasks.\n- It presents useful findings regarding decoder design, which are supported by the ablation study. In particular, a Fourier-based decoder combined with an attention layer yields better results, while a time-domain decoder with attention performs surprisingly worse." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces WavTokenizer, a GAN-based neural audio codec that uses a single vector quantization (VQ) codebook, in contrast to previous methods that rely on multiple residual vector quantization (RVQ) codebooks. This results in a bitrate as low as 480 bits per second, and to recover quality, the authors propose replacing the time-domain decoder with a Fourier-based decoder, preceded by attention layers. The ablation study confirms these design choices, and a comprehensive evaluation, including both subjective and objective metrics, shows that WavTokenizer maintains competitive reconstruction quality with state-of-the-art models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the motivation for scaling the single VQ codebook to many entries is clear, the paper falls short of achieving high codebook utilization when expanding beyond a size of 4096. What the authors list as contributions in VQ, such as k-means initialization and random restarts, are in fact well-established techniques in neural audio compression, and this paper doesn’t offer any novel methods to improve codebook usage. This is somewhat disappointing, given that a key focus of the paper is to provide a single quantizer. More experimentation to scale the VQ codebook is needed, as the current contribution feels more incremental and may be better suited for a different venue." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024wavtokenizer,\ntitle={WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yBlVlS2Fd9},\nnote={under review}\n}" }, "abstract": { "value": "Language models have been effectively applied to modeling natural signals, such as images, video, speech, and audio. A crucial component of these models is the codec tokenizer, which compresses high-dimensional natural signals into lower-dimensional discrete tokens. In this paper, we introduce WavTokenizer, which offers several advantages over previous SOTA acoustic codec models in the audio domain: 1) extreme compression. By compressing the layers of quantizers and the temporal dimension of the discrete codec, one-second audio of 24kHz sampling rate requires only a single quantizer with 40 or 75 tokens. 2) improved subjective quality. Despite the reduced number of tokens, WavTokenizer achieves state-of-the-art reconstruction quality with outstanding UTMOS scores and inherently contains richer semantic information. Specifically, we achieve these results by designing a broader VQ space, extended contextual windows, and improved attention networks, as well as introducing a powerful multi-scale discriminator and an inverse Fourier transform structure. We conducted extensive reconstruction experiments in the domains of speech, audio, and music. WavTokenizer exhibited strong performance across various objective and subjective metrics compared to state-of-the-art models. We also tested semantic information, VQ utilization, and adaptability to generative models. Comprehensive ablation studies confirm the necessity of each module in WavTokenizer. The demo is available at https://wavtokenizer.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": [ "speech representation", "discrete codec", "audio language model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/50b0c8ec4bdd0bae24d37b95064f738115dc0864.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": "WavTokenizer: an Efficient Acoustic Discrete Codec Tokenizer for Audio Language Modeling" }, "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": { "value": "No 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": 3 }, "primary_area": null, "questions": { "value": "- Could this balance of bound conservativeness and faster computation speed also be achieved if we approach smaller subproblems with classical approaches? I wouldn’t expect it to be as drastic as the HiQ-Lip, but it might make the classical computation times much more reasonable for the scales presented in the paper.\n\n- The paper refers to the paper (Bartlett et al., 2017) as justification for the improved scaling constant. It would be helpful to point to a specific result in that paper since there is no other discussion in the paper about where it comes from or how it's derived." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This is the first work of my knowledge to utilize quantum hardware models to accelerate Lipschitz constant estimation and is interesting in concept. The graph coarsening and refinement strategy appears to be a novel contribution of theirs which seems generally useful.\n\n- The computation times reported for 3-5 layer networks show two orders of magnitude speed up over GeoLip while only being slightly more conservative. This is a promising result.\n\n- The paper is generally well-written and the theory is accessible to most readers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a hybrid quantum-classical algorithm for estimating the $L_\\infty \\rightarrow L_1$ Lipschitz constant of a fully-connected neural network. This is achieved by converting the Lipschitz constant estimation problem into an equivalent cut-norm problem which can be solved efficiently on a quantum device with the so-called Quadratic Unconstrained Binary Optimization (QUBO) model. Due to hardware limitations of current quantum devices, a graph coarsening and refinement strategy is used to break the problem into subproblems that can be handled by about 100 qubits.Their experiments show significant speed up in computation time for 2-5 layer MLPs over existing $L_\\infty \\rightarrow L_1$ Lipschitz estimation approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-Although the methodology is interesting in concept, the results for 1-5 layer MLP model trained on MNIST are not very intriguing from a practical standpoint. The paper would be stronger if there was some evidence for making progress in some down-stream application such as certified-robustness of the MNIST classifier. In that case, there are currently 1-Lipschitz regularized layers which can already achieve very good robustness results for large CNNs trained on MNIST [Prach 2022, Araujo 2023].\n\n-I feel that the community is not practically limited by scales of 3-5 layers shown in the paper. In my opinion the current challenges in Lipschitz estimation are more for larger scale models like ResNets trained on ImageNET with the goal to achieve some non-trivial certified robustness for classification. At this scale it's more of a memory limitation issue for solving SDPs (which can also be broken into sub-problems). At this scale, I’m concerned that the quantum hardware limitations are going to make the bounds too conservative. There might be more potential for GPU acceleration on classical computers [Wang 2024].\n\n- It's difficult to interpret the computation times presented in the paper since each algorithm is supposedly using a different computing architecture. Since HiQ-Lip is a hybrid quantum-classical algorithm, is it using the same classical architecture as the baselines with an additional simulated 100 qubits? I think this practical aspect requires more explanation somewhere in the paper.\n\n[Prach 2022] Almost-Orthogonal Layers for Efficient General-Purpose Lipschitz Networks\n\n[Araujo 2023] A Unified Algebraic Perspective on Lipschitz Neural Networks\n\n[Wang 2024] On the Scalability and Memory Efficiency of Semidefinite Programs for Lipschitz Constant Estimation of Neural Networks" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 experimental setup described in Section 6 lacks clarity, particularly regarding the software and hardware configurations used.\n2. How is the computation time for HiQ-Lip determined? Since the experiments are conducted on a simulated CIM, how is the time for the quantum computation component accounted for?\n3. Comparing Table 4 with Table 2 reveals that networks with a larger number of layers appear to require less time for estimation than two-layer networks, which contradicts my expectations. Can you provide a more detailed explanation of this phenomenon?\n4. A more comprehensive explanation of the coefficients used for 3- to 5-layer networks (i.e., HiQ-Lip MP B) would be appreciated." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The work is the first attempt to employ quantum computing to handle the task of Lipschitz constant estimation.\n2. The hierarchical strategy for solving large-scale QUBO problems makes it possible to apply HiQ-Lip on small-scale devices.\n3. The simulation results show the advantage of HiQ-Lip in computation time.\n4. The presentation of this paper is clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors introduce HiQ-Lip, a hybrid quantum-classical hierarchical method for estimating the global Lipschitz constant of neural networks. The problem of Lipschitz constant estimation is first transformed into a QUBO problem. Subsequently, this QUBO problem is solved hierarchically on Coherent Ising Machines (CIMs) to accommodate the system size supported by quantum devices. Experimental results on multi-layer perceptrons (MLPs) demonstrate that their method provides estimations comparable to classical state-of-the-art methods while requiring less computation time." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. As the paper primarily focuses on applying quantum computing to global Lipschitz constant estimation, it is uncertain whether the ICLR community will find this topic compelling.\n2. The paper lacks discussion on the theoretical guarantee about the approximation ratio of the hierarchical strategy to the global optimal of original QUBO.\n3. The experimental results are derived entirely from simulations under ideal conditions, without consideration for practical aspects of quantum devices such as finite shots, device noise, and limited coherence time. These non-ignorable imperfections could significantly impact the quality of solutions obtained from quantum algorithms in practice." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "I don't see any concerns for this paper." }, "flag_for_ethics_review": { "value": [ "No 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. Is it possible to scale the proposed method for larger networks?\n\n2. In the l2 case, LipSDP has been extended for implicit networks (R1, R2), residual networks (R3) and more general structures (R4). See the following papers:\n\n [R1]: Revan et.al. Lipschitz bounded equilibrium networks.\n\n [R2]: Havens at al. Exploiting connections between Lipschitz structures for certifiably robust deep equilibrium models. NeurIPS 2023\n\n [R3]: Araujo et.al. A unified algebraic perspective on Lipschitz neural networks. ICLR 2023.\n\n [R4]: Fazlyab et. al. Certified robustness via dynamic margin maximization and improved lipschitz regularization. NeurIPS 2023\n\n It will be very interesting if the authors can address more general network structures for the l-infinity case. Can the authors comment on the possibility of such extensions?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The idea from this paper is new and original.\n\n2. The perspective on using small-scale quantum devices for Lipschitz estimation of neural networks is new." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims improving the scalability of SDP-based Lipschitz estimation of neural networks via developing HiQ-Lip, a hybrid quantum-classical hierarchical method that relies on Coherent Ising Machines (CIMs). The authors convert the problem into a Quadratic Unconstrained Binary Optimization (QUBO) problem and claim that this is more adaptable to the constraints of contemporary quantum hardware. Finally, the authors provide some experiments on fully connected neural networks on MNIST to show that their method is comparable to GeoLip but accelerates the computation process for such relatively small scale problem." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The scale of the experiments is quite small. Recently, for l_2 Lipschitz bounds, the SDP method has already been scaled to ImageNet by the following paper:\n\nZi Wang et.al. (ICLR2024): On the scalability and memory efficiency of semidefinite programs for Lipschitz constant estimation of neural networks.\n\n The authors seem not aware of the above result which achieves scalability to ImageNet. The authors studied the l-infinity case here, but the scale is on the MNIST level. This makes me think the contribution by the authors is very incremental in comparison to the original GeoLip paper. A few ways to make the contributions more significant include: 1. demonstrate the proposed method on large scale networks; 2) extend the method for more network structures, e.g. implicit models, residual network, etc." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": 4 }, "primary_area": null, "questions": { "value": "1. Any reason why you didn't consider Quantum Annealers (QAs) with larger number of qubits such as DWave Advantage system that have ~5000 qubits? Could you potentially work with larger neural networks given up to 5000 qubits? \n2. Are all the experiments limited only to 100 qubits or variables? Have you thought about how varying the limit on the number of qubits affects the quality of the estimation? \n3. Have you thought about whether there are limits in terms of the size of the neural networks that you can use this approach with? \n4. If you use this approach with neural networks of size larger than the ones considered in the experimental evaluation and fix the maximum number of qubits to 100, can you quantify the performance degradation from the additional graph coarsening and subsequent refinement through experiments or theory?\n5. In all your experiments, the estimates of the global Lipschitz constant come close to GeoLip's estimates, have you tried to increase the size of the neural network in terms of number of hidden neurons or number of layers till either your approach fails to give close estimates to GeoLip? I would expect some performance degradation with larger neural networks due to inexact nature of estimation using graph coarsening/refinement." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "One of the strengths of the paper is that it presents a systematic and scalable way to frame the neural network global Lipschitz constant estimation problem as a QUBO that can be solved with a limited number of qubits. This approach could be adapted to other problems related to neural networks such as neural network training or neural network verification. The experimental approach is also very sound as the authors compared with a number of different methods to estimate the global Lipschitz constant such as GeoLip, LipOpt, Matrix Product (MP), Sampling and Brute Force (BF) and tried two-differnt scaling . The time comparison with LipOpt and GeoLip show that the approach offers time-savings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the problem of finding the global Lipschitz constant of ReLU neural networks using a hybrid quantum-classical hierarchical method. The problem of estimating the global Lipschitz constant of a ReLU neural network is converted into a Quadratic Unconstrained Binary Optimization (QUBO) problem. To address address the issue of limited number of qubits, the paper proposes a new HiQ-Lip algorithm that works by first translating the structure of the neural network into a graph with each node representing a neuron and edges representing the connection strengths, employing graph coarsening to reduce the number of nodes by merging them until the resulting QUBO can be solved directly on a small quantum-annealing based computer, then solving the QUBO and finally mapping the approximate solution from the coarse graph back to the original graph by solving optimization subproblems. The paper finally presents experiments with a two-layer neural network with varying number of hidden neurons and deeper networks with varying number of layers and shows that HiQ-Lip doubles the solving speed of and provides more accurate upper bound (using GeoLip as gold standard) compared to the existing best method LiPopt for two-layer networks and GeoLip for multi-layer networks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper doesn't take into account the latest state-of-art in terms of Quantum Annealers (QA) such as DWave Advantage System (https://www.dwavesys.com/solutions-and-products/systems/) that have ~5000 qubits. They limited themselves only to 100 qubits and simulated CIM. They could have scaled out to larger number of qubits and explored what is the tradeoff between using larger number of qubits versus the graph coarsening/refinement strategy in terms of time saving or estimation quality. Given that larger of qubits are available, what is the value of the graph coarsening/refinement approach? There is a lack of assessment on how much performance degradation arises from graph coarsening and refinement." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. As far as I can tell, the evaluation was conducted on a simulated CIM rather than actual quantum hardware. Are all experiments run on the same architecture? I did not find this clarification in the paper. Clarification on the uniformity of the experimental setup across all evaluations is necessary for ensuring reproducibility and accurate interpretation of results, especially since the advantage of HiQ-Lip mostly comes from runtime.\n\n2. Given that GeoLIP's foundations in Grothendieck inequalities and the Unique Games Conjecture (UGC) suggest the difficulty of improvement within polynomial time, the precision similarity between HiQ-Lip (on simulated CIM) and GeoLIP on two-layer networks raises intriguing questions. Could the authors explain this similarity? This could imply a consistent performance on similar problems, such as the cut-norm problem, and might have crucial theoretical implications, for example, regarding the validity of the UGC." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The estimation of Lipschitz constants in neural networks is an important problem in deep learning research, with implications for model robustness and generalization.\n\n2. Applying quantum algorithms to deep learning remains an understudied domain, offering potentially promising avenues for exploration, particularly in light of the increasing sophistication of quantum computing devices." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study presents HiQ-Lip, a hybrid quantum-classical hierarchical approach for estimating the global Lipschitz constant of neural networks. The methodology involves reformulating the Lipschitz constant estimation as a Quadratic Unconstrained Binary Optimization (QUBO) problem, making it suitable for quantum algorithmic solutions. To address the limitations of current quantum devices, the authors introduce a multilevel graph coarsening and refinement strategy, enabling the adaptation of neural network structures to quantum hardware constraints. Empirical evaluations are conducted to validate the effectiveness of the proposed methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Novelty is limited: The conversion of the Lipschitz constant problem to QUBO and mixed-norm formulations was established in [1]; the application of CIM to QUBO was known. The primary algorithmic contribution, a graph coarsening and refinement strategy, is a heuristic and lacks theoretical guarantees.\n\n2. The baseline comparison is insufficient: Authors claim that current SDP methods face challenges such as high memory usage and slow processing speeds. This is true for generic SDP solvers. However, recent advancements in SDP methods have significantly improved efficiency for deep networks and convolutional architectures. For example, [2] has improved the SDP for very deep networks, and [3] has extended the SDP resolution to convolutional networks. Although these works focus on $\\ell_2$ Lipchitz constant estimation, I don't see why they cannot be extended to $\\ell_\\infty$ Lipschitz constant. [1] has pointed out that there are no fundamental differences between $\\ell_2$ and $\\ell_\\infty$ SDPs.\n\n3. The evaluation methodology raises concerns: HiQ-Lip demonstrates inferior precision compared to GeoLIP [1], with improvements primarily in runtime. However, runtime comparisons are implementation and architecture-dependent and do not account for more efficient, tailored SDP solvers (see above). Additionally, the reported runtimes exhibit inconsistencies, with more complex networks (Net3-Net5) showing significantly shorter processing times than simpler ones (Net2), casting doubt on the reliability of the performance metrics. For example, HiQ-Lip for Net-2 takes 30 seconds, while solving Net3 only takes 6.5 seconds.\n\n\nMinor: \n\nBecause converting the Lipschitz constant problem to QUBO and mixed-norm problems was already established in [1], the authors might consider properly crediting these to [1] in section 3. Most of the content was already presented in [1].\n\n[1] Zi Wang, Gautam Prakriya, and Somesh Jha. A quantitative geometric approach to neural-network smoothness.\n\n[2] Anton Xue, Lars Lindemann, Alexander Robey, Hamed Hassani, George J. Pappas, and Rajeev Alur. Chordal sparsity for lipschitz constant estimation of deep neural networks.\n\n[3] Zi Wang, Bin Hu, Aaron J Havens, Alexandre Araujo, Yang Zheng, Yudong Chen, Somesh Jha. On the Scalability and Memory Efficiency of Semidefinite Programs for Lipschitz Constant Estimation of Neural Networks" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "HiQ-Lip accelerates tight Lipschitz constant estimation for neural networks using small-scale quantum devices, outperforming state-of-the-art methods in speed." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024hiqlip,\ntitle={HiQ-Lip: A Quantum-Classical Hierarchical Method for Global Lipschitz Constant Estimation of Re{LU} Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yCAigmDGVy},\nnote={under review}\n}" }, "abstract": { "value": "Estimating the global Lipschitz constant of neural networks is crucial for understanding and improving their robustness and generalization capabilities. However, precise calculations are NP-hard, and current semidefinite programming (SDP) methods face challenges such as high memory usage and slow processing speeds. In this paper, we propose $\\textbf{HiQ-Lip}$, a hybrid quantum-classical hierarchical method that leverages Coherent Ising Machines (CIMs) to estimate the global Lipschitz constant. \nWe tackle the estimation by converting it into a Quadratic Unconstrained Binary Optimization (QUBO) problem and implement a multilevel graph coarsening and refinement strategy to adapt to the constraints of contemporary quantum hardware. \nOur experimental evaluations on fully connected neural networks demonstrate that HiQ-Lip not only provides estimates comparable to state-of-the-art methods but also significantly accelerates the computation process. \nIn specific tests involving two-layer neural networks with 256 hidden neurons, HiQ-Lip doubles the solving speed and offers more accurate upper bounds than the existing best method, LiPopt.\nThese findings highlight the promising utility of small-scale quantum devices in advancing the estimation of neural network" }, "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", "Lipschitz Constant", "Neural Network", "Quantum-Classical Hybrid Method", "Coherent Ising Machine", "QUBO" ] }, "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/a2683f090808e1114171c021b20612915cd67603.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "HiQ-Lip: A Quantum-Classical Hierarchical Method for Global Lipschitz Constant Estimation of ReLU Networks" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could the authors provide the specific prompts utilized in the experiments detailed in the paper? Additionally, could you further discuss how different prompts might affect the mechanism's performance?\n2. Could the authors extend their experiment to include comparison with other mechanisms from similar studies?\n3. Given the paper's claim that the proposed mechanism can operate without LLM fine-tuning or access to model weights, could the authors extend the experiments to include utilization of various closed-source models?" }, "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 structured, offering clear explanations of the problems addressed and the key terms used.\n2. The proposed mechanism has a firm theoretical grounding.\n3. The proposed mechanism seems to demonstrate strong applicability to real-world scenarios, especially in the realm of online advertising with LLM-generated content." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an auction mechanism, MOSAIC, to aggregate the preferences of multiple self-interested advertisers over LLM-generated replies. The authors claim that this mechanism can converge to the outputs of the optimally fine-tuned LLM as computational resources increase, without requiring fine-tuning or access to model weights. The authors also present context-aware versions of MOSAIC, which accelerates convergence and yields high advertiser value and platform revenue. Experiments with a publicly available LLM demonstrate that MOSAIC achieves high advertiser value and platform revenue with minimal computational overhead." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty of this paper is limited and the scientific contribution seems to be not obvious. The proposed mechanism addresses the problems using conventional approaches. The authors can strengthen the differences between their proposed mechanism and the algorithms used in standard RLHF and the rejection sampling.\n2. The experiments are insufficient. There are no comparative analyses between MOSAIC and other mechanisms, which makes it difficult to comprehensively assess its performance. The experiments are restricted to the use of Llama-2-7b-chat-hf, lacking generalizability across different LLMs. It would be nice to compare with other mechanisms mentioned in the related work and evaluate across different LLMs like Llama-3-8B, T5, and Mistral-7B." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* In equation (1) in line 152, why do you define KL-divergence of the two distributions and why $\\pi_{ref}$ is on the right hand? Will other choices fail?\n * In the paragraph between line 196 and 201, the author says, \"*The VCG allocation rule requires calculating the exact optimal solution to the optimization problem, which is intractable for choosing an LLM to maximize Equation (1) and is even difficult for choosing a single optimal sequence. If a sub-optimal solution is chosen, VCG’s strategyproofness is no longer guaranteed (Nisan & Ronen, 2007; 1999; Lehmann et al., 2002).*\"\n * In my understanding, choosing an LLM to maximize Equation (1) is equivalent to choosing an LLM that satisfies (2). Since we already have access on $\\pi_{ref}$, when $r(x,y) = \\sum_i r_i(x,y)$ is upper bounded given $x$, it's feasible to sample from (2) by rejection sampling. Besides, we can also use $\\pi_{con}$ replacing $\\pi_{ref}$ to decrease variations. In this sense, I do not find the advantage of using MOSAIC rather than using VCG directly. Could you make some clarifications on this concern?\n * Even in the case that $r(x,y)$ is not upper bounded and VCG can not be implemented, it's possible that sub-optimal solutions are chosen and strategyproofness still hold, MOSAIC is exactly one example. Besides, I think MOSAIC will also behave badly in this case because the probability of large value of $r(x, y)$ is small, and it's likely to sample $M$ candidates $y_1,...,y_M$ with all $r(x,y_i)$s are small." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The author provides an example throughout the paper, making the model easy to understand. The aggregation of LLMs is an important topic, while the introducing of regularization of user preference is novel. The idea of approximating VCG is also interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a mechanism called MOSAIC, that aggregates multiple advertising LLMs (represented by reward functions) and try to find a distribution on replies that maximizes the total rewards of advertisers together with a KL-divergence regularization term on user preference $\\pi_{ref}$. MOSAIC takes the reward functions as input, and output a (stochastic) reply to the user.\n\nMOSAIC first samples $M$ candidate replies from arbitrary general and pre-defined distribution $\\pi_{gen}$. After this, MOSAIC computes the distribution on replies (regarding the distributions as mechanism allocation) and sample one reply from this distribution as final output. The mechanism payments are computed by Rochet payment (1987). As $M$ tends to infinity, MOSAIC is guaranteed to converge to VCG mechanism.\n\nIn experiments, the paper specifies reward functions and $\\pi_{gen}$ by LLM distributions and contextually prompting LLMs, respectively. Experiments show that contextual-prompting MOSAIC performs far better than naive MOSAIC with $\\pi_{gen} = \\pi_{ref}$, with the log-probability of reply close to optimal distribution." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major Issues:\n\n* I think the main drawback of this paper is that MOSAIC is only an approximation of VCG mechanism. Moreover, such approximatio\nn is natural and not challenging to discover and define. Specifically, VCG outputs the distribution on full reply space, while MOSAIC\n first discretizes the full space into finite points and then outputs a distribution on finite points. This operation is simple a\nnd do not change the nature of VCG mechanism. Therefore, I believe that the contribution of MOSAIC is not significant.\n* The other contributions of this paper, eg, context-aware LLM in Line 237 $\\pi_{con}$, and payment offset in Line 335, are case-by-case operations and do not bring significance on the mechanism itself. These ideas are also natural. Besides, the payment contribution is a direct application of Rochet payment (1987). Overall, I believe that the contributions are incremental and limited.\n* Regarding the experiments, the authors only compare the contextual-prompting MOSAIC with naive MOSAIC. There is no comparison\n with baselines, thus I cannot evaluate whether the performance of MOSAIC is acceptable from these experiments. A suggestion is\n to compare MOSAIC with VCG mechanism directly when VCG is applicable.\n\nBesides, the presentation needs improvement. For example,\n\n* In Line 419, it appears $\\pi_i$. However, $\\pi_i$ is only mentioned in Line 146. The definition of $\\pi_i$ should be better placed between its first use, i.e., just before Line 419.\n* In Section 1.2, the authors mentioned 7 contributions of this paper. However, these contributions are incremental, or only the technical details appeared in this paper.\n* In the paragraph of Line 242, the notation $\\pi_r$ have never defined before these appearances. I think it should be $\\pi_{ref}$.\n* In line 313-314, I think that the term $\\log (\\frac{\\pi_{ref}}{\\pi_{gen}}) should be included in the $\\exp(\\cdot)$.\n* In the appendix, the sections are titled with \"Details from Section xx\", \"Omitted Proofs from Section xx\". It is conventional to use 'in' instead of 'from'." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "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 organized.\n- LLM and its application to online advertising, especially the domain of ad auctions, is a novel area to be studied with practical relevance.\n- The theoretical results provided appear sound. The authors also clearly stated the difference between their proposed mechanism and previous auction mechanisms such as VCG.\n- The numerical experiments (and the motivating example) provide interesting insights." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors studied a setting in which the advertisers would like to influence the LLMs to generate their preferred contents and the platform would need to satisfy both advertiser preference and user utilities. The authors proposed an auction mechanism called MOSAIC that enjoys a number of advantages, including ensuring that truthful reporting is a dominant strategy for advertisers. The authors also suggested that the proposed mechanism is equipped with technical feasibility and practicality. To validate their claims, the authors provided both theoretical results and numerical experiments with a publicly available LLM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I wonder if the authors can provide more explanations for why the ref LLM is not performing as well as the context-aware LLM. I understand that the authors have provided some intuitive explanations in Sections 4, but the ref LLM still appears to be an intuitive choice based on Corollary 4.1. Would it be possible to gain more insights into this difference from a theoretical point of view? Does the authors have any preliminary insights for how to introduce contextual information into the current model? \n- The current framework does not consider a number of constraints that could impact the advertisers' decision-making process, such as each advertiser's budget and/or ROI constraints, or the maximum length of the LLM output. I wonder if the authors can comment on whether their mechanism can incorporate any of the above 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": 3 }, "primary_area": null, "questions": { "value": "1. The paper claims that MOSAIC’s allocation rule converges to an optimal distribution as computational resources increase. How does the rate of convergence depend on the number of candidate sequences?\n\n2. The allocation rule is based on importance sampling to estimate the probability distribution over sequences. What is the mathematical form of the variance for this estimator?\n\n3. How does the choice of the hyperparameter \\tau in front of the KL term quantitatively influence the trade-off between advertiser reward maximization and alignment with the reference LLM?\n\n4. The strategyproofness guarantee hinges on honest reporting by advertisers, but how robust is MOSAIC if advertisers engage in complex forms of gaming or misreporting? \n\n5. In scenarios with advertisers holding strongly opposing interests, does MOSAIC risk generating responses that conflict with each other?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper introduces MOSAIC which effectively combines the preferences of self-interested advertisers while maintaining user-centric content. This is achieved without requiring direct access to LLM weights, which adds flexibility and broad applicability.\n\nThe mechanism ensures that truthful reporting is a dominant strategy for advertisers, thanks to its carefully constructed payment and allocation rules. This strategy-proofness is backed by theoretical proofs.\n\nThe design is computationally efficient, using only API access and avoiding expensive fine-tuning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces MOSAIC, an auction mechanism designed for aggregating advertiser preferences within the outputs of large language models (LLMs), particularly for applications in online advertising. The authors address the challenge of balancing the interests of multiple advertisers with user-centric content by creating a mechanism that incentivizes truthful reporting from advertisers. MOSAIC uses an approach that combines an allocation rule based on importance sampling, allowing it to converge to the optimal output distribution as computational resources increase, without requiring direct model fine-tuning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is no real-world testing or application with actual advertiser data to validate the effectiveness of MOSAIC in practical applications.\n\n2. Although MOSAIC is designed to be efficient, it may still face scalability challenges as the number of advertisers or candidate replies grows. An analysis of how the mechanism handles these scenarios under different computational limits would strengthen the paper.\n\n3. Given that the mechanism optimizes for advertiser rewards and alignment with a reference LLM, there’s a risk of unintentional bias in the final output.\n\n4. MOSAIC relies on advertisers truthfully reporting their preferences and reward functions, yet it does not fully address how misreporting could impact results. Further discussion on mid-reporting would be important." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "we give a truthful auction mechanism allowing agents to bid to influence LLM outputs, with a focus on advertising applications." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024truthful,\ntitle={Truthful Aggregation of {LLM}s with an Application to Online Advertising},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yCEf1cJDGh},\nnote={under review}\n}" }, "abstract": { "value": "The next frontier of online advertising is revenue generation from LLM-generated content. We consider a setting where advertisers aim to influence the responses of an LLM to align with their interests, while platforms seek to maximize advertiser value and ensure user satisfaction. The challenge is that advertisers may misreport their preferences. To address this, we introduce MOSAIC, an auction mechanism that ensures that truthful reporting is a dominant strategy for advertisers, and which aligns each advertiser’s utility with their contribution to social welfare. Importantly, the mechanism operates without LLM fine-tuning or access to model weights and provably converges to the output of the optimally fine-tuned LLM for the platform’s objective as computational resources increase. Additionally, it can incorporate contextual information about the advertisers, accelerating convergence. Via experiments with a publicly available LLM, we show that MOSAIC significantly boosts advertiser value and platform revenue with low computational overhead. While our motivating application is online advertising, our mechanism can be applied in any setting with monetary transfers, making it a general-purpose solution for truthfully aggregating the preferences of self-interested agents over LLM-generated replies." }, "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": [ "mechanism design", "llm", "auction", "online advertising" ] }, "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/f643fac9096b0ed41b118fb55ebfb16cd221e38d.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/0350cb2e1b39da55d4afc94c5cd46dde64cb9b40.zip" }, "title": { "value": "Truthful Aggregation of LLMs with an Application to Online Advertising" }, "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": "Refer to the weaknesses above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The proposed mixed multi-layer defense strategy in this paper demonstrates technical innovation by combining various defense mechanisms, such as feature-based prompting and real-time adjustments. This approach significantly enhances the system’s tolerance to adversarial attacks, offering a fresh perspective on the research of graphic prompting adjustment systems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel approach to graph prompt tuning focused on robustness in adversarial attack scenarios. It proposes a highly extensible shield defense system with a hybrid multi-defense prompt and robust prompt tuning strategy, demonstrating theoretical feasibility. Extensive experiments in few-shot scenarios under various adversarial attacks show that their strategies significantly enhance prompt tuning resilience in downstream biased tasks. The use of feature-based prompts allows real-time adjustments, and the paper highlights the effectiveness of their methods in both adaptive and non-adaptive attack scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One weakness of the paper is the absence of a comparative analysis with existing systems or defense strategies against adversarial attacks. \n\nThe paper primarily focuses on theoretical analysis and experimental validation in controlled settings. However, the lack of real-world implementation or case studies hinders the practical applicability assessment of the proposed method. To enhance the relevance and impact of the research, future work could involve implementing the system in real-world scenarios and evaluating its performance in practical applications.\n\nThe evaluation section of the paper might benefit from a more comprehensive selection of evaluation metrics to assess the proposed method’s performance accurately. Specifically, incorporating additional metrics such as computational efficiency, scalability, or user experience aspects could provide a more holistic evaluation of the system’s capabilities and limitations. Including a diverse set of evaluation criteria would offer a more thorough understanding of the method’s strengths and weaknesses.\n\nThe paper could further strengthen its impact by discussing the generalizability of the proposed method to different types of adversarial attacks or diverse datasets. Providing insights into how the approach could be adapted or extended to address a broader range of security challenges would enhance the paper’s contribution to the field." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Why other methods' results are getting close to the proposed methods in 10-shot experiments? Is it possible that other methods will outperform the proposed methods with even more shots?\n2. Adding the figures to represent prompt design and robust optimization strategies stages could be better." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The idea and structure is easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores enhancing the robustness of graph prompt-tuning methods against adversarial attacks in graph neural networks (GNNs). It identifies a vulnerability in current prompt-based methods, which are highly susceptible to adversarial alterations. The proposed solution is a \"shield defense system\" that enhances robustness through two strategies: Direct Handling and Indirect Amplification. Direct Handling customizes multi-defense prompts based on node-specific attributes, targeting biased nodes introduced by attacks, while Indirect Amplification leverages few-shot learning and selectively focuses on untainted information, thereby circumventing the misleading data introduced by adversaries. Theoretical validation demonstrates that this approach can maintain unbiased outputs despite adversarial perturbations. Experimental results show that the defense system enhances resilience under various attack scenarios, outperforming existing prompts in maintaining accuracy and robustness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some of the tables are confusing, for example, the table 1. Pre-Training Strategies & Prompts appear simultaneously in the top left corner. Should we design the table in a better way or use other charts to represent these results?\n2. Section 2 Related work, if there is no more subsection 2.1, is not necessary.\n3. The whole framework should have a figure to clearly show the stages of prompt design and robust optimization strategies" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 check the question raised in the comments concerning the weakness of this study." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "First of all, this is an interesting work in investigating robust graph prompt tuning to increase the stability of GPL against potential graph poisoning attacks. Especially, if the test graph structure is intentionally manipulated, this method offers a solution to suppress the impact of the poisoning efforts over the decision of GPL." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a robust graph prompt learning method based on adversarial training to defend Graph Neural Networks (GNNs) against adversarial attacks. The method begins by identifying nodes typically targeted in these attacks, such as low-degree nodes, nodes with low central similarity within multi-hop subgraphs, and nodes linked to adversarially perturbed edges. Next, the approach fine-tunes three distinct types of graph prompts, each corresponding to one of the targeted node types. The objective of the prompt tuning process is to ensure: (1) smoothness in node embeddings within multi-hop subgraphs across the three types of graph prompts, (2) consistency between prompt-enhanced and unprompted node embeddings, and (3) smoothness across embeddings from different types of prompts. Finally, the study applies this method to multiple graph prompt learning models, evaluating its defense effectiveness against non-adaptive attacks and its performance in combination with other defense strategies for adaptive and graph poisoning attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are several points that need to be strengthened: \n\n1/ This work only focuses on node classification-oriented GPL and does not discuss graph poisoning attacks targeting at manipulating node attributes. It is not clear if the proposed robust prompt tuning method can be applied to edge or graph level learning tasks and can be adapted to both node- and edge-manipulation based attacks. I am asking this as GPL is a multi-task system that can transfer knowledge between different types of GNN learning tasks (node, edge and graph classification). From this perspective, the evaluation in this study is not sufficient. Extending the coverage over more graph datasets and GNN learning tasks should be considered. I'd recommend the author refer to the dataset choice and learning task configurations in [1]. \n\n[1] All in One: Multi-Task Prompting for Graph Neural Networks, https://arxiv.org/pdf/2307.01504. \n\n2/ The theoretical proof in Section 4.3 is not associated with the main claim (robust prompt tuning) in this work. This theoretical analysis should be dedicated to explain how the robustness is improved via the proposed tuning approach. However, the link between Section 4.3 and the algorithmic design is unclear in the current submission. \n\n3/ The ablation study may also need to consider one situation: what if every node belonging to the three types (low degree, low central similarity and out-of-distribution nodes) has the same graph prompting function ? Besides, which types of the nodes should be given more weights in the robust tuning process, or they are handled equally in adversarial attacks ? \n\n4/ What are out-of-distribution nodes in a graph? Citing the paper (Li et al) is not enough. Further discussion regarding what these nodes are and why they are concerned in adversarial attacks should be involved. \n\n5/ How do you choose the threshold values $\\tau_{degree}$ and $\\tau_{sim}$ ? Are they set empirically, or is there any protocol to choose adaptively in different GNN adversarial attack scenarios ? Similarly, how do you choose $\\tau_{tune}$ in Eq.15 ? Does it mean you completely exclude the subgraphs if the embeddings of nodes in these subgraphs meet the condition in Eq.15? Will it also bring harm to the utility of the trained GPL as it excludes information from the training process. \n\n6/ As an adversarial training method, it is not surprised to see the proposed defensive graph prompts can mitigate one attack method when they are trained to be resilient to this attack. I am wondering if the proposed method has the transferability. In particular, if it is trained with the perturbed graph using one of the attack methods (MetaAttack, Heuristic attack, Random attack, DICE) can be also resilient to the other three attacks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024gpromptshield,\ntitle={{GP}romptShield: Elevating Resilience in Graph Prompt Tuning Against Adversarial Attacks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yCN4yI6zhH},\nnote={under review}\n}" }, "abstract": { "value": "The paradigm of ``pre-training and prompt fine-tuning\", with its effectiveness and lightweight characteristics, has rapidly spread from the language field to the graph field. Several pioneering studies have designed specialized prompt functions for diverse downstream graph tasks based on various graph pre-training strategies. These prompts concentrate on the compatibility between the pre-training pretext and downstream graph tasks, aiming to bridge the gap between them. However, designing prompts to blindly adapt to downstream tasks based on this concept neglects crucial security issues. By conducting covert attacks on downstream graph data, we find that even when the downstream task data closely matches that of the pre-training tasks, it is still feasible to generate highly misleading prompts using simple deceptive techniques. In this paper, we shift the primary focus of graph prompts from compatibility to vulnerability issues in adversarial attack scenarios. We design a highly extensible shield defense system for the prompts, which enhances their robustness from two perspectives: Direct Handling and Indirect Amplification. When downstream graph data exhibits unreliable biases, the former directly combats invalid information by adding mixed multi-defense prompts to the input graph's feature space, while the latter employs a training strategy that circumvents invalid part and amplifies valid part. We provide a theoretical derivation that proves their feasibility, indicating that unbiased prompts exist under certain conditions on unreliable data. Extensive experiments across various adversarial attack scenarios indicate that the prompts within our shield defense system exhibit enhanced resilience and superiority. Our work explores new perspectives in the field of graph prompts, offering a novel option for downstream robust prompt fine-tuning." }, "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": [ "pre-training; prompt tuning; robustness; adversarial attacks." ] }, "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/67c8ccfdd3a8c0838aaa8ff4715b94cc399ffaaa.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": "GPromptShield: Elevating Resilience in Graph Prompt Tuning Against Adversarial Attacks" }, "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. Could the authors provide a theoretical explanation or analysis to clarify why NODEDUP’s node duplication approach is effective for improving cold-node representation?\n2. Since the authors claim that their method does not compromise warm-node performance, could they provide comparisons with some of the latest graph contrastive learning baselines, such as GCLMI [1], Sp^2GCL [2], POT [3], and GraphACL [4], proposed in 2024?\n\n\n\n[1] Xu, Yuhua, et al. \"Graph contrastive learning with min-max mutual information.\" *Information Sciences* 665 (2024): 120378.\n\n[2] Bo, Deyu, et al. \"Graph contrastive learning with stable and scalable spectral encoding.\" *Advances in Neural Information Processing Systems* 36 (2024).\n\n[3] Yu, Yue, et al. \"Provable training for graph contrastive learning.\" *Advances in Neural Information Processing Systems* 36 (2024).\n\n[4] Xiao, Teng, et al. \"Simple and asymmetric graph contrastive learning without augmentations.\" *Advances in Neural Information Processing Systems* 36 (2024)." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Comprehensive Experimental Evaluation**: \n\n The paper is well-supported by extensive experimental results, covering a wide range of datasets and comparisons. This thorough empirical analysis demonstrates NODEDUP’s efficacy in addressing the cold-start problem, adding substantial credibility and depth to the findings.\n\n2. **Clear Presentation and Visualization**: \n\n The paper is well-organized and clearly written, with visual aids that effectively communicate the experimental results. The figures and tables are particularly helpful in understanding NODEDUP’s impact across different settings, making the paper accessible and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces NODEDUP, a novel augmentation technique aimed at enhancing the performance of graph neural networks (GNNs) in link prediction (LP) tasks, specifically for cold-start or low-degree nodes. By duplicating low-degree nodes and linking each to its duplicate, NODEDUP provides a “multi-view” representation that improves embeddings for cold nodes while retaining the performance of high-degree nodes. Extensive experimental results demonstrate NODEDUP's ability to achieve significant LP performance gains on cold nodes across multiple datasets, highlighting its potential to address limitations in existing GNN-based LP methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While NODEDUP demonstrates effectiveness in addressing the cold-start link prediction problem, several critical limitations should be considered:\n\n1. **Limited Scalability**: \n\n NODEDUP's design focuses heavily on the \"duplication of cold nodes,\" making it overly specialized for the cold-start problem and potentially too simple for broader link prediction tasks. The technique lacks flexibility to handle other challenges, such as heterophilic graphs or highly dynamic networks, which require adaptable augmentation strategies. This limits NODEDUP’s versatility as a universal approach in graph contrastive learning.\n\n2. **Lack of Theoretical Foundation**: \n\n Despite extensive empirical results, the paper does not provide a theoretical explanation for why NODEDUP improves cold-node representation. A theoretical analysis would strengthen the validity and interpretability of the method, clarifying how and why node duplication facilitates performance gains for cold nodes in GNNs.\n\n3. **Outdated Baselines**: \n\n Although the authors claim that NODEDUP does not compromise overall performance while addressing the cold-start problem, they compare it only against older baselines rather than the latest advancements in graph contrastive learning. This raises questions about how NODEDUP’s performance aligns with state-of-the-art methods and whether it may lag behind the latest graph contrastive learning approaches in overall link prediction performance.\n\n4. **High Dependency on Hyperparameters**: \n\n The proposed NODEDUP method relies heavily on the selection of the hyperparameter δ, which determines the distinction between cold and warm nodes. Although the authors provide a heuristic method and a hyperparameter search experiment, they do not offer a clear rationale for selecting a specific value of δ. This raises concerns that an optimal δ may need to be tuned individually for each dataset, potentially limiting the method’s practical applicability and 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": 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 Section 3.2, you mention that duplicating cold nodes provides additional view for aggregation. However, could you clarify how this extra view differs fundamentally from simply adjusting the weights assigned to a node's own representation? What specific advantages does node duplication offer that cannot be achieved through weight adjustments alone? And how is it related to helping cold-start LP\n\nIn Section 3.2, you argue that more supervision signals for cold nodes can lead to better-quality embeddings. Can you provide a more detailed explanation of how the link between a cold node and its duplicate functions as a meaningful supervision signal? What evidence or reasoning supports the claim that this relationship enhances the learning process in a way that traditional training methods (e.g. self-loop) do not?\n\nIn your paper, you define cold-start nodes using a fixed threshold of 2 for node degrees. Given that the effectiveness of this threshold may vary based on different graph properties, such as node degree distribution, and overall connectivity, how do you justify the choice of this specific threshold? Have you considered the impact of varying this threshold on the identification of cold-start nodes and the subsequent performance of your proposed method?\n\nIs there any reasoning why NodeDup and NodeDup(L) consistently outperforms GSage in the warm nodes and overall settings. As explained in the paper, the data augmentation is done on cold-start nodes only. How does it affect the performance of the warm nodes and overall settings?" }, "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 stands out for its simplicity and ease of integration with existing GNN architectures. By requiring only the duplication of low-degree nodes and the establishment of connections between these duplicates, it provides a straightforward augmentation technique that can be seamlessly applied to various GNN models.\n\nThe paper has extensive experiments to demonstrate the effectiveness and superiority of the proposed method. By evaluating the approach across multiple benchmark datasets, the authors provide strong empirical evidence supporting their claims." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a simple yet effective graph data augmentation method called NodeDup to improve the performance of Graph Neural Networks (GNNs) in link prediction tasks, particularly for low-degree or \"cold\" nodes. The method duplicates low-degree nodes and creates links to their duplicates, thereby providing a \"multi-view\" perspective during training. It addresses the well-known cold-start problem in recommendation systems and other applications by augmenting the training data without sacrificing the performance of well-connected nodes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper appears to lack sufficient novelty in its contributions to the field of link prediction. The underlying idea of augmenting data through duplication has been explored in various contexts. Additionally, the paper does not convincingly demonstrate how NodeDup outperforms or fundamentally enhances prior methods, leading to concerns about the overall impact of the contribution.\n\nAlthough the paper provides some justification for the proposed method, it could delve deeper into the theoretical foundations of why node duplication is particularly effective for cold-start link prediction. More comprehensive discussions could help solidify the rationale behind the approach.\n\nThe paper lacks a thorough investigation into the node degree distribution of the graphs used in the experiments, which can significantly impact the efficiency of the proposed method. In cases where the graph exhibits a highly skewed degree distribution, the additional complexity introduced by duplicating cold-start nodes could potentially lead to a doubling of the original computational complexity." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see my previous comment." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The paper is clearly written and easy to understand.\n\n2. The authors claim that the proposed method demonstrates significant improvements in performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose NODEDUP, a simple yet effective augmentation technique aimed at improving link prediction (LP) performance for low-degree nodes in Graph Neural Networks (GNNs) while preserving performance for high-degree nodes. NODEDUP works by duplicating low-degree nodes and creating links to their duplicates before applying standard supervised LP training. This “multi-view” approach significantly enhances LP performance for low-degree nodes without negatively impacting high-degree nodes. As a plug-and-play module, NODEDUP integrates easily into existing GNNs with minimal computational cost. Extensive experiments show average relative improvements of 38.49%, 13.34%, and 6.76% for isolated, low-degree, and warm nodes, respectively, compared to traditional GNNs and state-of-the-art cold-start methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The problem addressed in this paper does not align with the experimental datasets used. The cold start issue is a significant challenge in recommendation systems, and while the authors frequently mention the intent to tackle this problem, they rely on publicly available general graph datasets, such as citation networks, for their experiments. Crucially, they do not utilize any recommendation system datasets, such as Movielens-1M or Amazon-Books.\n\n2. The work lacks theoretical guarantees. Although the authors attempt to explain in Section 3.2 how node duplication aids cold start link prediction, I would prefer to see definitive theorems or lemmas presented in the paper. This would provide a more solid theoretical foundation for the work.\n\n3. In Table 18, GCN+NodeDup shows a decline in overall metrics compared to GCN. What is the reason for this phenomenon?\n\n4. There are several works designed to solve link prediction problems using GNN methods, such as references [1], [2], and [3]. However, the authors do not compare NodeDup against these methods in their experiments.\n\n5. The method description lacks clarity. In the NodeDup approach, determining which nodes are cold starts is a crucial step. However, in the core algorithm (Algorithm 1), the authors do not demonstrate how to deterministically obtain the set of cold start nodes V_{cold}.\n\n6. The authors should provide a direct comparison with adding self-loops as an augmentation baseline is missing, which could help clarify the advantages of NodeDup and NodeDup(L) over simpler alternatives.\n\n7. The authors state that OGB datasets were not used because they ``lack a substantial number of isolated or low-degree nodes'' (Lines 892-893). However, even though OGB datasets may primarily include high-degree nodes, they still contain a subset of lower-degree nodes, which would allow for a relevant evaluation of the proposed method's performance across both node types. Additionally, testing on these datasets could strengthen the claim that the method does not compromise warm node performance, as OGB is a widely recognized benchmark.\n\n[1] Bai Q, et al. HGWaveNet: A Hyperbolic Graph Neural Network for Temporal Link Prediction. WWW-23.\n\n[2] Zhu Z, et al. Neural Bellman-Ford Networks: A General Graph Neural Network Framework for Link Prediction. NeurIPS-21.\n\n[3] Cai L, et al. Line Graph Neural Networks for Link Prediction. TPAMI-21." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "**Q1**: What is the key insight and contribution of the proposed heuristic strategy?\n\n**Q2**: Figure 2 shows that NodeDup significantly improves performance compared to GraphSage on cold nodes while achieving results comparable to GraphSage on warm nodes. However, it raises the question of why the overall performances of two models are nearly identical.\n\n**Q3**: The additional time complexity of the decoder should be $O( |\\mathcal{V}\\_{cold}| D)$ rather than $O((M+|\\mathcal{V}\\_{cold}|) D)$?\n\n**Q4**: Whether self-loops are included in the adjacency matrix of the datasets used or not?\n\n**Q5**: Further explanation about the differences in performance between NodeDup and NodeDup(L) on warm nodes is needed.\n\n**Q6**: Additionally, there are minor typos, such as in line 400, where \"did\" is incorrectly used." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1: This paper focuses on addressing the cold-start problem in link prediction tasks, a significant issue with numerous practical applications in real-world scenarios.\n\nS2: The authors conduct extensive experiments to demonstrate the effectiveness of their method compared to traditional GNNs, cold-start GNNs, and other data augmentation techniques.\n\nS3: The proposed method is simple and can be easily applied to different GNNs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces NODEDUP, a data augmentation method designed to enhance the performance of GNNs on low-degree nodes in downstream link prediction tasks. By duplicating low-degree nodes and creating links between nodes and their own duplicates, NODEDUP can significantly improve prediction performance on *cold nodes* without compromising overall performance on *warm nodes*. Extensive experimental results demonstrate the effectiveness of NODEDUP compared to base GNNs, cold-start GNNs, and other data augmentation GNNs. Additionally, NODEDUP is a plug-and-play module that can be easily applied to different GNNs with minimal additional cost." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**W1**: The primary concern of this paper is its limited novelty. The main idea of duplicating cold nodes is not a novel strategy within existing GNN frameworks. Additionally, the distinction between NodeDup and the common data pre-processing step of adding self-loops is not clear. When adopting general message passing GNNs, isolated nodes can also treat themselves as neighbors and update node representations using both $\\boldsymbol{W}\\_1$ and $\\boldsymbol{W}\\_2$. Without a clear differentiation, it is challenging to justify the novelty of NodeDup over existing techniques.\n\n**W2**: It is unclear why such a simple strategy significantly enhances overall GNN performance on warm nodes, particularly on the *citeseer* dataset, which contains a large number of isolated nodes with zero degrees. Consider an example in online social networks, where a node $u$ represents a famous individual with millions of followers, and most of $u$’s followers may have few connections with other users, serving as *cold nodes*. In this scenario, adding self-loops using NodeDup(L) on these cold nodes may introduce additional noise in predicting the *warm node* $u$. This could potentially degrade the model's performance rather than improve it. \n\n**W3**: Additional theoretical analysis from a spectral perspective, such as examining changes in the graph Laplacian after self-augmentation, could enhance the understanding of the self-augmentation strategy. Such analysis might provide deeper insights into why and how the self-augmentation strategy works, potentially revealing underlying mechanisms that contribute to its effectiveness or limitations.\n\n**W4**: More experiments on larger datasets, such as collab, ppa, and citation2 from the OGB benchmark datasets [R1], should be included in the main text. The current experimental setup may not be sufficient to generalize the findings across different types of graphs and scales. Including these larger datasets would provide a more comprehensive evaluation of the method scalability and robustness. Additionally, a comparison between NODEDUP and more recent link prediction models, such as MPLP [R2], is necessary.\n\n---\n\n[R1] W. Hu, M. Fey, M. Zitnik, Y. Dong, H. Ren, B. Liu, M. Catasta and J. Leskovec. Open Graph Benchmark: Datasets for Machine Learning on Graphs. 2020. NeurIPS(33): 22118-22133.\n\n[R2] K. Dong, Z. Guo, N. V. Chawla. Pure Message Passing Can Estimate Common Neighbor for Link Prediction. arXiv:2309.00976." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024node,\ntitle={Node Duplication Improves Cold-start Link Prediction},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yCr55EjC1d},\nnote={under review}\n}" }, "abstract": { "value": "Graph Neural Networks (GNNs) are prominent in graph machine learning and have shown state-of-the-art performance in Link Prediction (LP) tasks. Nonetheless, recent studies show that GNNs struggle to produce good results on low-degree nodes despite their overall strong performance. In practical applications of LP, like recommendation systems, improving performance on low-degree nodes is critical, as it amounts to tackling the cold-start problem of improving the experiences of users with few observed interactions. In this paper, we investigate improving GNNs' LP performance on low-degree nodes while preserving their performance on high-degree nodes and propose a simple yet surprisingly effective augmentation technique called NodeDup. Specifically, NodeDup duplicates low-degree nodes and creates links between nodes and their own duplicates before following the standard supervised LP training scheme. By leveraging a ``multi-view'' perspective for low-degree nodes, NodeDup shows significant LP performance improvements on low-degree nodes without compromising any performance on high-degree nodes. Additionally, as a plug-and-play augmentation module, NodeDup can be easily applied on existing GNNs with very light computational cost. Extensive experiments show that NodeDup achieves 38.49%, 13.34%, and 6.76% improvements on isolated, low-degree, and warm nodes, respectively, on average across all datasets compared to GNNs and state-of-the-art cold-start methods." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Graph Neural Network", "Link Prediction", "Cold-start", "Graph Augmentation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/75ce8b1701660713dee4dc8c7b8f06933413a560.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Node Duplication Improves Cold-start Link Prediction" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. **Test Distribution Estimation:** In line 212 and from equations (4) and (6), it is described such that a batch of test samples arrive at each time step. However, prior baselines TPT, TDA perform single image TTA. Further, in Implementation details, in line 706, it is mentioned batch size is 1. Please clarify if single image TTA is done here as well, for fair comparison. If so, what does $n$ refer to in equations (5) and (6). How are $\\mu_k$ and $\\sigma_k$ estimated in single image TTA. Are you storing features, as done in TTA as well, along with the statistics?\n\n2. **Selection of uncertain samples:**\nConfidence is softmax applied over the similarity scores only right? Why is there such a large discrepancy using these two similar metrics (Table 7)? Also, is this similarity and confidence estimated from zero shot classifier or the classifier proposed? And why'd you choose what you choose?\n\n3. **Sensitivity to hyperparameters:** How sensitive is the method on the choice of the parameters $\\sigma, \\eta, \\rho$, as it's not practical to assume access to validation data before actually doing TTA? \n\n4. **Human in the loop TTA:** Please refer to the weaknesses and address the relevant concerns raised." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- They propose to estimate the class distributions in an online manner. \n- The adaptive fusion of zero shot text classifier based predictions and the distribution based feature similarities is simple and intuitive. \n- This being a backpropogation free approach, is very light-weight computationally, which is a great advantage for TTA.\n- The paper is well written and easy to follow, however several clarifications are required." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors address the problem of Test Time Adaptation of Vision Langugage models. They propose to continuously estimate the distribution of test samples, which they leverage through Bayes theorem, to make the final test predictions. They also collect human feedback to receive supervision for uncertain samples." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Distributional Test Time Adaptation:** In a single image TTA setting, the whole section 3.2 is quite unclear. There is no **batch** of samples in this setting. So, how are the class distributions actually updated at each time step.\n\n2. **Samples for distribution estimation:** Are all test samples used for updating the class distributions? Wouldn't the use of low confident/uncertain samples lead to bad parameter estimates in eqn(4)?\n\n3. **TTA with human feedback:** While this is one of the major contribution of the paper, it appears to only result in modest improvements. 5% and 15% is a lot of data to ask labels for, from a human. However, the results improve only of the order of 1-2%. This makes the efficiency of this whole process questionable.\n\n4. **Performance evaluation of TTA with human feedback:** As the test samples arrive in an online manner and based on uncertainty, how is the final accuracy evaluated here? Are these samples inclusive when evaluating accuracy? If so, you should be using the ground truth as predictions for actively labeled samples. Then the accuracy should be up by about 5% or 15%. As this is not the case in the results reported, are the labeled samples excluded from evaluation? This needs to be clarified. For fair comparison, all results should be reported on the complete test set, even when using human feedback.\n\n5. **Need stronger baselines for TTA with human feedback:** To study the role of TTA with human feedback, stronger baselines need to be established, using different selection strategies, like random, confidence, entropy etc. and report the accuracy of complete test set. As all strategies have same amount of labeled samples included, the performance improvement due these strategies as well as the gains wrt no human feedback can be assessed. \n\n6. **Amount of human feedback:** 5% and 15% is a lot of supervision and this may not be feasible during test time. It's more practical to ask labels for about 1-2% of test data. Experiments with stronger baselines, with lesser supervision, with correct evaluation method, is required to actually understand and evaluate the role of human feedback in TTA.\n\n7. **Choice of hyperparameters:**\nIn Implementation details, it is mentioned that validation sets are used to choose the hyperparameters. However, in TTA, one does not have access to any data from the test distribution apriori. Hence, validation data is not accessible in practice. Well, if one had access to validation data for test data, it provides a lot more information and could be used for more than just hyperparameter tuning." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. This paper is novel and interesting; would you consider making the code open source?\n2. In Table 1, for ResNet-50, DOTA’s performance is only slightly better than TDA, with an average improvement of just 0.15%. I would like to understand the reason for this marginal gain." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper is well written and easy to follow.\n2. The idea is interesting, the motivation of this paper is clear as well as the novelty of the method.\n3. The proposed method is extensively tested against prior work and outperforms on a variety of tasks/baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The author presents DOTA, a method that adapts to deployment conditions by continually estimating test sample distributions rather than memorizing them. Using Bayes’ theorem, Dota computes posterior probabilities for real-time adaptation. A human-in-the-loop feature also gathers feedback on uncertain samples, enhancing adaptability. Experiments show Dota outperforms current methods with continuous test-time learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper does not include a detailed case study focused on a particular domain or a challenging dataset.\n2. This paper does not include experiments assessing the model's sensitivity to hyperparameters. A detailed analysis of hyperparameter tuning could offer valuable insights into the robustness and generalizability of the proposed approach.\n3. The paper could benefit from additional visualizations illustrating Test-time adaption with human feedback" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How does the performance change w.r.t. the data stream?\n2. What is the impact of the various hyper parameter?\n3. How does the test-time forgetting phenomenon happen?\n4. How does the model compare with other baselines, e.g., DMN?\n5. How does the work relate to existing ones on prototypical networks and the various TTA settings?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. DOTA revisits principles in the literature on continual learning via nearest class mean classifiers (e.g., [a,b]) for improving the performance of VLMs at test time. Overall the approach is easy to implement and can be considered a valid baseline for future works, performing continuous TTA without the need for storing a cache, as in TDA.\n\n2. The article is well-structured and easy to follow, guiding the reader through all the design choices. \n\n3. DOTA is effective (as shown in the comparisons with TDA, e.g., Tab. 1 and Tab. 2) and computationally cheap (as shown in Tab. 3). \n\n4. Fig. 3 provides an analysis of how the performance of the two models (TDA and DOTA) vary w.r.t. the number of samples, showing the advantages of the latter.\n\n**References**:\n\n[a] Mensink, Thomas, et al. \"Distance-based image classification: Generalizing to new classes at near-zero cost.\" IEEE transactions on pattern analysis and machine intelligence 35.11 (2013): 2624-2637.\n\n[b] Bendale, Abhijit, and Terrance Boult. \"Towards open world recognition.\" Proceedings of the IEEE conference on computer vision and pattern recognition. 2015." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The article addresses the problem of test-time adaptation of vision-language models. Differently from previous works (e.g., TDA, Karmanov et al. 2024) that use a cache to store samples, the proposed method, DOTA, stores an online estimate of the statistics (mean and variance) of each class of interest. These statistics are then used during inference to refine standard CLIP predictions. An active learning strategy exploiting this statistic is also proposed to improve the performance of the model further, asking a user to annotate the least confident examples. Experiments on a wide range of tasks show the efficacy of this approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. DOTA continually updates its estimates of the statistics. Those might be affected by various factors linked to the experimental protocol, (e.g., order of the classes in the stream, batch size) as well as hyperparameters choice (i.e., the initial variance value, the shrinkage $\\epsilon$, $\\lambda$'s hyperparameters). Currently, the article does not provide too many insights on these factors, with the analysis mostly limited to the active learning percentile (i.e., Fig. 3). To assess the robustness of the model and provide a thorough study of its performance, it would be interesting to show results across multiple data ordering (i.e., currently it is not clear how many orders have been tested) and whether the performance changes w.r.t. the particular stream considered, even on the edge-cases where the data is non-i.i.d. [c]. Moreover, the hyperparameters may impact the speed of adaptation (e.g., variance initialization, $epsilon$) as well as how much the pretrained model is considered (e.g., $\\lambda$s): studying their impact is essential to fully evaluate the complexity of the approach and potential difficulties in applying it on real-world scenarios. \n\n2. While TPT and DiffTPT are strong models for test-time adaptation, they work on the episodic setting, i.e., where adaptation is held out on a single sample, and then the model is reset to its previous state. The possibility of storing/using test-time data (assuming coherence in the sequence) is a non-negligible advantage that DOTA has (and that it shares with TDA). This makes both the \"continual adaptation\" mark on TPT (Tables 1 and 2) potentially misleading, as well as TDA the only true baseline acting under the same priors of DOTA. To make the results stronger, it would be beneficial to add more baselines, such as DMN [d]. \n\n3. Following on the previous points, adapting to an evolving stream is a much more nuanced problem, where correlation between consecutive data may play an important role. Thus, various TTA settings with different types of stream and data dependencies (e.g., practical TTA [e], universal TTA [f]) could have been considered to further show the effectiveness of the approach. \n\n4. A key motivation behind DOTA relies on the test-time forgetting of TDA (lines 52-56). However, there are no experiments demonstrating this point (beyond the quantitative advantages of DOTA). An analysis clearly showing this phenomenon (and how DOTA is more robust to it) would strengthen the motivation behind the approach. \n\n5. The active learning strategy proposed to refine the performance for uncertain samples (Section 3.3) is a nice addition to make the approach more coherent but it lacks competitors. For instance, also TDA could employ a similar strategy (as the update of the cache is based on the confidence of the predictions). Moreover: (i) the accuracy with random feedback is also very close to those achieved with the proposed strategy (e.g., 0.6% gap on average in Tab. 6, 5% percentile); (ii) for the confidence-based scoring to work, the model is assumed to be calibrated, something not always true and that needs a proper discussion [g]; (iii) in Tab. 7 the accuracy of the random baseline is not reported: this is an important reference to put results into perspective. \n\n6. Related work (Section 2) provides a limited discussion on the various types of TTA settings (e.g., [e,f]) as well as on previous methods employing online updates of statistics for continual learning/open world recognition [b] or prototype-based few-shot learning [I,j]. Expanding the discussion would help to better contextualize the work in the current literature.\n\n**Minors**: \n\n- Footnote 1 hints that the model could be applied beyond CLIP. However, there are no experiments confirming this claim. It would have been more thorough to show other models (e.g., SigLIP [h]) to support it.\n\n- Table 4 shows the results only for DOTA. It would be interesting to see the same analysis for the other baselines (e.g., TDA) to contextualize/provide a reference for the results.\n\n\n**References**:\n\n[a] Mensink, Thomas, et al. \"Distance-based image classification: Generalizing to new classes at near-zero cost.\" IEEE transactions on pattern analysis and machine intelligence 35.11 (2013): 2624-2637.\n\n[b] Bendale, Abhijit, and Terrance Boult. \"Towards open world recognition.\" Proceedings of the IEEE conference on computer vision and pattern recognition. 2015.\n\n[c] Gong, Taesik, et al. \"Note: Robust continual test-time adaptation against temporal correlation.\" Advances in Neural Information Processing Systems 35 (2022): 27253-27266.\n\n[d] Zhang, Yabin, et al. \"Dual memory networks: A versatile adaptation approach for vision-language models.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2024.\n\n[e] Yuan, Longhui, Binhui Xie, and Shuang Li. \"Robust test-time adaptation in dynamic scenarios.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2023.\n\n[f] Marsden, Robert A., Mario Döbler, and Bin Yang. \"Universal test-time adaptation through weight ensembling, diversity weighting, and prior correction.\" Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2024.\n\n[g] Tu, Weijie, et al. \"An Empirical Study Into What Matters for Calibrating Vision-Language Models.\" International Conference on Machine Learning. 2024.\n\n[h] Zhai, Xiaohua, et al. \"Sigmoid loss for language image pre-training.\" Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.\n\n[i] Snell, Jake, Kevin Swersky, and Richard Zemel. \"Prototypical networks for few-shot learning.\" Advances in neural information processing systems 30 (2017).\n\n[j] De Lange, Matthias, and Tinne Tuytelaars. \"Continual prototype evolution: Learning online from non-stationary data streams.\" Proceedings of the IEEE/CVF international conference on computer vision. 2021." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 3 }, "strengths": { "value": "1. The writing and figures are good and easy to understand.\n\n2. The DistributiOnal Test-time Adaptation (DOTA) method for Vision-language foundation models without BP is simple yet effective during testing in new target domain, achieving a significant improvement compared to current state-of-the-art methods in most of the datasets.\n\n3. This paper first define the test-time adaptation problem with human feedback, allows the test-time adaptation for uncertain samples with human feedback.\n\n4. An adaptive final fusion probability is introduced to mitigate the potential negative impact when the number of test samples is insufficient." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a Distributional Test-time Adaptation (DOTA) method, which adapt the pretrained Vision-language foundation models (e.g., CLIP) to the test target domain by estimating the distributions of different categories for test samples continually. The authors further introduce a human feedback collaboration method which identifies uncertain samples to further enhance the adaptability. Extensive experiments on diverse datasets validate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed method estimate the data distribution of samples in the current test environment during testing, but there lack some evidence or visualization. Why updating the feature distribution for different categories works better?\n\n2. The DOTA method seems somewhat similar to the T3A method [R1], which continually maintains a memory bank for prototypes during the testing stage. Could the authors clarify and analysis the difference and the advantages of the proposed method?\n\n[R1] Test-Time Classifier Adjustment Module for Model-Agnostic Domain Generalization\n\n3. The proposed method seems works for not only VLMs but also other models in all classification tasks. Is it suitable for traditional TTA or Domain Generalization tasks?\n\n4. There could give more details and explanation about the $f_k(x)$ in Eq.(3).\n\n5. The uncertainty estimation method is simple. Is there any other uncertainty estimation method (like entropy) better?\n\n6. There missing an ablation study for the adaptive fusion probability." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We proposes a new Distributional Test-time Adaptation (DOTA) method, which continuously estimates the distribution of test samples and incorporates human-machine collaboration to handle uncertain samples." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024dota,\ntitle={{DOTA}: Distributional Test-Time Adaptation of Vision-Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yD2JMeKumt},\nnote={under review}\n}" }, "abstract": { "value": "Vision-language foundation models (e.g., CLIP) have shown remarkable performance across a wide range of tasks. However, deploying these models may be unreliable when significant distribution gaps exist between the training and test data. The training-free test-time dynamic adapter (TDA) is a promising approach to address this issue by storing representative test samples to guide the classification of subsequent ones. However, TDA only naively maintains a limited number of reference samples in the cache, leading to severe test-time catastrophic forgetting when the cache is updated by dropping samples. In this paper, we propose a simple yet effective method for DistributiOnal Test-time Adaptation (DOTA). Instead of naively memorizing representative test samples, DOTA continually estimates the distributions of test samples, allowing the model to continually adapt to the deployment environment. The test-time posterior probabilities are then computed using the estimated distributions based on Bayes' theorem for adaptation purposes. To further enhance the adaptability on the uncertain samples, we introduce a new human-machine collaboration paradigm which identifies uncertain samples, collects human-feedback, and incorporates it into the DOTA framework. Extensive experiments validate that DOTA enables CLIP to continually learn, resulting in a significant improvement compared to current state-of-the-art methods." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Test-time", "uncertainty", "vision-language models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/261bbc6ea262d382c59592ddefbfea93b7ceb710.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "DOTA: Distributional Test-Time Adaptation of 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": 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 would the author expect $k$ to scale with context length and the size of the model?" }, "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 investigates the timely question on how to speed up attention calculation and offers a new perspective on when attention calculation can be speed up.\n\n2. While focusing on the theoretical perspective, this paper present empirical validation on the correctness of the presented algorithm and required $k$ to approximate attention matrix." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper examines the possibility to compute attention using Fast Fourier Transformer when attention matrix can be well approximated by the sum of some convolution matrices. They show that when the Attention matrix can be broken down in this way, both inference and training can be done in almost linear time when $kd = n^{o(1)}$." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The algorithm 3 is deferred to appendix while being a key building block of the algorithm. In general, a more detailed explaination on how the binary search is carried out will improve the clarity of the paper.\n\n2. Because the current work is built upon the previous work with low rank assumption, a more direct comparison (either empirically or theoretically) between the current method and previous one should be presented." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- What relationship does the conv-basis rank ($k$) have to the context length?\n- Can your method be applied on top of existing pretrained transformers without any retraining or finetuning?\n- Why is the one provided experiment performed on such a small scale. Given that the method reduces the quadratic complexity of a transformer, shouldn't it be able to scale to huge context lengths without much trouble? It would be interesting to compare this to another method such as Streaming LLM [1] on a streaming perplexity task (like PG19) or other long context benchmark which scales to the order of millions of tokens. \n\n## Overall\n\nOverall, I think the work attempts to make an ambitious change to the attention operation in order to lower the complexity. However, I believe anyone reading this work will be left with many questions about the workings of the algorithms as well as the performance characteristics. I believe adding more experiments, ablations, and answering the questions posed here will aid in understanding the proposed method more fully." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The proposed method offers a way to reduce the prohibitive quadratic complexity of modern transformers.\n- As compared to previous works, the method appears to be much less restrictive on the structure of the attention mask which allows it to be applied to modern LLM settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a k-basis system for convolutional matrices which allows for sub-quadratic inference in attention with fewer constraints on the form of the attention as compared to previous works." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experiments section is rather light, and contains only one experiment. Given that the proposed method differs quite drastically from previous transformers, I would like to see more empirical validation and ablations on different aspects of the performance.\n\n- It is not clear to me if the experiment which was performed on Llama3 required any training from the proposed method or not.\n - If it did not require training, I would be interested to see the performance of the proposed method which did include training or finetuning.\n\n- The rank $k$ in k-conv is given, however after reading I am curious as to how $k$ relates to performance on long context tasks. Does $k$ need to scale commensurately with the context length?\n\n- I am left confused by the following items in Algorithm 2:\n - What is $u$ and where is it introduced in the method?\n - What is $v$ and where is it introduced in the method?\n - What is the $s$ subscript in algorithm 2 ($\\tilde{H}_s \\leftarrow M_s \\odot (QK^\\top_s)$) supposed to denote?\n - What is happening on line 8 after achieving $\\tilde{H}_s$\n\n- The algorithmic complexity looks compelling, but what is the actual wall clock time of the experiment compared to the quadratic tranformer? Can you run an additional experiment which looks at this, even if it only utilizes random inputs? I suspect that at smaller context lengths such as 2K which is provided in the experiment, the proposed method might perform worse in terms of wall clock time, but it would be interesting to know at what length the crossover happens to where the lower complexity starts to show real gains in 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": 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": "- As the authors mention, the proposed concept of the conv-basis approximation to sequence models seems quite related to prior work around sub-quadratic alternatives to attention, e.g. SSMs [1], Mamba [2]. Is there some way a conv-approximation of an existing attention matrix can be interpreted as a set of specific kernels for a subquadratic model?\n - In particular, I wonder if the conv-basis concept relates to the semiseparable matrices of Mamba2 [3]?\n- It would be interesting to see how the approximability of attention vs. number of conv-basis elements changes based on the task, e.g. do math reasoning tasks (GSM8K) require more basis elements than simple question-answering?\n\n1. Efficiently Modeling Long Sequences with Structured State Spaces\n2. Mamba: Linear-Time Sequence Modeling with Selective State Spaces\n3. Transformers are SSMs: Generalized Models and Efficient Algorithms Through Structured State Space Duality" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The authors investigate an important problem: how to speed up the efficiency of the self-attention mechanism, the computational bottleneck in Transformers.\n- The theoretical results are interesting, thorough, and clearly presented." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work investigates the problem of approximating matrices using a lower-triangular Toeplitz matrix basis, or \"conv-basis\", focusing specifically on the self-attention mechanism. The authors provide theoretical results about the approximability of lower triangular matrices into such a convolution basis, and provide algorithms for efficiently recovering this basis and using it to efficiently perform the self-attention operation using FFTs. Finally, the authors provide empirical results showing that their proposed method requires few basis elements to recover performance on the IMDB dataset when applied to Llama3 8B." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I think the theoretical results are quite interesting and deserve to be presented. However, the main weakness seems to be that the empirical results are quite limited.\n - The empirical speedup results are limited. Although asymptotically, the authors argue that the conv-basis approximation is more efficient, as far as I can tell, an empirical comparison of their method's speed vs. standard self-attention is not provided. It could be clarifying to provide such a comparison in the case of e.g. Llama3 8B, so that it's clear at what sequence length such an approach becomes more efficient in practice.\n - Additionally, the expressivity vs. efficiency tradeoff inherent to the method could be explored more thoroughly. The main theoretical result in the paper is that any lower triangular matrix can be well approximated in the conv basis, but accurate recovery may take as many as n basis elements in the worst case. The authors further support this result with experimental validation on the IMDB dataset with Llama3 8B, but this is only one task and relatively simple. It could be clarifying to provide additional experimental results on more recent LM benchmarks (e.g. HellaSwag, WinoGrande?)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024convbasis,\ntitle={Conv-Basis: A New Paradigm for Efficient Attention Inference and Gradient Computation in Transformers},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yD7oAhFEtD},\nnote={under review}\n}" }, "abstract": { "value": "The self-attention mechanism is the key to the success of transformers in recent Large Language Models (LLMs). However, the quadratic computational cost $O(n^2)$ in the input sequence length $n$ is a notorious obstacle for further improvement and scalability in longer contexts. In this work, we leverage the convolution-like structure of attention matrices to develop an efficient approximation method for attention computation using convolution matrices. We propose a $\\mathsf{conv}$ basis system, analogous to the rank basis, and show that any lower triangular matrix can always be decomposed as a sum of structured convolution matrices in this basis. We then design a fast algorithm to approximate the attention matrix via a sum of such $k$ convolution matrices. This allows us to compute the attention {\\it inference} via Fast Fourier Transforms (FFT) in $O(knd \\log n)$ time, where $d$ is the hidden dimension, and thus achieve almost linear time $n^{1+o(1)}$ in the practical scenario where $kd = n^{o(1)}$. Furthermore, the attention {\\it training forward} and {\\it backward gradient} can be computed in $n^{1+o(1)}$ as well. We provide theoretical guarantees on the run time and approximation error and conduct preliminary experiments to evaluate its effectiveness. We hope our new paradigm for accelerating attention computation in transformer models can help their application to longer contexts." }, "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": [ "Attention Acceleration", "Fast Fourier Transforms", "Gradient Computation" ] }, "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/a1c2d21ac4bbe2498d480ea6aa88f38ce9a8cd39.pdf" }, "presentation": null, "primary_area": { "value": "learning theory" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/34f735500c75cc23eba9063d9a3be5509f9ba160.zip" }, "title": { "value": "Conv-Basis: A New Paradigm for Efficient Attention Inference and Gradient Computation in Transformers" }, "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": "See above in weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Originality: The paper introduces a novel approach that uses causal model editing to generate faithful-unfaithful explanation pairs, offering a rigorous basis for assessing faithfulness in natural language explanations. This approach combines causality with faithfulness evaluation and tries to get to the model’s true reasoning processes.\n\n2. Quality: The paper is rigorous, with comprehensive experiments across three tasks and multiple language models. The inclusion of alternative model editing techniques and ablation studies further strengthens the evaluation framework.\n\n3. Clarity: The paper is well-organized, with clear, detailed explanations. Tables and figures present complex concepts in easy to understand manner. \n\n4. Significance: This work is significant in that its findings highlight specific improvements needed in LLM explanation fidelity.\n\nIn summary, this paper is a good contribution to the field of LLM interpretability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new framework to assess the faithfulness of natural language explanations generated by large language models (LLMs). To evaluate existing metrics, the framework employs model editing to create pairs of faithful and unfaithful explanations and tests various metrics using a benchmark of three tasks: fact-checking, analogy, and object counting. The study finds that while the CC-SHAP metric consistently outperforms others, many metrics fail to reliably capture faithfulness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The use of synthetic explanations may be limiting, as these pairs might not fully represent actual model-generated explanations. It would be helpful if the authors provided an analysis of how well synthetic explanations align with actual ones.\n2. The focus on three specific tasks (fact-checking, analogy, object counting) may not generalize well to more complex contexts. Adding diverse tasks or discussing broader applicability would be helpful. Have the authors considered experimenting with other complex contexts?\n3. Relying on diagnosticity as a faithfulness measure is overlooking other aspects of reasoning, like consistency and coherence. Including complementary metrics or discussing proposed framework’s limitations would be helpful." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Can the author say more on how do they generate synthetic explanations?\n2. Can the authors justify the use of word \"causal\" in the metric/framework they introduce?\n3. Fig. 6, seems to have a bug in the legend?" }, "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 - the motivation of the work is clearly presented, related works are well discussed, proposed approach and experiments are clearly described, and results are well discussed.\n\n2. The topic the paper focuses on is extremly important. Given the widespread usage of LLMs, it is very important to develop faithful methods to explain their predictions, but it is equally important to benchmark them.\n\n3. The experiments are diverse and include ablation studies to understand if and how the conclusion generalises." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents an empirical study to analyse the **reliability of faithfulness metrics for explanations of LLMs**. The authors propose \"Causal Diagnosticity\" as a metric to evaluate seven faithfulness metrics (for both post-hoc and Chain of Thought explanations) on three NLP tasks across four different LLMs. The results suggest that CC-SHAP is the most reliable metric among all the evaluated faithfulness metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper seems very applied to me with limited novelty. The authors expand an existing metric (called diagnosticity) to natural language explanations by arguing that random text cannot work as meaningful explanation (line 188..). However, this argument needs more backing/examples as random text can be considered as unfaithful explanation as done previously by Chan et al. 2022b.\n\n2. Secondly, the authors introduce model editing as a way to generate pair of explanations (faithful and unfaithful), but this might be limiting the analysis as a given model editing method may not work perfectly (because the knowledge may be incorrectly learned), hence, model prediction may change, but the reason is incorrect. The authors argue to use \"synthetic explanations\" to handle this, but it is not clear if insights on synthetic explanations are generalisable to the real world. For e.g., are the syntetic explanations guranteed to not hallucinate?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The framework on evaluating faithfulness metrics for natural language explanations is quite novel.\n\nThe use of model editing to create the three synthetic tasks is also very novel. \n\nExtensive evaluations of several different faithfulness metrics are used." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes Causal Diagnoticity, a benchmark to evaluate metrics that evaluate natural language explanations generated by LLMs. In particular, this benchmark creates carefully edited models for three tasks that lead to unusual reasoning in three tasks: fact-checking, analogy and object counting. For each model response, two explanations are synthetically generated, with one containing the correct reasoning and the other containing the incorrect reasoning. Then different faithfulness metrics are computed to evaluate this pair of explanations, and the score difference between the correct and incorrect explanations is taken as the quality (i.e., diagnosticity) of the metric. Experimental results suggest that CC-Shap is the best performing, but still leaves much room for improvement." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My biggest concern is with the generation of synthetic explanations, and the assumption that one is correct and the other is incorrect. In particular, while the model is edited on the particular fact, it is unclear that the particular editing causes the model to use the \"intended\" reasoning path, or the model is actually using some very different reasoning paths. For example, in the Rihanna example, it could be that the model editing removes \"Rihanna\" entity from the \"singer set\", and hence results in the model predicting no. In this case, the \"correct\" explanation should be something like \"No, because I do not find Rihanna in the singer set\" (though this language may be highly unlikely to be generated by an LLM). How to carefully eliminate such possibilities is, in my opinion, the most crucial issue in establishing the soundness of the framework.\n\nIn addition, for a slightly related question, can we really assess the faithfulness of natural language explanations that are not generated by the models themselves (i.e., having low or almost-zero probability of being decoded), or in other words, is such evaluation fundamentally meaningful? In \"traditional\" interpretability, all the explanations, such as feature attribution, concept and counterfactual, are generated using well-defined algorithms, and hence the \"meaningfulness\" of the explanation is of little doubt. However, with synthetic explanations, I am less sure, so I would hope that the authors could convince me on this aspect.\n\nOf a minor note, there are also alternative ways to generate and evaluate LLM-generated explanations, such as [1] and [2], that could be discussed in related work. Furthermore, there is even work that outright assert that LLMs cannot explain via natural language explantions [3].\n\n[1] https://arxiv.org/abs/2310.11207\n\n[2] https://aclanthology.org/2024.findings-acl.19.pdf\n\n[3] https://arxiv.org/pdf/2405.04382" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 following questions are about perplexity described in Section 5.2:\n - What is perplexity? Could the authors provide a definition and explain how it is calculated?\n - Why can low perplexity indicate a faithful explanation?\n - If low perplexity does indeed correlate with faithful explanations, then perplexity itself could serve as a metric for evaluating faithfulness. Why is it not evaluated with **Causal Diagnosticity**? Without assessing the **Causal Diagnosticity** of perplexity, how can it be sufficient to determine whether an explanation is faithful?\n\n2. Are the synthetically generated explanations produced by the LLMs themselves? In line 203, the authors mention that \"$\\overline\\theta$ generates the explanation $\\overline\\epsilon$ and $\\widetilde\\theta$ generates the explanation $\\widetilde\\epsilon,\" which is confusing.\n\n3. Is the use of $\\widetilde\\theta$ necessary?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Natural language explanations are becoming increasingly relevant due to advancements in LLMs, making them more accessible for end users. This work aids users in selecting a metric to evaluate the faithfulness of natural language explanations.\n- The idea of extending diagnosticity from feature attribution to natural language explanations is straightforward and relevant.\n- The authors conducted empirical experiments and discussed the effects of model knowledge editing methods, explanation types, and the reliability of these edits. The results demonstrate that CC-SHAP outperforms other metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose **Causal Diagnosticity**, a metric designed to evaluate the diagnosticity of existing faithfulness metrics for natural language explanations, where diagnosticity indicates how often a faithfulness metric favors faithful explanations over unfaithful ones.\n\nThrough evaluations of several post-hoc and Chain of Thought (CoT) faithfulness metrics across four tasks and four large language models (LLMs), the authors conclude that CC-SHAP outperforms the other metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Equation (5), the assumption is made that $\\overline\\epsilon_i$ is faithful to $\\overline\\theta$, while $\\widetilde\\epsilon_i$ is not. However, this is not guaranteed in the experiments. Though the authors discuss this in Section 5.2, it's unclear why perplexity can indicate faithfulness.\n\n\n2. The model editing discussed in Section 2.2 involves modifying the internal weights of LLMs, which restricts the experiments to open-source LLMs, excluding closed-source models like GPT-4.\n\n3. The model $\\widetilde\\theta$ is exclusively used to generate unfaithful explanations and is not incorporated in Equation (5). Since directly modifying $\\overline\\epsilon$ could also yield an unfaithful explanation related to $\\overline\\theta$, the necessity of using $\\widetilde\\theta$ is unclear.\n\n4. The placement of figures throughout the paper is disorganized, which hurts readability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Regarding model editing, is changing a single fact (i.e. “is paris the capital of france”) sufficient to override all pretrained knowledge in the LLM? I ask because this is a key part of the evaluation framework.\n- How might this framework, especially model editing aspects, be adapted to more complex/realistic scenarios?\n- Minor: I may be misunderstanding something here, but do you need to modify two models for causal diagnosticity? Is it not sufficient to change the knowledge in one LLM, and compare the modified LLM to the original?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Tailors existing datasets to the new tasks.\n- Comprehensive evaluation across a range of tasks, with many existing metrics.\n- The paper is well written. The main idea of using two models for causal diagnosticity is emphasized throughout. Separate applications are detailed well.\n- Decent evidence is provided to demonstrate that many existing faithfulness metrics are staggeringly weak, as they fail to pass tests on basic tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new framework to assess existing faithfulness metrics for natural language explanations (NLEs) from Large Language Models (LLMs). It adapts a previous test for faithfulness metrics, diagnosticity, which uses random feature attributions as unfaithful explanations. The new framework, causal diagnosticity, edits the LLM's knowledge in order to generate explanations that can be determined as unfaithful. Several existing metrics are evaluated on modified datasets across a range of tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper points out that many existing faithfulness metrics from [1] are flawed, which is useful extra evidence, though it is fundamentally unclear to me how this is going to lead to substantial later improvements.\n- While the datasets used are systematic enough to gain insights, a big fear is that they are overly simplistic. The examples given are for very short chain-of-thought responses to basic questions. Moreover, most of the faithfulness metrics in [1] are quite premature and have been criticized in the literature [2, 3]. It makes sense to me therefore that CC-SHAP might perform well, but the performance here does not really shed insight into the performance of the metric on more subtle scenarios (nor is it clear how well the evaluation framework is able to handle real-world tasks such as medical question answering).\n- These are my fundamental issues with the current setup, alongside relatively small/similar models being assessed. The faithfulness problem is due mostly to the fact that we do not have good ground truth of what faithful explanations are internally. This paper offers ground truths but in very simple settings.\n\n[1] Measuring Faithfulness in Chain-of-Thought Reasoning, Lanham et. al., 2023\n\n[2] Chain-of-Thought Unfaithfulness as Disguised Accuracy, Bentham et. al., 2024\n\n[3] On Measuring Faithfulness or Self-consistency of Natural Language Explanations, Parcalabescu and Frank, 2023" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Causal Lens for Evaluating Faithfulness Metrics},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yDICgRUj5s},\nnote={under review}\n}" }, "abstract": { "value": "The increasing capabilities of Large Language Models (LLMs) have made natural language explanations a promising alternative to traditional feature attribution methods for model interpretability. However, while these explanations may seem plausible, they can fail to reflect the model's underlying reasoning faithfully. The idea of faithfulness is critical for assessing the alignment between the explanation and the model's true decision-making mechanisms. Although several faithfulness metrics have been proposed, they lack a unified evaluation framework. To address this limitation, we introduce Causal Diagnosticity, a new evaluation framework for comparing faithfulness metrics in natural language explanations. Our framework extends the idea of diagnosticity to the faithfulness metrics for natural language explanations by using model editing to generate faithful and unfaithful explanation pairs. We introduce a benchmark consisting of three tasks: fact-checking, analogy, and object counting, and evaluate a diverse set of faithfulness metrics, including post-hoc explanation-based and chain-of-thought (CoT)-based methods. Our results show that while CC-SHAP significantly outperforms other metrics, there is substantial room for improvement. This work lays the foundation for future research in developing more faithful natural language explanations, highlighting the need for improved metrics and more reliable interpretability methods in LLMs." }, "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": [ "faithfulness", "diagonsticity", "natural language explanations", "interpretability", "model editing" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/cbe029e6bbba7195d53ece5de12df825bebbd5d4.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": null, "title": { "value": "A Causal Lens for Evaluating Faithfulness Metrics" }, "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 I would like to hear the authors discussi the related works on Hyperbolic spaces for dealing with hierarchical data and Hierarchical clustering, where data also have latent tree structure. The goal here is to compare different trees and assign a loss to each so that lower loss means better tree for the dataset. The approaches there are also differentiable so how do you compare with them?\n\n-For your metrics, is there any hope to prove something about the quality of the metric found?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+the paper studies a natural problem on hierarchies which is how to define suitable metrics that are differentiable and modular. \n\n+the authors present some natural candidate and apply it to different types of hierarchical data\n\n+the authors present experimental results showcasing properties of their proposed metrics and benefits over prior methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies hierarichically structured data and introduces a tree-distance with differentiable parameters weighting the importance of different subspaces. The paper presents experimental evidence that their approach achieves similar performance to SOTA methods based on neural networks while having orders of magnitude fewer parameters, and also has some benefits for heterogeneous Graph Neural Networks compared to prior methods.\n\nThe paper is motivated by the fact that there are many structured data formats such as JSON/XML/Protobuffer but not a good way of defining a reasonable notion of distance between them, which is in contrast with what happens when we deal with more standard objects like vectors in Euclidean space.\n\nThis paper proposes a particular distance called HTD distance, which exploits the recursive nature of the previously-mentioned data formats. The ultimate goal is to have a modular construction by combining potentially different metrics on different levels of the given tree. HTD has weight parameters, which control importance on different parts, is differentiable, and requires orders of magnitude fewer parameters than neural networks with similar guarantees (based on experiments). The authors perform a series of experiments with supervised learning, ianomaly detection, heterogenous GNNs, clustering and UMAP for visualization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-the theory is very straightforward in this paper. In fact, the two theorems stated as Th1 and Th2 could be obserations or propositions as they follow from the basic definitions.\n\n-there have been recently approaches to define differentiable objectives suitable for doing optimization over trees and hierarchies, especially to deal with problems on relational data coming from networks (e.g. facebook or other social networks) with the goal of performing hierarchical clustering. The first such works were 1) Nickel et al. \"Poincaré embeddings for learning hierarchical representations\" and later 2) \"Hyperbolic graph neural networks\" of Nickel et al. and later the works of 3) Chami et al. \"Hyperbolic graph convolutional neural networks\" and 4) \"From Trees to Continuous Embeddings and Back: Hyperbolic Hierarchical Clustering\" and of 5) Monath et al. \"Gradient-based hierarchical clustering using continuous representations of trees in hyperbolic space\" have dealt with similar questions. I am surprised the authors do not cite such works as the problem of optimization over trees was addressed using differentiable methods in all of these works.\n\n-omission of discussion for use of hyperbolic techniques and hyperbolic spaces in the present paper which is known to be suitable for hierarchical relations, much more than euclidean spaces." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "see weakness" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is overall well-written and easy to follow.\n2. This paper demonstrates theoretical superiority, as shown in Table 1." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the Hierarchically-structured Tree Distance (HTD), a novel metric for structured data in formats like JSON and XML. Designed for rooted heterogeneous trees, HTD is modular and differentiable, allowing it to adapt to tasks like classification, clustering, and anomaly detection. Experiments show that HTD-based algorithms perform competitively with neural network methods while using far fewer parameters and are more effective for analyzing heterogeneous Graph Neural Networks than the Tree Mover’s Distance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The authors demonstrate the effectiveness mainly on distance-based tasks, which shows better performance compare to other distance-based method but does not appear comparable to GNN classifiers.\n\n2. I am also concerned about the contribution and scope of this paper; however, I acknowledge that I am not an expert and am open to other opinions.\n\n3. Although some limitations are mentioned in the submission (e.g., in the caption of Table 4), there is no comprehensive discussion of the proposed method's limitations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the weaknesses above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper is well-written.\n- The proposed HTD generalizes the tree mover’s distance, making it applicable to both heterogeneous graphs and tree-structured data.\n- Extensive experiments conducted across multiple tasks—classification, clustering, and anomaly detection—show HTD's superiority over state-of-the-art methods for tree-structured data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the Hierarchically Structured Tree Distance (HTD), a metric designed to measure distances between tree-structured data commonly stored in formats like JSON and XML. HTD effectively represents message passing in heterogeneous graph neural networks (GNNs). Experimental results show that this distance metric is capable of addressing various machine learning tasks, including classification, visualization, and clustering." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have several concerns regarding the novelty of the proposed distance:\n- HTD appears to be a straightforward extension of the tree mover’s distance, replacing the optimal transport (OT) distance with the Hausdorff and Chamfer distances, which may introduce cheaper computational complexity.\n- It is unclear why HTD outperforms the tree mover’s distance on homogeneous graph datasets, such as MUTAG and BZR, as shown in Table 3.\n- Even on heterogeneous datasets, it seems feasible to apply the tree mover’s distance by constructing separate computational trees for each node type in the graph. So we can improve the performance of tree mover's distance on heterogeneous datasets like MUTAG and BZR.\n- Similar to the tree mover’s distance, HTD does not meet the criteria for defining a valid kernel for tree-structured data, as it is not conditionally negative definite.\n- Including standard deviations (STD) in the results of Tables 3 and 4 would be beneficial, as the variances are large; for instance, the STD of accuracy values for MUTAG and BZR might be around 5." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Only until Figure 2 the authors demonstrate examples of Schema. However, the sample 1 and sample 2 are essentially trees, it still remains unclear how HS-Trees to GNN in Section 2.1. Could the author provide an illustrative figure to show the relation?\n- The font size is a bit bigger than usual?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors introduce a new distance for hierarchically-structured trees based on Leaves, Bags, and Dicts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces hierarchically-structured tree distance (HTD) between HS-Trees." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The motivation and explanation of the use of HS-trees and their distance are unclear in the introduction. The authors first say that the properties of HT-Trees are used in existing work. Then, in the next paragraph, the authors mention that the distance on HS-Trees has been studied very little. Among the properties used in previous work, were there HS-Tress distance?\n- In the introduction and background, it’s unclear if the term HS-Trees is from previous work or if the term is first given by the authors. In addition, the background of HS-Trees, sample, and schema are very confusing. It would be clearer if the authors could provide an example of what they are here. Moreover, it’s unclear how bags are assumed to be “permutation invariant; therefore, the position has to be encoded through position encoding” and how the “universal approximation theorem for HS-Trees has been proved in Pevny & Kovarik (2019).”\n- The authors mention the motivation of the work is “measuring the distance between samples emerging from popular data storage formats (e.g., JSON, XML, and ProtoBuffer)”. However, in the experiment section, there is no application in such data format. It is misleading to use the data storage formats as motivation in the introduction, but there is no related experiment\n- In Section 2.1, it is unclear if the authors are trying to claim that GNNs are “hierarchically-structure data”. In the context of GNN, it’s hard to see what sample, schema, and even HS-Trees are.\n- The authors claim that the proposed distance is differentiable, however, there is no theoretical proof\n- The term Leaves is commonly used in tree structure data. It will be important to differentiate the difference between common tree structures and HS trees. In addition, it’s unclear how and why the definition of Leaves in Section 3. 1 is outside of the scope\n- There is no proof for Theorem 1. The justification right after Theorem 1 is hard to follow.\n- It is hard to link the relation between Eq. (5) and the distance in Table 2, even with the brief introduction in Appendix B. A simple proof or derivation could have helped to understand the connection.\n- It’s unclear what the relation between HTD and TMD is. A simple proof or derivation could have helped to understand the connection.\n- The README.md files in the provided link to the code are not sufficient to reproduce the HTD and experimental results.\n- The experimental setup is unclear in the main texts. It’s unclear what the actual classification tasks and anomaly detection are. Also, while Appendix A includes the implementation details, there is no reference in the main text link to the appendix. In addition, in Appendix A, the authors claim that the experiments are repeated five times, and there is no variance or standard deviation reported in the performance in the main text.\n- It’s unclear what the colors represent in Figure 3. It’s unclear why and what “—” represents in Tables 3-5.\n- The paper needs more proofreading: i) additional ) in line 029, ii) line 250 missing a period and there is an additional ), iii) notation is very hard to follow; a notation is given with confusing comma\n\n## Minor\n- There if no reference and introduction when the Mutagenesis dataset is first mentioned in line 317" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024differentiable,\ntitle={Differentiable Distance Between Hierarchically-Structured Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yDlvteYBbF},\nnote={under review}\n}" }, "abstract": { "value": "Many machine learning algorithms solving various problems are available for\nmetric spaces. While there are plenty of distances for vector spaces, much\nless exists for structured data (rooted heterogeneous trees) stored in popular\nformats like JSON, XML, ProtoBuffer, MessagePack, etc. This paper\nintroduces the Hierarchically-structured Tree Distance (HTD) designed\nespecially for these data. The HTD distance is modular with differentiable\nparameters weighting the importance of different sub-spaces. This allows\nthe distance to be tailored to a given dataset and task, such as classification,\nclustering, and anomaly detection. The extensive experimental comparison\nshows that distance-based algorithms with the proposed HTD distance\nare competitive to state-of-the-art methods based on neural networks with\norders of magnitude more parameters. Furthermore, we show that HTD is\nmore suited to analyze heterogeneous Graph Neural Networks than Tree\nMover’s Distance." }, "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": [ "Distance", "Distance function", "Tree-structured data", "Heterogenous Graphs", "JSONs", "Multiple Instance 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/66137360ff65a367c9b78659e3cabfa7a4f37f96.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Differentiable Distance Between Hierarchically-Structured Data" }, "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": 1 }, "primary_area": null, "questions": { "value": "This has already been addressed in the Weaknesses section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The overall research motivation is well-articulated, and the proposed idea is both interesting and has promising applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The main contributions of this work, as reported, are:\n\nA. Development of a VeriDistill framework: The first truly end-to-end machine learning model that processes raw Verilog code directly, without preprocessing, to accurately estimate circuit area and delay metrics.\n\nB. Innovative Knowledge Distillation (KD): A novel knowledge distillation method is applied during training, transferring low-level circuit insights (as LUT graphs) back into the model for enhanced predictions. Experiments show that this approach surpasses previous state-of-the-art (SOTA) baselines on a large-scale Verilog dataset, with improved generalization to out-of-distribution (OOD) data. The use of both LLM representations and knowledge distillation is critical to the model’s performance, as omitting either component decreases performance below baseline levels." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The manuscript lacks clear description of the proposed approach. Both figure 1 and figure 2 should be improved in both schematic and description. Methodology section 3.1 must be elaborated. I will advice to break the training strategies and steps at first and finally combine them towards final goal/results. It will enhance readability, understanding of the general audience with more detailed explanation of the intermediate training flows. Figure 1 & 2 and section 3.3 description seems disjointed. Reproducibility from this text description is problematic.\n\n2. in Section 4.1 EXPERIMENTAL SETUP, \"The model’s decoder feedforward network depicted\nin Fig. ?? is designed so that the 512-dimensional representations from the...\", Fig. reference is missing.\n\n3. The overall quality of writing and presentation could be improved. The writing should maintain a smoother flow, as some paragraphs feel disjointed. For instance, in the experimental setup section 4.1 , certain details—such as the explanation of 512-dimensional state tokens—are important for understanding the methodology and would benefit from clearer integration.\n\n4. It is recommended to compare the performance of the proposed model with previous models/approaches to better emphasize its novelty, particularly in terms of computational complexity, data preprocessing demands, resource utilization, and the metrics of area and delay. For instance, could you provide an analytical comparison of your approach against the work by Fang et al. (2023/2024b), addressing the computational time for SOG and the complexity of converting linguistic data into bit-level operators using the logic synthesis tool Yosys (Wolf et al., 2013)? This comparison would further clarify the advantages of your approach, especially given that accuracy, precision, and sensitivity are crucial metrics for circuit quality estimation in addition to computational efficiency.\n\n5. In Fig. 4, please clarify the presence of scattered outlier points for both area (lower values) and delay (higher values) in relation to the fitting curve. These sparse data points appear to indicate cases where VeriDistill’s performance aligns with that of other baseline models. Could you explain why VeriDistill performs similarly to the baselines for these specific outliers, and discuss any factors that might contribute to this behavior in terms of model limitations or specific data characteristics?\n\n6. Authors explained \"Finally, in Figure 5, we present the t-SNE projection of the last hidden space representations on the test data from the teacher model (Z_teacher) trained for predicting log-area, alongside those from the LLM-based models. As can be seen, the resulting t-SNE representation of the VerDistill model appears very similar to the one of the LUT-GNN teacher model, albeit slightly rotated by roughly 30 degrees to the right (which is arbitrary and immaterial). The AST-GNN with KD model also can be seen as a rotation of the LUT-GNN plot, roughly 90 degrees to the left. In contrast, the plot of the CodeV + Decoder appears much more like an undefined mass.\" \n\nIt seems hypothetical assumption about the statement \"VerDistill model appears very similar to the one of the LUT-GNN teacher model, albeit slightly rotated by roughly 30 degrees to the right (which is arbitrary and immaterial). The AST-GNN with KD model also can be seen as a rotation of the LUT-GNN plot, roughly 90 degrees to the left.\". AST-GNN with KD is quite different than LUT-GNN and VeriDistill, as well it does not seem \"t-SNE representation of the VerDistill model appears very similar to the one of the LUT-GNN teacher model\"...and if the authors knew already \"slightly rotated by roughly 30 degrees to the right\".. please adjust this rotation to make it fit overlapped and compared.\n\"\"\n7. Please be consistent with proposed approach name .......authors mentioned \"VerDistill\" in one place and \"VeriDistill\" in other place(s). Please correct any typos. Also, grammatical errors have been noticed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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) Does the knowledge distillation help with a specific subset of Verilog designs (more than others)? Could be interesting to exploit this.\n2) Please elaborate on the training times\n3) There are many ways to code (e.g an adder) in Verilog, does the framework handle this well? \n4) Curious to hear whether you think this approach can scale to handle other RTL (e.g VHDL)" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "As a hardware engineer, I greatly appreciate the contributions in the work as this can provide a lot of benefit to practioners.\n\nI find the knowledge distillation approach between the GNNs and LLMs to be novel. The work showcases strengths of GNNs and LLMs in a creative fashion. The work also shows that LLMs have remarkably learned about the underlying circuit representation underneath the verilog code.\n\nThe results are also impressive, showing improvements over baselines including the out-of-distribution datasets like OpenABCD." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Introduction of a new ML framework: VeraDistill to predict circuit QoR metrics like area/delay bypassing costly logic synthesis processes. The framework uses a verilog-trained LLM (codeV-7B) to produce a representation (final layer) to feed into the FFN to make predictions for the QoR metrics during inference. To train, they take the AIG LUT outputs of the synthesis flow and use GNNs to produce embeddings where the GNN acts as a \"teacher\" resulting in knowledge distillation. The loss metric used for final training is a simple MSE involving not only the FFN predictions and synthesis QoR data, but also the final layer data from the GCN(teacher) and LLM (Student). The LLM is frozen in the training process." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Its unclear to me how reliant the framework is on a particular LLM (CodeV) and how versatile a single instance (trained) will be in accommodating more unseen RTL code. The authors do acknowledge the lack of Verilog/RTL code for training, so while not a weakness per se, it does make me question the generalization capabilities.\n\nNot much details are shared about the training resources (outside of the 8 V100 GPUs used)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. For Table 1, you state that knowledge distillation has almost no effect on the AST-GNN model. But in Figure 5, you show that the knowledge distillation makes the t-SNE representation of AST-GNN more like that of the teacher model. These statements seem contradictory and confusing. Could you provide further clarification?\n2. For Table 2, could you also report the performance of the teacher model on the out-of-distribution dataset? This would help clarify the current gap in performance.\n3. It seems VeriDistill can only handle small Verilog designs due to the context window length constraint of the LLM base model. Is there a way to scale VeriDistill for larger designs? Given that logic synthesis for small designs is usually fast and the error remains significant, the current model may not be very attractive.\n4. When the logic synthesis recipe changes, the delay and area metrics of circuits typically change as well. How can VeriDistill handle these variations?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This work introduces VeriDistill, an innovative end-to-end machine learning model that predicts circuit quality directly from raw Verilog code, which is an important and challenging task in the domain of electronic design automation (EDA). It applies a novel knowledge distillation method, transferring low-level circuit insights via LUT graphs into an LLM-based predictor. And VeriDistill demonstrates robust performance, outperforming state-of-the-art baselines on large-scale and out-of-distribution datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents VeriDistill, the first end-to-end machine learning model that processes raw Verilog code to predict circuit quality metrics. The model uses an innovative knowledge distillation technique, transferring low-level circuit insights via LUT graphs into an LLM-based predictor. Experiments show that VeriDistill surpasses current state-of-the-art methods on large-scale Verilog datasets and demonstrates transferability to out-of-distribution datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are several limitations for VeriDistill. Firstly, as shown in Table 1, VeriDistill’s performance lags significantly behind the teacher model. To improve this, I think a possible way is to unfreeze some parameters in the base LLM and fine-tune them with the decoder. Secondly, the paper only uses CodeV as its base model without exploring other Verilog LLMs that might provide different insights. Conducting more experiments with various models could better demonstrate the method's effectiveness and robustness. Additionally, the paper needs careful proofreading; for instance, there is a missing figure referenced in Line 293." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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.Given the token count limitations of LLMs, which may prevent the framework from handling large Verilog designs, and the relatively quick and low-cost process of transforming smaller designs into netlists, wouldn’t it be more effective to use a GNN to process the netlist representation directly?\n2.Why transform aiger into LUT? What’s the advantage of LUT? And in Advanced integrated circuit (IC) manufacturing process, whether LUT is widely used?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper is the first to implement knowledge transfer of circuit quality estimation from netlist-level to RTL-level. Considering that netlist-level circuit performance and area estimation is more precise and RTL-level circuit quality prediction is faster, this work strike a balance between the precision and inference speed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework, which utilize a GNN trained on LUT netlist-level graph as a teacher in order to transfer knowledge to LLM encoder based circuit quality estimator. It enables model to get more accurate performance and area of circuits end-to-end with RTL as input." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "In this work, the LLM is used as an encoder, leveraging its embeddings to predict circuit quality. However, a potential limitation arises from the token count restrictions of LLMs, which may prevent the framework from handling large Verilog designs. For smaller designs, which can be processed within these token constraints, transforming them into netlists is relatively quick and incurs minimal computational cost. Given this, why not employ a GNN to process the netlist representation directly? Predictions at the netlist level are generally more accurate than those at the RTL level, making this a potentially more effective approach. Moreover, the authors mainly compare there work with models trained on AST. Such an experiment lacks persuasive power.\nThings to improve: The ‘Fig. ??’ in section 4.1." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024the,\ntitle={The Graph's Apprentice: Teaching an {LLM} Low-Level Knowledge for Circuit Quality Estimation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yDy9fZXNJV},\nnote={under review}\n}" }, "abstract": { "value": "Logic synthesis is a crucial phase in the circuit design process, responsible for transforming hardware description language (HDL) designs into optimized netlists. However, traditional logic synthesis methods are computationally intensive, restricting their iterative use in refining chip designs. Recent advancements in large language models (LLMs), particularly those fine-tuned on programming languages, present a promising alternative. This work proposes augmenting LLMs with predictor networks trained to estimate circuit quality directly from HDL code. To enhance performance, the model is regularized using embeddings from graph neural networks (GNNs) trained on Look-Up Table (LUT) graphs, thereby incorporating lower-level circuit insights. The proposed method demonstrates superior performance compared to existing graph-based RTL-level estimation techniques on established benchmarks, such as OpenCores and OpenABCD, while providing instant feedback on HDL code quality." }, "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", "Knowledge Distillation", "Verilog", "Graph Neural Network" ] }, "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/64c6dce5011e7098b9ee10b08b7343d0dd636511.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": "The Graph's Apprentice: Teaching an LLM Low-Level Knowledge for Circuit Quality Estimation" }, "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": { "value": "At the review stage, I could easily find the corresponding materials of this submission : (1) GitHub repository is publicly available at https://github.com/TencentARC/SEED-Story, and (2) arxiv paper at https://arxiv.org/ab/s2407.08683.\n\nEven it seems rather popular these days, \nuser double-blind review, I think that there is a possibility of making effects on reviewers biased in any forms." }, "flag_for_ethics_review": { "value": [ "Yes, Other reasons (please specify below)" ] }, "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 weaknesses" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The samples of generated outputs demonstrated in the manuscript look impressive, especially in long story generation, character consistency, and narrative text quality. \n This work would be significant when these examples objectively represent general performance.\n\n2. The introduction of the StoryStream dataset is a valuable contribution that could advance future research in this field.\n The authors present StoryStream Dataset, which is already downloadable at https://huggingface.co/datasets/TencentARC/StoryStream." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a method, SEED-Story, for the task of image-text multimodal story generation.\nThey design architectures that tokenize images for tuning large language models (LLMs) and detokenize them for image generation. They also employ a multimodal attention sink to facilitate long story generation. \nThey evaluate experimentally the quality of generated images and assess story quality with a baseline method, MM-interleaved. \nAdditionally, a novel dataset, StoryStream, is introduced." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper lacks essential elements such as a clear motivation and problem definition, making it difficult to judge whether the research goal is successfully achieved. \n Also, those make readers confused about whether the evaluation metrics and comparative methods are appropriate or sufficient.\n\n2. In order to validate academic value, the manuscript should provide evidence theoretically or experimentally. \n The manuscript shows the quality of generated images with FID and CLIP score as an ablation study, and relative comparison story quality with a baseline method, MM-interleaved. \n I think the materials above are NOT enough to clarify the validity of improvement to keep fair comparison and reproducibility. \n \nThe results shown are promising, but clearer motivation, explicit research questions, and rigorous comparative experiments would strengthen the paper. The contribution of a new dataset is notable, yet the lack of comprehensive evaluation limits the ability to objectively assess the work's significance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- How may participants were involved in user study presented in Section A of Appendix?\n- My understanding is dense attention uses attentions between all tokens, and the proposed multimodal attention sink mechanism uses a subset of the tokens. So I think dense attention may perform (almost) equally well as multimodal attention sink mechanism even though it is more expensive. Can you clarify this?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The proposed dataset has high-quality and high-resolution images with longer story length, which would be useful for research community.\n- The proposed model seems to make sense, and the presented analysis for multimodal attention sink mechanism is nice.\n- The experimental results show that the proposed approach outperforms the baseline on the proposed dataset and demonstrate its effectiveness even though it still has a limitation (please see the weaknesses below)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the task of multimodal story generation which generates both story text and corresponding images from give input image and text, by proposing SEED-Story model utilizing Multimodal Large Language Model. Additionally, authors extend attention sink mechanism and introduce a multimodal attention sink mechanism to enable the model to generate long sequences in an efficient manner. Finally, a new dataset called StoryStream was proposed for training and benchmarking the task of multimodal story generation with longer story length. In experiments, it has been shown that the proposed approach outperforms the baseline, MM-interleaved (Tian et al. (2024)), in both quantitative and qualitative evaluations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The novelty of proposed approach seems somewhat incremental as it looks to share some ideas with MM-interleaved (Tian et al. (2024)). Also the proposed multimodal attention sink mechanism seems just a slight modification of existing attention sink mechanism even though it includes nice analysis.\n- The proposed approach was only evaluated on the proposed StoryStream dataset. However, as MM-interleaved (Tian et al. (2024)) was also evaluated on both Pororo and Flintstones datasets in their paper, it would make the paper stronger with additional evaluations on more datasets.\n- Another major concern is baseline choice. The authors just chose MM-interleaved, but I think it is still possible to use more natural baselines. e.g., generating story text first and then generate corresponding images with existing story visualization approaches.\n- The experimental result is not so convincing. e.g., in Fig. 7 (bottom), in the generated image sequence, the hat appeared and disappeared, and the color of hat was changed, so its consistency seems still not very good." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. For Figure 8, at what size sequence length does the sliding window approach stop working? I would also like to know for Table 2 what sequence length this is? \n\nIt would be interesting to show the FID/clip score at different sequence lengths. E.g. for seq length = 10 I expect all the methods are equivalent, but I am not sure how the numbers compare with seq length = 20, 30, 40, 50 . Does the multimodal attention sink only help when the test seq length is much larger than the training? \n\n2. For section 5.1 please clarify the differences between MM-interleaved and seed-story (e.g. architecture/base models)\n\n3. Does the mm attention sink affect the quality of the text? Figure 8 / Table 2 are mostly concerned with the quality of the images" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The strongest contribution of the paper is the new dataset (StoryStream). There seems to be a need for this new dataset, given that the existing datasets look quite low-resolution and simple. It will be useful for other papers to use this dataset to train and evaluate.\n\nThe major technical contribution of the paper is the multimodal attention sink. This is quite an interesting non-obvious observation that the model attributes high attention to tokens near BoI / EoI. Then the authors take this observation and propose a multimodal attention sink which seems to be useful for generating longer stories. This contribution could be useful in other problems outside of story generation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present a new method for story generation (seed-story), including the generation of images along with text. The model to generate the stories is based on a multimodal LLM. Images are tokenized with a ViT based tokenizer. Then images can be fed into an LLM to predict the next tokens, which could be images or text. A detokenizer, based on SDXL is then used to generate viewable images from the tokens. The image detokenizer and the LLM are finetuned for this specific task, with LORA being used for the LLM.\n\nThe authors introduce a new dataset, StoryStream, which consists of interleaved image/text stories to train their model. To create the dataset, the authors first sample keyframes and corresponding subtitles from children's cartoons. Then a caption model is used to caption frames. Finally GPT-4 is used to generate a consistent story given the subtitles and captions. The final dataset is larger in terms of story length and higher resolution compared with existing datasets.\n\nSeed-story also attempts to generate longer stories. Long stories are difficult to generate given the limited training data. Therefore the authors use a \"train-short-test-long\" approach to generate longer videos. The authors introduce a variation on attention-sink for multimodal data. Attention sink uses a sliding attention window plus a few initial tokens at the beginning of the sequence. The proposed multimodal attention sink also includes the tokens corresponding to the beginning and end of the image tokens. It is shown that this approach generates much higher quality story images.\n\nSeed-story is primarily compared with MM-interleaved, another approach that can be used to generate stories. Compared with MM-interleaved seed-story achieves lower FID, better style consistency, story engagement, and image-text coherence." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The long generated stories don't seem to be very good in terms of storytelling. For example, Figure F does not really suggest any logical story (e.g. with a beginning, middle, end). It seems more like random captions were generated, especially near the end. The llm also starts repeating itself often (e.g. \"A man in a hat and shirt looked surprised\", \"George, the cartoon monkey, stood in a grassy area\". The authors in the paper say that this plot is \"engaging\", but IMO it is not engaging. \n\nThis is a difficult task, but a simple baseline is to have a text only LLM generate a curious george story. The story in Figure F looks quite poor compared to a pure text baseline. Another simple baseline would be to have a text-llm generate a story and then have a diffusion model generate an image for each paragraph. This might lack consistency between images, but overall might be better, especially if rated by humans. It feels like there is potential in this paper but I feel more iteration is needed to generate reasonable stories.\n\nI'm also wondering how good the StoryStream dataset actually is for this task. From figure 5 it actually doesn't seem to have a very coherent story either, which might be why the trained model struggles. Figure 5 looks like a bunch of image captions with very little narrative storytelling. To be fair, it does seem that the other existing datasets have the same issue.\n\nThe main comparison with previous works is with MM-interleaved, but I am concerned about the fairness of these comparisons. MM-interleaved uses stable-diffusion-2-1 and Seed-story use SDXL. SDXL should be much better than stable-diffusion-2-1, and it is possible that the results in Figure 6a,b,d are strongly affected by this. Also, MM-interleaved use ViT-L and seed-story uses ViT-G as the image encoder which might also affect these numbers. MM-interleaved uses Vicuna-13b and seed-story uses llama2 which could affect Figure 6c. The authors do not mention these important differences between MM-interleaved and seed-story in the paper.\n\nNitpicks:\nReferences need to be cleaned up, there are multiple cases of duplicate references. E.g. for Qwen-vl, \"Improved baselines with visual instruction\ntuning\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- How do ensure the quality of the proposed dataset, and what are the criteria measured if any?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The tokenizer and de-tokenizer training coupled with multimodal instruction finetuning is neat.\n- The adaptation stage seems useful.\n- The extended multimodal attention sink mechanism is interesting and should be studied more." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work tackles the challenging problem of multimodal story generation – generating coherent textual stories with corresponding images interleaved with them. The work proposes a three-stage procedure, where the first stage trains an image de-tokenizer to regress towards target image features, while the second stage unifies the visual tokenizer and detokenizer with a multimodal LLM to generate simultaneously the image and textual features. Then, the trained model will undergo a final de-tokenizer adaptation step where the diffusion loss will be used to refine the outputs in their pixel quality. The authors also extend the attention sink mechanism with “focused” tokens on several special positions to preserve story length and consistency.\nThe work features additionally a collected multimodal story generation dataset, which is also used to evaluate the above proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the proposed method is evaluated well on the featured StoryStream dataset, the method should still be evaluated on existing ones such as Pororo and Flint Stones.\n- While I appreciate the proposed dataset, the visual consistency test is still lacking and a well defined metric particularly for visual consistency is much needed.\n- The attention sink mechanism is not well-studied yet in this manuscript, for example, how does the performance with its introduction scale with lengths, story complexities, and character consistencies?\n- Some human evaluation is needed for the results comparisons." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 3 }, "primary_area": null, "questions": { "value": "+ Why attention sink always choose the beginning of the text/image tokens, what is the insight behind this. How about long sequence and the current generation is not quite relevant to the beginning.\n+ The evaluation metric is not very comprehensive for the story generation setting. It only shows FID and CLIP score. First of all, CLIP score is not very reliable for identifying the fine-grained vision-language similarity. Except for the human evaluation, is there any metric for demonstrating the image consistency and character accuracy?\n+ The paper targets on multimodal story generation setting. How to demonstrate the text generation follows a coherent and logical storytelling progression?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "+ The paper curated a high-resolution long story dataset StoryStream that contains 257k images with story length up to 30 frames.\n+ It proposes a multimodal attention sink mechanism for efficient long-sequence generation.\n+ The qualitative results are impressive for generating long multimodal stories." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presented SEED-Story, aiming to generate long, coherent stories combining narrative texts with visually consistent images, addressing challenges in multimodal story generation. It proposes a multimodal attention sink mechanism for efficient long-sequence generation, and release the StoryStream dataset to benchmark the model. Through comparison and user studies, the authors claim SEED-Story outperforms existing models in terms of coherence, visual quality, and style consistency in multimodal storytelling." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "+ The model architecture is quite similar SEED, with proposed multimodal attention sink.\n+ The evaluation is not very comprehensive, giving existing literature also evaluates image consistency and character accuracy in the story generation setting." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nyang2024seedstory,\ntitle={{SEED}-Story: Multimodal Long Story Generation with Large Language Model},\nauthor={Shuai Yang and Yuying Ge and Yang LI and Yukang Chen and Yixiao Ge and Ying Shan and Ying-Cong Chen},\nyear={2024},\nurl={https://openreview.net/forum?id=yEPNPbF8E7}\n}" }, "abstract": { "value": "With the remarkable advancements in image generation and open-form text generation, the creation of interleaved image-text content has become an increasingly intriguing field. Multimodal story generation, characterized by producing narrative texts and vivid images in an interleaved manner, has emerged as a valuable and practical task with broad applications. However, this task poses significant challenges, as it necessitates the comprehension of the complex interplay between texts and images, and the ability to generate long sequences of coherent, contextually relevant texts and visuals. In this work, we propose SEED-Story, a novel method that leverages a Multimodal Large Language Model (MLLM) to generate extended multimodal stories. Our model, built upon the powerful comprehension capability of MLLM, predicts text tokens as well as visual tokens, which are subsequently processed with an adapted visual de-tokenizer to produce images with consistent characters and styles. We further propose multimodal attention sink mechanism to enable the generation of stories with up to 25 sequences (only 10 for training) in a highly efficient autoregressive manner. Additionally, we present a large-scale and high-resolution dataset named StoryStream for training our model and quantitatively evaluating the task of multimodal story generation in various aspects. \nAll models, codes and datasets are released in https://anonymous.4open.science/r/SEED-Story." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Shuai_Yang7", "~Yuying_Ge2", "~Yang_LI82", "~Yukang_Chen1", "~Yixiao_Ge2", "~Ying_Shan2", "~Ying-Cong_Chen1" ] }, "authors": { "value": [ "Shuai Yang", "Yuying Ge", "Yang LI", "Yukang Chen", "Yixiao Ge", "Ying Shan", "Ying-Cong Chen" ] }, "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", "Story telling", "multi-modal generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "yang|seedstory_multimodal_long_story_generation_with_large_language_model" }, "pdf": { "value": "/pdf/21bf8ccf6d62f46403c12cf7f9d54386d665569a.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/fcd4359397ddf2eac3ffe411a8d312171ff0b725.zip" }, "title": { "value": "SEED-Story: Multimodal Long Story Generation with Large Language Model" }, "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": 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. Does the order of class pairs impact the iterative process for finding optimal weights?\n2. CWPLUGIN currently does not take class frequency or class imbalance in the validation dataset into account. Could incorporating these factors enhance performance, particularly in label-shift scenarios?\n3. Why was BERT-FT chosen as a baseline for comparison instead of fine-tuning other used black-box models like DistilBERT, RoBERTa, or DistilRoBERTa? For example, section 4.2.1, black-box model is DistilBERT, but we are comparing to BERT-FT." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "**Originality:** While CWPLUGIN shares the concept of post-processing prediction probabilities through reweighting, it is uniquely motivated by the need to adapt black-box models to specific target domains, providing a fresh perspective on model adaptation for restricted-access models.\n\n**Quality:** CWPLUGIN demonstrates strong performance across a range of scenarios without requiring large amounts of validation data. And its design can be adapted to many settings. Their experiment shows promising results.\n\n**Clarity:** The motivation for CWPLUGIN is clear to me. However, the algorithmic section could benefit from further clarification (see below).\n\n**Significance:** This paper addresses a critical and emerging challenge in the field: adapting black-box models for task-specific needs. The proposed method shows effectiveness and efficiency by adapting predictions through simple probability reweighting, providing a practical solution for the community’s growing demand for adaptable AI tools." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Black-box or proprietary models often limit users to accessing only predictions via API calls, making it difficult to adapt the model behaviers to align with specific user preferences. This paper introduces CWPLUGIN, a plug-and-play method that reweights prediction probabilities to match users' desired metrics or target domains. CWPLUGIN uses a coordinate-wise search algorithm that iteratively finds optimal relative weights for each class pair to achieve the desired reweighting. Empirical results show that CWPLUGIN outperforms existing calibration methods and fine-tuning approaches, especially when data size is limited. Additionally, CWPLUGIN proves effective in handling scenarios with label shift and label noise. This framework offers new opportunities for users to adapt black-box models to their tasks, allowing for a degree of customization previously unavailable." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Algorithm 1:** The CWPLUGIN algorithm iterates over each class pair; however, the notation used is unclear. The symbol m denotes the number of classes but is also used as a class index, which creates confusion. Additionally, in line 3 of the algorithm, a single for-loop is shown, which does not accurately reflect the process of iterating over class pairs. Using a nested for-loop structure would better clarify this.\n\n**Class Size:** The experiments were conducted on datasets with a relatively small number of classes. It would be beneficial to include results on datasets with a larger number of classes, such as adapting to specific metrics on CIFAR-100, TinyImageNet, ImageNet, to demonstrate the scalability of CWPLUGIN when m is high.\n\n**API Call Expense for Validation:** The paper could discuss how the cost of API calls increases with the size of the validation dataset, which may be a limitation in scenarios where extensive validation data is needed to reach desired performance.\n\n**Suggested References:** Two relevant papers are suggested to read and include:\n* The first paper leverages relational information in label space to reweight prediction probabilities without requiring validation datasets: https://arxiv.org/abs/2307.12226.\n* The second paper adapts black-box models by steering model outputs using a combination of tuned and untuned models: https://arxiv.org/abs/2401.08565." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "**Clarification regarding Section 4.1 setting**: In Section 4.1, could you elaborate on how the regression model is used for classification tasks? What is the process for converting regression outputs into discrete class labels? Are there predefined thresholds or ranges used to map continuous outputs to specific classes?\n\n**Clarification regarding section 4.2.1 setting**: distilBERTbased is used a the base model while a finetuned version of BERT is used as a baseline. What is the reason behind choosing a different model as a basline? Would it be more appropriate to use a fine-tuned version of distilBERT (distilBERT-FT) as the baseline for a more direct comparison? Alternatively, why not use BERT as the base black-box model to maintain consistency with the baseline (perhaps a more apples-to-apples comparison)? How might these different choices impact the interpretation of the results and the conclusions drawn from the comparison?\n\n**Clarification regarding section 4.2.2 setting**: Why test with different base models for lmemotions and lmemotionsOOD.? how about using the same base model across all tasks? If different base models are necessary, could an additional ablation study be conducted using a single base model across all tasks to isolate the effect of the model choice?\n\n**Clarification regarding section 4.3 setting**: The authors mention on lines 484-485 that ‘For both the label shift and label noise settings, we utilize a model trained on GLUE (Wang et al., 2019) and ANLI as our base, black-box predictor’ but did not mention which model used. What model/LLM is used for this task?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The method is computationally very efficient as it solely relies on output probabilities of the black-box model. \n\nThe authors also propose ways to further save costs when data is balanced across classes by using parallelization or when the metric it is optimizing has a specific quasi-concave shape; these are interesting observations.\n\nThe method is also sample efficient, requiring only a few additional samples to optimize the weights w. \n\nThe paper is well-written and the presentation is good overall." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the author propose CWPLUGIN, a post-hoc method that adapts black-box machine learning models to new target distributions and optimizing specific performance metrics without requiring access to the model's internals (e.g. training details, metric optimizes, hyperparameters used, etc.). CWPLUGIN is a post-hoc method as it takes in as input the set of probabilistic multiclass predictions (e.g. softmax outputs) on a target domain and their corresponding true labels, thereby disregarding any feature information. CWPLUGIN optimizes metrics that are simple functions of the confusion matrix and outputs new class weights, one for each class. The method can be made efficient when the data is balanced across classes or when the metric it is optimizing has a specific quasi-concave shape." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Limited novelty**: The authors mention the “probing classifier” approach Hiranandani et al. (2021) that solves a, global linear system in order to find the weights which optimize a particular metric. CWPLUGIN is instead local in that it considers only pair-wise comparisons between classes. Is this the only difference between 'probing classifier' Hiranandani et al. (2021) and CWPLUGIN? \nIt is unclear if this difference constitutes a significant advancement, especially since the probing classifier already performs very well (e.g. in Figure 6, Probing in fact outperforms in terms of MCC and G-mean. Similarly, Figure 2 shows only marginal improvements in F-measure and accuracy compared to other methods, especially the probing classifier).\n\n**Key literature missing**: The paper could benefit from citing some very important studies in the calibration literature that are relevant to the discussion, including:\n\n1)\tZihao Zhao, Eric Wallace, Shi Feng, Dan Klein, and Sameer Singh. Calibrate before use: Improving few-shot performance of language models. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th ICML 2021.\n\n2)\tAbbas, M., Zhou, Y., Ram, P., Baracaldo, N., Samulowitz, H., Salonidis, T., and Chen, T. Enhancing in-context learning via linear probe calibration. In Proceedings of The 27th AISTATS 2024.\n3)\tZhixiong Han, Yaru Hao, Li Dong, Yutao Sun, and Furu Wei. Prototypical calibration for few-shot learning of language models. ICLR 2023.\n\n4)\tHan Zhou, Xingchen Wan, Lev Proleev, Diana Mincu, Jilin Chen, Katherine Heller, and Subhrajit Roy. Batch calibration: Rethinking calibration for in-context learning and prompt engineering. ICLR 2024.\n\n5)\tZhongtao Jiang, Yuanzhe Zhang, Cao Liu, Jun Zhao, and Kang Liu. Generative calibration for in-context learning. In Houda Bouamor, Juan Pino, and Kalika Bali (eds.), Findings of EMNLP 2023.\n\n6)\tM. Shen, S. Das, K. Greenewald, P. Sattigeri, G. Wornell, and S. Ghosh. Thermometer: Towards universal calibration for large language models. ICML 2024.\n\nThese papers focus on calibration for LLMs and could also serve as valuable baselines, especially because the authors do studies on language tasks. Including them could strengthen the paper's comparative analysis and contextualize its contributions within the broader field of language model calibration. \n\nNotably, Zhao et al. (2021) and Abbas et al. (2024) are especially pertinent, since these are also very simple post-hoc methods that use a handful of samples in the range comparable to the one used by the authors (i.e. on the order of tens or hundreds of examples). While these papers don't directly address the specific metrics used by the authors, they demonstrate high utility in terms of accuracy. It would be valuable to evaluate their performance on the metrics employed in this study, such as F-measure, MCC, and G-mean. Comparing CWPLUGIN against these methods would provide a more comprehensive assessment of its effectiveness.\n\n\n**Marginal improvement over baselines**: Figure 2 shows only marginal improvements in F-measure (e.g. 0.579 vs 0.576) and accuracy (e.g. 0.619 vs 0.617) compared to other methods, especially the probing classifier. Moreover, in Figure 6, Probing in fact clearly outperforms CWPLUGIN in terms of MCC and G-mean.\n\nSimilarly, in Figure 3, on Language tasks, we observe marginal improvement over FullDirich baseline on most metrics except F-measure (e.g. 0.563 vs 0.562 on accuracy, 0.256 vs 0.255 on MCC).\t\n\nCompared to the baselines, CWPLUGIN is the least performing method in terms of accuracy and MCC for lmemotions (figure 8) and one of the least performing methods on lmemotionsOOD (Figure 9). The results are perhaps good for applications where different evaluation metrics (e.g. F-measure) are more critical than mere predictive accuracy. However, I think even improvements on these metrics are marginal in most cases.\n\n**Figure 5 table ANLI with label noise results for G-mean do not match results in Figure 13 in the appendix. In Figure 13, Clean baseline always outperforms CWPLUGIN for all Validation Set sizes where as in Figure 5, Clean has lower mean of 0.528 as compared to 0.541 of CWPLUGIN.**\n\n**Reproducibility**: The code for reproducing the results has not been provided, which may raise concerns about the reproducibility of the findings. Making the code available would enhance transparency.\n\n**Minor comments**: \nLine 263: typo: “…linear-diagonal metric Both results…” has a missing full-stop." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 algorithm’s description is confusing. The algorithm is only iterating over k, but the caption shows iteration over class pairs. The symbol m is abused, sometimes indicating number of classes and sometimes a specific class. \n2. What metrics are the original models optimizing? This is important for understanding the results. \n3. The authros should re-evaluate whether a number should be bolded in each table. When there is overlap in the error intervals, the result is not usually not considered significant. \n4. Table captions should be above the table. Figure 5 is a table. Figures and tables should not be combined together as is the case for Figure 3 and Figure 2. Please present them individually. \n5. The abstract on openreview is different from that of the paper. Please converge on a single version of the abstract." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. CWPlugin is a simple and explainable optimization method, without relying on black-box procedures. Non-ML experts could adopt this algorithm. \n2. The algorithm in theory recovers the Bayes optimal classifier. \n3. Empirical performance is consistent across datasets, and equally importantly, across varying validation set sizes. It works on several metrics related to the confusion matrix and under distribution shift settings. \n4. Relvant to LLM practitioners, the method beats fine-tuning on BERT when data is limited." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes CWPlugin, a post-hoc algorithm to optimize a black-box ML model’s performance on a specifci metric. The algorithm iterates over class pairs and optimizes the decision boundaries à la Bayes optimality. The authors proves that the method is theoretically more efficient than brute force search with a binary search implementation in the average case. Empirically, the method is compared against several post-hoc and calibration baselines on prediction tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Novelty is limited in the case of CWPlugin. The method’s core is finding an optimal vector w to reweight a black-box model’s output. The novel contribution within the algorithm is the local search procedure over pairs of classes. The improvement here is rather incremental. As a more subtle point, the baselines used in Section 4 is inadequantely discussed in the related work subsection. Only probing is mentioned earlier in the paper. This results in a weak qualitative comparison with prior work.\n2. A major section of the paper shows CWPlugin’s superior efficiency. This advantage is not demonstrated in the empirical section: there are no concrete runtime comparisons. Speedup from parallelization is also unverified. \n3. The presentation of empirical results is not convincing. The tables adopt a specific validation set size, all different depending on the dataset but lacks justification. Sometimes the tables show 4 metrics, and sometimes only 2. This might indicate cherry-picking. The figures show more comprehensive results, but their presentation is poor. Legends are blocking the lines. Figure and axes titles are not formatted properly, and the color bands adds to the confusion. Based on the color bands, CWPlugin does not have an advantage in performance over the baselines due to overlap. \n4. Further justification of why the datasets are chosen is necessary. Most of them also require proper citations. More importantly, the authors need to show these datasets as standard for previous works in the same field. For example, the income predicion dataset is uncommon in deep learning from a quick search. \nOverall, some of these aspects can be fixed and improve the quality of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "**Minor concerns:**\n1. Are there any theoretical analyses for the proposed optimization/adaptation procedure if the metric is not linear-diagonal? I realize this might be a big ask, but additional comments on this would be appreciated.\n2. Clarifying question: is the validation set/test set sampled from the same target distribution (different from the set the black-box model is trained on)?\n3. An important naive baseline to include is directly optimizing $\\mathbf{w}$ on some likelihood objective defined over the validation set $\\mathcal{S}$. This does not optimize for the metrics of interest $f$, but it is important to see the effects this would have on performance.\n4. There is a concern for overfitting or reward hacking of $f$ since $\\mathbf{w}$ is tuned solely on (a small) $\\mathcal{S}$. The empirical results from the test set seem to indicate this is not a major concern. But can the authors comment on this point with more detail?\n5. Importantly, what are the effects of different choices of $\\varepsilon$ on post-hoc adaptation performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper focuses on an interesting and relevant problem, namely post-hoc adaptation of a black-box classifier, with only a small sample from a target distribution of interest.\n2. The method proposes a novel way to optimize weights to re-weigh predicted probabilities and applies to multi-class classification.\n3. The paper reads well, with intuition and formal depictions interleaved in a nice way." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method to perform post-hoc adaptation of a black-box classifier, by tuning class weights to optimize a metric of interest on a target distribution. Specifically, the authors introduce CWPlugin, which performs a line search to find the best reweighting coefficient for each (m-1) class, based on pairwise comparisons against a reference class." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Main concerns:**\n1. The authors only compare empirically against calibration techniques. The proposed problem of finding a set of weights $\\mathbf{w}$ to adapt a pre-trained model to optimize a black-box metric (which the authors refer to as `query-access only` metrics) is an **instantiation of black-box optimization**. As such, mature BBO techniques, including Bayesian optimization [R1], population-based search (e.g., evolution strategies) [R2], and even random search, should be included in the empirical comparisons.\n2. The proposed line-search method has quite high complexity $\\mathcal{O}(mn/\\varepsilon)$, and scales linearly with number of classes $m$, and number of samples $n$, and inversely with search resolution $\\varepsilon$. Techniques like BO and ES remove this dependence on $m$, $n$ and might be more suitable. They also do not require the pairwise comparison that is the basis of the line search.\n3. Fundamentally and theoretically, what are the benefits of the author's proposed approach over these BBO techniques (including BO, ES)? Based on my understanding, the proposed technique has some theoretical underpinnings for linear-diagonal metrics, but as the authors emphasize that they focus on 'query-access only' metrics, it is not clear whether there are broader theoretical guarantees.\n4. How is the reference class selected? The authors claim that the choice has `little impact to the algorithm`, but this is not justified. More specifically, how does the choice of reference class impact the optimization, since the line search is performed based on pairwise comparisons (against the reference class)?\n\n[R1] Snoek, J., Larochelle, H. and Adams, R.P., 2012. Practical bayesian optimization of machine learning algorithms. Advances in neural information processing systems, 25.\n\n[R2] Salimans, T., Ho, J., Chen, X., Sidor, S. and Sutskever, I., 2017. Evolution strategies as a scalable alternative to reinforcement learning. arXiv preprint arXiv:1703.03864." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A simple coordinate-wise optimization method to adapt black-box models by post-processing their predictions with a scaling vector." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024an,\ntitle={An Efficient Plugin Method for Metric Optimization of Black-Box Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yEnJvc7ogD},\nnote={under review}\n}" }, "abstract": { "value": "Many machine learning algorithms and classifiers are available only via API queries as a ``black-box'' --- that is, the downstream user has no ability to change, re-train, or fine-tune the model on a particular target distribution.\nIndeed, a downstream user may not have any knowledge of the training distribution or performance metric used to construct and optimize the black-box model.\nWe propose a simple and efficient plugin method which takes as input arbitrary multiclass predictions and post-processes them in order to adapt them to a new target distribution, while simultaneously optimizing for a particular metric of the confusion matrix.\nImportantly, the plugin method is \\textit{post-hoc}, does not rely on feature information, and only requires a small number of probabilistic predictions along with their corresponding true label.\nWe empirically demonstrate that plugin has performance competitive with related methods on a variety of tabular and language 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": [ "optimization", "black box systems", "domain adaptation", "distribution shift", "classification" ] }, "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/a5916c6360f04daed29de9c88907533849166537.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "An Efficient Plugin Method for Metric Optimization of Black-Box 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": "Please discuss the issues mentioned above.\n\nL. 134: in the formula brackets \\{...\\} are missing.\n\nL. 186: How do you define the adjacency matrix A? Do you mean really that A_{i,j} = 1 iff i is connected with j in G?\n\nL. 814: Lemme D.1 --> Lemma D.1" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Counting the number of paths and cycles (of specific lengths) in graphs is an important task in algorithmics and combinatorics with many applications to different fields. The authors show that using the DRL approach in combination with the Monte Carlo Tree Search method allows the discovery of more efficient matrix-based formulae for counting paths (of lengths up to six) than the best known so far. This is a nice achievement of the work.\n\nThe authors present MCTS-based DRL algorithm, termed Grammar Reinforcement Learning (GRL), that started with context-free grammar uses an equivalent Pushdown Automaton (PDA) for generating searching trees. To learn policy and value functions in the GRL a transformer architecture is proposed which models the PDA. This provides an interesting connection between the use of grammars, PDAs and reinforcement learning. Using this approach, the algorithm were able to discover formulae that are more efficient than those proposed by Voropaev and Perepechko (2012)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a Reinforcement Learning based method to discover an algebraic formula involving (adjacency) matrix of an undirected graph and some constant matrices, for counting the number of paths and cycles of specific (short) lengths in the graph. The main experimental achievement is finding (simple) formulas for path lengths l = 2, 3, 4, 5, 6. In the case of l = 2 the discovered formula is equal to the best known formula proposed by Voropaev and Perepechko (2012), and for l = 3, 4, 5, 6 the algorithm found even more efficient alternative expressions than those introduced by Voropaev and Perepechko.\n\nThe algorithm implements a searching strategy to find an optimal formula generated by an appropriate context-free grammar (CFG), or equivalently by a Pushdown Automaton (PDA). To this aim the authors propose a Monte Carlo Tree Search (MCTS) based Deep Reinforcement Learning algorithm, termed Grammar Reinforcement Learning (GRL)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper is largely based on the work of Piquenot et al. (ICLR 2024), that introduced a methodology based on CFGs that led to the construction of a new Grammatical Graph Neural Network model. It is provably 3-WL equivalent. As stated in the submitted work the generative framework of Piquenot et al. (ICLR 2024) already produces all formulae identified by Voropaev and Perepechko (2012). Hence, the innovative aspects of this submission are somewhat limited. Furthermore, it is not clear to what extent the proposed approach is generic. It would be interesting to have a discussions on such applications. It would be also interesting to compare the methods using the obtained matrix-based formulae with other methods for counting paths and cycles of lengths up to six." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How are the proofs of equivalence for the grammar-generated formulae derived? Is it challenging to verify correctness once a candidate formula is generated?\n2. By characterizing various rewriting rules used in these proofs, could it be possible to automate the generation of more compact formulae using traditional methods, such as a rewriting system?\n3. If so, has the proposed approach been compared against exhaustive explorations that utilize such rewriting rules?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper addresses a problem of foundational significance, exploring the automation of mathematical creativity through machine learning. \n- The approach is well-motivated and effectively presented, with a clear architecture and a detailed explanation of generating formulae from a given grammar. \n- The framework successfully recovers several known formulae while also discovering new, more compact ones." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a grammar-based reinforcement learning (RL) approach to synthesize novel formulae that describe path and cycle counts in graphs. Building on prior work, particularly by Voropaev and Perepechko, which provided matrix multiplication and Hadamard product-based formulae for counting expressions, this study defines a context-free grammar (CFG) capable of generating these expressions. Using a reinforcement learning and transformer-based approach, the authors search for more compact formulae with improved time complexity. They report multiple successes with this framework, though they note that training the model is time-intensive." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The importance of path and cycle counting is not sufficiently argued, particularly for an ICLR audience that may require more context on its broader relevance.\n- It remains unclear whether the RL and transformer framework is essential to the solution, as it seems more like a general-purpose tool for exploring terms generated by a given CFG..\n- The framework does not provide a proof of correctness for the generated formulae, requiring a time-consuming, manual derivation for each result. Automating the generation of explanations (a sequence of rewriting rules) to demonstrate correctness would significantly improve the approach's appeal.\n- The method appears more general beyond its current application domain. The paper could be strengthened by exploring additional applications of the proposed framework." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Regarding Proofs:\n- In the proof of theorem 3.1/A.1 it says that $(N \\times \\mathrm{J} \\times \\mathtt{diag}(w)) \\cdot \\mathrm{I} = \\mathtt{diag}((N \\cdot \\mathrm{J}) \\times w)$ but if $n=2$, $N = \\begin{bmatrix} a & b \\\\\\\\ c & d \\end{bmatrix}$, $w = \\begin{bmatrix} x \\\\\\\\ y \\end{bmatrix}$, we have $(\\begin{bmatrix} a & b \\\\\\\\ c & d \\end{bmatrix} \\times \\begin{bmatrix} 0 & 1 \\\\\\\\ 1 & 0 \\end{bmatrix} \\times \\begin{bmatrix} x & 0 \\\\\\\\ 0 & y \\end{bmatrix}) \\cdot \\begin{bmatrix} 1 & 0 \\\\\\\\ 0 & 1 \\end{bmatrix} = (\\begin{bmatrix} b & a \\\\\\\\ d & c \\end{bmatrix} \\times \\begin{bmatrix} x & 0 \\\\\\\\ 0 & y \\end{bmatrix}) \\cdot \\begin{bmatrix} 1 & 0 \\\\\\\\ 0 & 1 \\end{bmatrix} = \\begin{bmatrix} bx & 0 \\\\\\\\ 0 & cy \\end{bmatrix}$ and $(\\begin{bmatrix} a & b \\\\\\\\ c & d \\end{bmatrix} \\cdot \\begin{bmatrix} 0 & 1 \\\\\\\\ 1 & 0 \\end{bmatrix}) \\times \\begin{bmatrix} x \\\\\\\\ y \\end{bmatrix} = \\begin{bmatrix} 0 & b \\\\\\\\ c & 0 \\end{bmatrix} \\times \\begin{bmatrix} x \\\\\\\\ y \\end{bmatrix} = \\begin{bmatrix} by \\\\\\\\ cx \\end{bmatrix}$, hence, $\\mathtt{diag}(\\begin{bmatrix} by \\\\\\\\ cx \\end{bmatrix}) = \\begin{bmatrix} by & 0 \\\\\\\\ 0 & cx \\end{bmatrix}$. Therefore, I do not yet see how the induction can work.\n- In the proof of theorem D.1, it is used that $\\mathrm{J} \\cdot (A\\times(\\mathrm{I} \\cdot (A \\times A))) = A\\times(\\mathrm{I} \\cdot (A \\times A))$ but if $A=\\begin{bmatrix} 1 & 1 \\\\\\\\ 1 & 1 \\end{bmatrix}$, we have $A \\times A = 2A, \\mathrm{I} \\cdot A = \\mathrm{I}, \\mathrm{J} \\cdot A = \\mathrm{J}$ and hence $A\\times(\\mathrm{I} \\cdot (A \\times A)) = A \\times 2\\mathrm{I} = 2A \\neq 2\\mathrm{J} = \\mathrm{J} \\cdot (A\\times(\\mathrm{I} \\cdot (A \\times A)))$.\n\nRegarding PDA:\n- The usage of nondeterministic PDA is a bit confusing as PDAs are usually language acceptors rather than language generators. Still, of course, this is technically just a relabelling of the input as output. This happens in the text without clarification.\n- There is also an inconsistent use of the terms 'transition relation' vs. 'transition function'.\n- The PDAs could be omitted entirely using generative syntax trees induced by context-free grammar as there is no build-up on automata theory. For Gramformers, this would mean defining a variable token for every nonterminal character, a rule token for every production, and a terminal token for every terminal character. This would make the approach much easier to follow." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper introduced an exciting setup of transformer-guided MCTS on context-free grammar in a reinforcement learning setting for generating specific words of the corresponding context-free language.\n- The chosen application of efficient counting cycles is quite versatile in real-world scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a reinforcement learning setting for the generation of formulae that allow for efficient counting paths and cycles up to length six in directed graphs by implementing a transformer-guided Monte Carlo tree search on a generative syntax tree of a suitable context-free grammar.\n\nWhile the application of efficiently counting cycles is very intuitive, and the method of transformer-guided MCTS on formal grammar is very interesting, the paper misses any discussion of related work concerning RL methods of deep neural network-guided MCTS or MCTS on formal grammar. The paper would also heavily benefit from discussing other use cases of transformer-guided MCTS on formal grammar for GRL such that the paper could have focused more on the method of GRL rather than a specific use case only. The explanations and proofs also need clarity and seem incomplete or partly wrong." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper does not mention any other work related to the used/introduced machine learning method of grammar/language/llm-guided MCTS (the cited work is only about the chosen application of efficiently counting paths/cycles in graphs).\n- The paper could include other GRL applications to strengthen the introduced method's flexibility.\n- The relevance of counting cycles could be better illustrated with a few examples next to the literature reference.\n- Some explanations need coherence and clarity, especially the usage of PDAs in the paper, which could be more accurate or even omitted. (see Questions)\n- At least two proofs must be completed or corrected, and a proper list of necessary conditions/definitions could improve readability in most cases. (see Questions)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "C1. Given the usage of a reinforcement learning approach, the Markov Decision Process (i.e., the state/action/transitions/rewards) should be clearly specified in mathematical terms. The current textual description is ambiguous.\n\nC2. The parameters used in your model (e.g., exploration parameter for MCTS, architectural parameters and optimizer / learning rate etc. for the transformer) should be clearly specified in an Appendix. The reproducibility is limited otherwise.\n\nC3. There are a few clashes in notation because of trying to unite RL and CFG/DFA notations. For example, Q is used to refer to both the set of states of the automaton as well as the Q-value in RL. I'd suggest checking that each symbol is used consistently.\n\nC4. I would suggest discussing whether and how the approach could apply to directed graphs as well (it seems undirected graphs are assumed).\n\nC5. Could you also comment on what would be needed to generalise the approach to cycles of arbitrary length (more than 7)? Is it simply a matter of applying your method more computational power, or is the approach limited in this sense to predefined path lengths?\n\nC6. Typos: \"ouputs\" (Fig 6 caption)" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1. The authors treat an important problem with an interesting and novel methodology, and use it to discover demonstrably better path and cycle search algorithms.\n\nS2. The paper is fairly well-written and the technique appears sound." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers the problem of path and cycle counting in graphs. It aims to do so efficiently using formulae that perform a set of pre-defined matrix operations, specified as a context-free grammar. The authors frame this task as a combinatorial optimization problem and propose a reinforcement learning method to solve it. The method is made up of a Monte Carlo Tree Search with neural networks for function approximation. Particularly, the authors propose a transformer architecture that can process tokens from the grammar. The authors demonstrate that their method discovers more efficient formulae for path and cycle counting than those that were previously known." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. The biggest weakness of the work is its very limited evaluation. The paper proposes a very complex methodology for operating in this discrete search space, but does not compare it with meaningful baselines. It is therefore not possible to determine whether the proposed method is indeed a better means of navigating this search space than other more standard methods, especially given the very large reported computational cost. At the very least, in my opinion, the paper should include empirical comparisons with:\n\n- Online Monte Carlo Tree Search without any function approximation (to demonstrate the benefit of using the transformer within the search);\n- The learned transformer policy at convergence;\n- A classic metaheuristic such as simulated annealing;\n- A simple random search that samples from the grammar and is given the same amount of rollouts as MCTS.\n\nW2. Another weakness of the work is that the method does not generate algorithms that are provably correct. Indeed, the authors have to resort to proving the correctness of the algorithms themselves, which they do. This stems from the fact that, to evaluate a given formula, the method needs to compare the output over a limited set of graphs with generated ground truth values. It could be the case that a formula produces the correct outputs for this set of examples while missing some edge cases. This is a limitation that should be acknowledged and discussed.\n\nW3. The literature review omits many related works on reinforcement learning for combinatorial optimization over graphs. I would suggest including at least [1], one of the first recent works to treat this type of problem with RL, and which also uses token-like sequences; [2], a work that made substantial leaps in RL for combinatorial optimization in terms of performance and scalability; as well as other recent works in this space (see [3] for a survey).\n\n[1] Vinyals, O., Fortunato, M., & Jaitly, N. (2015). Pointer networks. Advances in neural information processing systems, 28.\n\n[2] Khalil, E., Dai, H., Zhang, Y., Dilkina, B., & Song, L. (2017). Learning combinatorial optimization algorithms over graphs. Advances in neural information processing systems, 30.\n\n[3] Darvariu, V.-A., Hailes, S., & Musolesi, M. (2024). Graph Reinforcement Learning for Combinatorial Optimization: A Survey and Unifying Perspective. Transactions on Machine Learning Research." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Could you help to answer my concern on the first point of weakness?\n- Could you clarify the purpose of the cubic shapes in Figures 3, 4, and 5? Are they intended to convey a particular meaning beyond their function as symbols?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Great literature work: This paper references existing literature, which demonstrates an awareness of previous work in the area and situates the study within the broader research context.\n- Full details of the proposed method with detailed theoretical justification." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new approach to efficiently count paths and cycles in graphs. GRL uses a pushdown automaton (PDA) approach to generate and optimize mathematical formulas for path and cycle counting, addressing computational challenges in fields such as network analysis, biology, and social sciences. By framing path/cycle counting as a CFG-constrained search problem, GRL discovers new matrix-based formulas, improving computational efficiency by two to six times over current methods. Key contributions include a generic framework for generating efficient formulas within a CFG, the development of Gramformer to learn policies and values within a PDA model, and the identification of novel, efficient counting formulas for graph substructures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The focus of this work is on re-solving a problem that requires a polynomial-time algorithm, as indicated in Line 99: \"As mentioned in Section 1, path/cycle counting has been extensively tackled in the literature.\" This context renders the proposed methods less impactful. I recommend that the authors provide a more compelling rationale for why this issue is considered challenging within the fields of network analysis, computer science, biology, and social sciences.\n\n- There are no experiments conducted on datasets from network analysis, computer science, biology, or social sciences. Given that the results in Figure 7 indicate an algorithm running in 0.2 seconds, which is already considered \"very fast,\" it is unclear what specific challenge is being addressed in this figure.\n\n- The mathematical notation used throughout the paper tends to create confusion rather than enhance the reader's understanding of the problem." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024grammar,\ntitle={Grammar Reinforcement Learning: path and cycle counting in graphs with a Context-Free Grammar and Transformer approach},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yEox25xAED},\nnote={under review}\n}" }, "abstract": { "value": "This paper presents Grammar Reinforcement Learning (GRL), a reinforcement learning algorithm that uses Monte Carlo Tree Search (MCTS) and a transformer architecture that models a Pushdown Automaton (PDA) within a context-free grammar (CFG) framework. Taking as use case the problem of efficiently counting paths and cycles in graphs, a key challenge in network analysis, computer science, biology, and social sciences, GRL discovers new matrix-based formulas for path/cycle counting that improve computational efficiency by factors of two to six w.r.t state-of-the-art approaches. Our contributions include: (i) a framework for generating transformers that operate within a CFG, (ii) the development of GRL for optimizing formulas within grammatical structures, and (iii) the discovery of novel formulas for graph substructure counting, leading to significant computational improvements." }, "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": [ "Graph", "Reinforcement Learning", "Grammar", "Cycle Counting" ] }, "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/3def87a8eba80de2f61a06bb1723ffdbd54d04f1.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": "Grammar Reinforcement Learning: path and cycle counting in graphs with a Context-Free Grammar and Transformer 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": 3 }, "primary_area": null, "questions": { "value": "1. When converting various combinatorial optimization problems to TSP instances, are there cases where the reduction process fails or underperforms due to problem characteristics? How does the method handle these instances?\n\n2. Given that MatDIFFNet has longer inference times due to complex diffusion steps, are there plans to optimize the model architecture or algorithm to improve inference speed and computational efficiency?\n\n3. In multi-task training, does the interaction between different tasks lead to performance drops in any specific task? Is there a clear mechanism in the model to handle task weight allocation and interdependencies?\n\n4. How robust are MatPOENet and MatDIFFNet when the input data contains noise or incomplete information? Were there any robustness tests conducted?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The introduction of MatPOENet and MatDIFFNet, which use Transformer-based and diffusion-based models, respectively, showcases the application of advanced neural network structures to solve matrix-encoded TSP problems.\n\n- The RedCO framework offers a novel approach by unifying different combinatorial optimization (CO) problems through reduction to a general TSP format. This reduction expands the scope of neural solvers to tackle diverse problem types in a single architecture.\n\n- RedCO's capability to handle non-metric, asymmetric, and discrete TSP instances, unlike traditional Euclidean-focused TSP solvers, significantly broadens its applicability.\n\n- The RedCO framework is designed to incorporate various solver types, including existing methods like DIMES, showing the framework's modularity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a unified neural solver framework called RedCO, which uses problem reduction techniques to map different combinatorial optimization (CO) problems to the general Traveling Salesman Problem (TSP) format. Two novel neural solvers, MatPOENet and MatDIFFNet, are introduced to handle matrix-encoded inputs and solve these problems efficiently. This work aims to extend neural combinatorial optimization beyond specific problem types by providing a scalable solution for problems like asymmetric TSP (ATSP), directed Hamiltonian cycle problems (DHCP), and 3-Satisfiability (3SAT)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the framework performs well for medium-scale problems, its efficiency and feasibility for large-scale, real-world instances (e.g., with tens of thousands of nodes) are not thoroughly demonstrated or tested.\n\n- The use of complex neural models like MatPOENet and MatDIFFNet makes it difficult to understand the inner workings and decision-making processes of these solvers. More interpretability features or case studies would be beneficial.\n\n- The paper mainly focuses on synthetic data for testing, with limited discussion on how the models would handle real-world problem instances that could have different statistical properties.\n\n- There is little exploration into how the proposed solvers manage noisy or incomplete data, which is common in practical applications.\n\n- The MatDIFFNet, while powerful for certain problem types, is computationally intensive, which may hinder its use for larger instances or require additional optimization strategies." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Table 2 shows that MatPOENet and MatDIFFNet outperform LKH in solving 3SAT problems, but they tend to produce worse results in most other scenarios. Could you provide some explanations for this?\n2. How does RedCO perform on standard TSPs (Symmetric TSPs)?\n3. In line 268, the POE is based on $f(x) = 1/cosh(100x)$, can you give more introduction of the empirical function?" }, "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 easy to understand.\n2. As far as I know, this is the first study attempting to create a general framework for learning various COPs in a unified manner.\n3. The experiments conducted are thorough, and the results effectively showcase the framework's capability to handle arbitrary matrix-encoded TSPs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors propose RedCO, to unify a set of CO problems by reducing them into the general TSP form featured by distance matrices. RedCO demonstrates the potential to efficiently train a neural TSP solver using a diverse range of CO instances, and can also be adapted to specialize for specific problem types." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The organization of the experimental section is lacking. With seven research questions (RQs) presented, the lack of clear categorization makes this part of the paper somewhat difficult to navigate.\n2. The results for DIFUSCO and T2T are not included. It is noted that MatDIFFNet performs well on 3SAT problems, which is developed upon DIFUSCO and T2T.\n3. While the specific problem reduction is detailed in Appendix A, it would be helpful to have a more detailed introduction to the reduction principles and the applicable COPs. Specifically:\n - What types of COPs (or what properties must COPs have) can be reduced to a general TSP?\n - What considerations should be taken into account when performing this reduction?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. MatDIFFNet is trained on 8 NVIDIA H800 80GB GPUs with Intel Xeon (Skylake, IBRS) 16-core CPU is super computational resource consuming, I am curious about the ablation experiments on computational resources.\n2. The test questions in the article experiments are too limited, this article mentions applicability for ``P ≤P general TSP or P in matrix format (VC, Clique, VRPs, FFSP, MIS, etc.)`` etc., I would highly recommend to introduce evaluations on more CO problems to respond to my concerns on applicability.\n3. In Line 1137, you mention that ``Also, they generally evaluated their proposed methods on no larger than 100 nodes of TSP/VRP instances, with a major emphasis of methodological innovations rather than eager pursuit of scalability at sheer engineering level.`` What means the sheer engineering level? Also, it seems that this paper also mainly focuses on no larger than 100 nodes. Please provide a clear explanation.\n4. This paper uses a unique test problem design, which I think requires the authors to implement more comparative algorithms (e.g., GOAL, MVMOE) on the problems they cover for experimental validation.\n5. The results in Table 5 are not sufficient to illustrate performance on larger-scaled data, I think you should provide experiments without the aid of an external process to explore whether the model has the ability to scale up. Also, this paper does experiments on scales of 20-100, and I doubt that it makes sense to compare methods that address large-scale CO problems such as GLOP." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This article is well-written and reasonable.\n2. The author has carried out abundant experiments and discussions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an interesting approach to dealing with multi-task CO by transforming several CO problems into equivalent TSPs. This paper also proposes two new solvers, MatPOENet and MatDIFFNet, to solve the following TSP." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I think this article has two major weaknesses that should be considered.\n1. This paper adopts a quite special modeling approach, and its applicability is worrying. I doubt the effectiveness of reducing multi-CO problems into the general TSP form featured by distance matrices. \nAccording to Fig. 1, you show that NP problems can be transformed into SAT, I am concerned if this part can be proved. For some problems, the transformation into TSP is itself an NP problem if you want to maintain the Found Rate ``FR`` (e.g., CVRP, as mentioned in Appendix D.2.4, ``first clustering points and then solving each cluster as a TSP`` will harm the FR), and even if it can be transformed into TSP, the time complexity of such transaction may increase dramatically.\n2. Some of the experiments in this paper are not clearly described. I tried my best to find out but it is still not clear what the exact settings of the * version, single, and mixed in Table 2 are." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. In RQ2 you show a comparison of solution times and results with the LKH method. The reason you showed efficiency in this experiment compared to LKH seems to come entirely from superior performance on the 3-SAT problem. I don't think it is fair to take Average L in this case. I am more curious as to why LKH performs poorly on the 3-SAT problem and where MatPOENet and MatDIFFNet excel in this problem. Can you provide a visual example to help me understand this result intuitively?\n2. For solving efficiency, I think this paper should be compared more with Gurobi for efficiency. I am very curious about the results of this part. I also suggest you add time as reference in Table1.\n3. I am having trouble understanding the specific N and d settings for variants in the ii) part of RQ3 and especially in Table 4. I need more explanation about it.\n4. What is the significance of MatDIFFNet? Based on the results so far (ignoring the future work you mentioned) it looks like its lagging behind in performance and efficiency as well as training efficiency. I would suggest deleting this section or putting it in the appendix. Also The authors say in RQ7 that MatDIFFNet has the potential for more accurate solution space for larger scale instances while mentioning in the limitation that ``MatDIFFNet has the potential for direct solving of larger instances but is currently yet to be implemented.`` But I can't find any evidence for this. But I can't find any evidence for this , please explain this." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The Problem reduction of this paper has theoretical support.\n2. The proposed methods of this paper has advantages in terms of effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This article focuses on multi-task CO problems, proposes a solution method that is general for several CO problems and presents two efficient solvers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The RedCO approach proposed in this paper is not intuitively applicable to a wide range of CO problems. I think the value of multi-task CO should be reflected in its applicability to most CO problems.\n2. The contribution of this paper is weak, translating these problems into a TSP is not a new idea and TSP solver is quite well developed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024unified,\ntitle={Unified Neural Solvers for General {TSP} and Multiple Combinatorial Optimization Tasks via Problem Reduction and Matrix Encoding},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yEwakMNIex},\nnote={under review}\n}" }, "abstract": { "value": "Various neural solvers have been devised for combinatorial optimization (CO), which are often tailored for specific problem types, ranging from TSP, CVRP to SAT, etc. Yet, it remains an open question how to achieve universality regarding problem representing and learning with a general framework. This paper first proposes RedCO, to unify a set of CO problems by reducing them into the general TSP form featured by distance matrices. The applicability of this strategy is dependent on the efficiency of the problem reduction and solution transition procedures, which we show that at least ATSP, HCP, and SAT are readily feasible. The hope is to allow for the effective and even simultaneous use of as many types of CO instances as possible to train a neural TSP solver, and optionally finetune it for specific problem types. In particular, unlike the prevalent TSP benchmarks based on Euclidean instances with 2-D coordinates, our focused domain of general TSP could involve non-metric, asymmetric or discrete distances without explicit node coordinates, which is much less explored in TSP literature while poses new intellectual challenges. Along this direction, we devise two neural TSP solvers with and without supervision to conquer such matrix-formulated input, respectively: 1) MatPOENet and 2) MatDIFFNet. The former is a reinforcement learning-based sequential model with pseudo one-hot embedding (POE) scheme; and the latter is a Diffusion-based generative model with the mix-noised reference mapping scheme. Extensive experiments on ATSP, 2DTSP, HCP- and SAT-distributed general TSPs demonstrate the strong ability of our approaches towards arbitrary matrix-encoded TSP with structure and size variation. Source code and data will be made public." }, "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": [ "Travelling Salesman Problem", "Neural Combinatorial Optimization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/af162805c8600bab281edcce3e6f0eb428c69872.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Unified Neural Solvers for General TSP and Multiple Combinatorial Optimization Tasks via Problem Reduction and Matrix Encoding" }, "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": 4 }, "primary_area": null, "questions": { "value": "Three models are presented in the paper (Tiny, Small and Base) except the overfitting large model. What is the model presented in the supplementary files?\n\nIs there any reason to use an image VAE instead of audio VAEs? \n\nDid the authors observe any advantage of GLA over a neural vocoder?\n\nIs it possible to run a subjective listening test, and see the consistency between human evaluation and the audio-video alignment accuracy measured by a DNN model?\n\nCould the authors measure the FAD scores on top of the current FID scores?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "Most contributions of MDSGen are about micro design aspects.\n1. Channel selection of Mel-spec\n2. Time-aware masking strategy for generative models\n3. Reduced dimension of the video features\n4. Small model size and fast inference speed" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes \"MDSGen\", an efficient model based on Masked Diffusion Transformer, for video-to-audio generation.\n\nThe challenges of video-to-audio generation are mainly:\n1. Heavy computation and memory usage;\n2. Requirements for the audio quality;\n3. Requirements for the audio-video alignment;\n\nMDSGen reduces the resource consumption by using very light-weight Transformer coupled with fast diffusion samplers such as DPM solver, as well as a dimension reduction module to reduce the size of the video conditioning embeddings.\n\nMDSGen improves audio quality and audio-video quality by introducing a time-aware masking strategy into the mask DiT framework, together with other efforts.\n\nConceptualy, MDSGen looks like a framework that replaces the \"text prompt\" in text-to-audio DiT [StableAudioOpen],[MakeAnAudio2] by a video feature embedding. Hence the technical contributions are more in micro aspects.\n\nHowever, some design choices may have severely affected the audio quality, making the work less solid or reusable to the community. Audio quality observed in the supplementary files is far from the level in modern text-to-audio models such as [AudioLDM], [MakeAnAudio2], [SpecMaskGIT], [StableAudioOpen]." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "## Major issues\n### 1. Improper audio reconstrution pipeline\nMDSGen ustilizes the VAE from Stable Diffusion, which is not trained for Mel-spec. Although the authors carefully discussed how to take the most advantage of this image VAE, the discussion itself is **NOT** reusable for the audio community, as there have been plenty of audio VAE designs, some of which are publicly available such as [AudioLDM], [MakeAnAudio2], [StableAudioOpen], [DAC].\n\nAnother improper choice is that, MDSGen utilizes the Griffin-Lim Algorithm (GLA) to convert mel-spec back to wave forms. GLA has almost been abandoned by audio community, due to the recent advance in neural vocoder, e.g., [HiFiGAN], [UnivNet], [BigVGAN]. I believe the apparent phase distortion in the supplementary files might have been caused by GLA.\n\nThere is a rough comparison on the quality of audio reconstruction pipeline in a recent paper [SpecMaskGIT], I hope it could be useful for the improvement of MDSGen.\n\nI strongly recommend the authors to consider audio-specified reconstruction pipelines for improved audio quality. Even in audio-visual generation community, we can see the usage of such audio VAE for excellent audio quality, e.g., [VisualEchoes]\n### 2. Invalid claims on the result\nBecause the audio quality is far from the baseline in audio generation community, it is improper to claim that MDSGen is better in \"audio-video alignment\".\n\nI believe, the audio-video alignment can be evaluated only when the audio quality is sufficiently good. Given the current audio quality, I don't think the model is ready for further evaluation.\n## Minor issues \n### 1. Evaluation metrics\nThe FID used in this paper comes from the implementation of SpecVQGAN, a pioneer of audio generation. However, the FAD implementation ([AudioLDM],[FAD_github]) has been more widely accepted in audio community. Evaluating with the widely adopted FAD metric can also help the readers to compre the audio quality with other audio generation models.\n### 2. Insufficient ablation study\nMDSGen trains a learnable module to reduce the video feature sequence into a single vector. From Figure 6, we can observe that the learned weights are quite evenly distributed (except the beginning and ending frames).\n\nThe observation posts a question: How much improvement can the learnable reducer bring compared to a naive average pooling?\n\n[StableAudioOpen]: https://arxiv.org/abs/2407.14358\n[MakeAnAudio2]: https://arxiv.org/abs/2305.18474\n[AudioLDM]: https://audioldm.github.io/\n[SpecMaskGIT]: https://arxiv.org/abs/2406.17672\n[DAC]: https://github.com/descriptinc/descript-audio-codec\n[HiFiGAN]: https://github.com/jik876/hifi-gan\n[UnivNet]: https://github.com/rishikksh20/UnivNet-pytorch\n[BigVGAN]: https://github.com/NVIDIA/BigVGAN\n[VisualEchoes]: https://arxiv.org/abs/2405.14598\n[AudioLDMEval]: https://github.com/haoheliu/audioldm_eval\n[FAD_github]: https://github.com/gudgud96/frechet-audio-distance" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper introduces a novel framework for video-to-audio sound generation that effectively combines a temporal-aware masking strategy with a redundant feature removal module. \n\nMDSGen demonstrates significant improvements in model efficiency by using a smaller masked diffusion transformer architecture. The framework achieves high alignment accuracy on benchmark datasets with a fraction of the parameters, memory usage, and inference time compared to baselines. \n\nThe paper provides a structured explanation of MDSGen’s architecture and mechanisms, including the Temporal-Awareness Masking (TAM) and the Reducer module for filtering out redundant features. Extensive experimental results on VGGSound and Flickr-SoundNet datasets clearly validate the method’s effectiveness, with MDSGen achieving superior performance across alignment accuracy and efficiency metrics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents MDSGen, an efficient framework for vision-guided sound generation that minimizes model size, memory usage, and inference time. Key innovations include a temporal-aware masking strategy to enhance alignment accuracy and a redundant feature removal module to filter unnecessary video information. Using a lightweight masked diffusion transformer, MDSGen outperforms larger Unet-based models on VGGSound and Flickr-SoundNet, achieving high synchronization and alignment with significantly reduced computational costs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Lack of Novelty and Contribution: The paper presents the primary contributions are the Temporal-Awareness Masking (TAM) strategy and the visual Reducer module. However, masking strategies have been widely explored in audio generation research, as seen in works like [1, 2], and the specific concept of Temporal-Awareness Masking has been studied in [3, 4]. The visual Reducer module, primarily a 1x1 convolutional layer (line 181), lacks detailed design innovations, which limits its distinctiveness and impact.\n\n2. Insufficient Exploration of Design Choices: For video-to-audio generation, the choice of video encoder plays a crucial role in understanding the video content. Clarification on the selection of CAVP as the video encoder would add valuable insight. Additionally, the paper could explore using more video encoders, such as CLIP [5], VideoMAE [6], ViVit [7], and TAM [8], which could enrich the technical depth of the proposed method.\n\n3. Presentation and Writing:\nSome claims in the paper lack supporting evidence, such as the statements in lines 183-185 that the proposed method “minimizes redundant features that could lead to overfitting” and in line 224 that setting N_2 = 4 “gives better performance for audio data.” These points would benefit from empirical support to substantiate their validity.\n\n4. Supplementary Material: The quality of generated audio samples in the supplementary material raises concerns regarding the overall quality of results produced by the proposed method, which may affect its effectiveness and appeal.\n\n[1]Pascual S, Yeh C, Tsiamas I, et al. Masked Generative Video-to-Audio Transformers with Enhanced Synchronicity[J]. arXiv preprint arXiv:2407.10387, 2024.\n\n[2]Borsos Z, Sharifi M, Vincent D, et al. Soundstorm: Efficient parallel audio generation[J]. arXiv preprint arXiv:2305.09636, 2023.\n\n[3]Bai H, Zheng R, Chen J, et al. A $^ 3$ T: Alignment-Aware Acoustic and Text Pretraining for Speech Synthesis and Editing[C]//International Conference on Machine Learning. PMLR, 2022: 1399-1411.\n\n[4]Garcia H F, Seetharaman P, Kumar R, et al. Vampnet: Music generation via masked acoustic token modeling[J]. arXiv preprint arXiv:2307.04686, 2023.\n\n[5]Radford A, Kim J W, Hallacy C, et al. Learning transferable visual models from natural language supervision[C]//International conference on machine learning. PMLR, 2021: 8748-8763.\n\n[6]Tong Z, Song Y, Wang J, et al. Videomae: Masked autoencoders are data-efficient learners for self-supervised video pre-training[J]. Advances in neural information processing systems, 2022, 35: 10078-10093.\n\n[7]Arnab A, Dehghani M, Heigold G, et al. Vivit: A video vision transformer[C]//Proceedings of the IEEE/CVF international conference on computer vision. 2021: 6836-6846." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "No concern." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. The reducer is designed to have fixed weights, serving as a weighted average of all the frames. However, the sound events of different videos are distinct, so why not adopt a dynamic weighted average strategy, e.g., attention pooling?\n2. What is the exact implementation of the model? How is the visual conditioning (i.e., $p(x|v)$) implemented?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The idea of compressing visual representations into one single vector is bold and intriguing, which reduces significant computing pressure on the DiT side.\n2. The proposed TAM strategy is interesting and makes sense. Previous works about masking audio representations, such as AudioMAE, have drawn conclusions that the unstructured masking strategy is superior, which contradicts the conclusion in this paper. I believe this paper brings more insights into this topic. \n3. The authors have conducted tons of ablation experiments to support their model design and parameter decision, making the conclusions plausible." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces MDSGen, an innovative framework designed for vision-guided open-domain sound generation with a focus on optimizing model parameter size, memory consumption, and inference speed. It features two major innovations: a redundant video feature removal reducer and a temporal-aware masking strategy. Utilizing denoising masked diffusion transformers, MDSGen achieves efficient sound generation without the need for pre-trained diffusion models. On the VGGSound dataset, the smallest MDSGen model demonstrates a 97.9% alignment accuracy while using fewer parameters, consuming less memory, and performing faster inference compared to the current state-of-the-art models. The results underscore the scalability and effectiveness of this approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The major concern is about the modeling of audio representations, as I am familiar with this field. I believe that Mel spectrum is more of a 1D feature rather than 2D, because the spectrum does not satisfy translational invariance (if a formant chunk in a spectrum is moved from the bottom left to the top right, the semantics of the sound are likely to be completely destroyed), and the frequency domain and time domain cannot simply be simulated by spatial coordinates. A relevant observation in this paper is that a complete random masking strategy is underperformed by the temporal-aware masking strategy. Therefore, I believe that considering Mel spectrograms as gray-scale images and modeling them using 2D VAE pretrained with real images is suboptimal, which further prevents modeling sounds with varying lengths. There are already approaches that model audio using 1D VAE, such as Make-an-audio 2. So can the authors provide justifications for choosing 2D rather than 1D? In my view, choosing 1D combined with the TAM strategy could form a more compelling motivation.\n2. The idea of compressing visual representations into one single vector is intriguing. However, I don't understand why this could work. How does one single vector provide accurate information about temporal position? I believe Diff-foley works because it adopts a sequence-to-sequence cross-attention mechanism, which provides rough sequential and positional information for the audio to follow. Could the author provide further analysis and discussion on this point? For example, analyzing the components related to temporal position within that vector, or the relation of the learned weights of reducer between key frames of videos. \n3. Similar concern: the learned weights of reducer seem to be focused more on the head and tail frames of videos. Does this imply that the reducer is more focused on global video information? How can it be determined that it is capable of extracting local positional information?\n4. The alignment classifier proposed in Diff-foley only reaches 90% accuracy on their test set. However, the best performance in this paper reaches 98+. How could this happen? Is the classifier involved during the training process?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "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. MDSGen achieves strong results with a small model size, making it useful for real-time applications. Compared to larger models, it is faster and more memory-efficient.\n2. TAM is an interesting approach that focuses on time-based information in audio, aiming to improve alignment by using masking based on temporal patterns rather than spatial patterns (commonly used for images).\n3. The paper provides extensive experiments with detailed comparisons against other models. Ablations for each key component further clarify the model’s design choices." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents MDSGen, an efficient and compact framework for generating open-domain sounds guided by visual input. MDSGen uses masked diffusion transformers instead of the usual U-Net architectures, optimizing for faster speed, lower memory use, and parameter efficiency. Key features include a module to remove redundant video data and a Temporal-Awareness Masking (TAM) strategy, both aimed at efficient, high-quality audio-visual alignment. Tests on the VGGSound and Flickr-SoundNet datasets show that MDSGen achieves strong alignment accuracy with much lower computational demands than larger models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I question the decision to use an image-trained VAE (from Stable Diffusion) rather than an audio-specific VAE, such as those in AudioLDM [1]. An audio-dedicated VAE could better capture the temporal and spectral nuances inherent to sound, which are often lost when treating audio as an image. Relying on an image-based VAE reduces the model’s potential to fully leverage audio-specific features and may affect TAM’s performance.\n2. The authors highlight channel selection within the RGB output as a means of optimizing the final mel-spectrogram. While using the G channel showed marginal improvements, I question if relying on such RGB channel selection can sufficiently address the nuances of audio spectrogram representation (similar to 1). A more audio-specific solution that doesn’t require treating spectrograms as RGB images would likely be more consistent with the needs of audio data, as these channels are meant for pixels, not spectral data.\n3. I would suggest that the authors conduct a human perceptual study to better assess audio quality. Relying solely on quantitative metrics may not fully capture perceptual quality, as these measures can sometimes be unreliable.\n4. In Section 5.4, various masking strategies are explored, and TAM shows a clear improvement over random masking and FAM. However, the reasons for TAM’s superiority are not fully explained. It would be beneficial to discuss why TAM outperforms FAM in this context, particularly since FAM is intuitively suitable for audio data.\n5. I noticed a missing citation for SpecAugment [2] and AudioMAE [3], a masking approach relevant to TAM proposed here.\n\nReferences\n\n[1] Liu et al. AudioLDM: Text-to-Audio Generation with Latent Diffusion Models.\n\n[2] Park et al. SpecAugment: A Simple Data Augmentation Method for Automatic Speech Recognition.\n\n[3] Huang et al. Masked Autoencoders that Listen." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A novel approach is presented for highly efficient vision-guided sound synthesis using masked diffusion models" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024mdsgen,\ntitle={{MDSG}en: Fast and Efficient Masked Diffusion Temporal-Aware Transformers for Open-Domain Sound Generation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yFEqYwgttJ},\nnote={under review}\n}" }, "abstract": { "value": "We introduce MDSGen, a novel framework for vision-guided open-domain sound generation optimized for model parameter size, memory consumption, and inference speed. This framework incorporates two key innovations: (1) a redundant video feature removal module that filters out unnecessary visual information, and (2) a temporal-aware masking strategy that leverages temporal context for enhanced audio generation accuracy. In contrast to existing resource-heavy Unet-based models, MDSGen employs denoising masked diffusion transformers, facilitating efficient generation without reliance on pre-trained diffusion models. Evaluated on the benchmark VGGSound dataset, our smallest model (5M parameters) achieves 97.9% alignment accuracy, using 172x fewer parameters, 371% less memory, and offering 36x faster inference than the current 860M-parameter state-of-the-art model (93.9% accuracy). The larger model (131M parameters) reaches nearly 99% accuracy while requiring 6.5x fewer parameters. These results highlight the scalability and effectiveness 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": [ "vision-guided audio generation", "fast inference", "open-domain sound synthesis", "masked diffusion models", "temporal learning", "visual sound source localization", "generative AI" ] }, "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/fada5da88a48955dabddda9ffa955b128b00c16e.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/15cf2434cce953c18de4f78948a1e596b8f8c6b4.zip" }, "title": { "value": "MDSGen: Fast and Efficient Masked Diffusion Temporal-Aware Transformers for Open-Domain Sound 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": 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": "I think it will be useful if the paper can present accuracy numbers in similar training time, or training time numbers at roughly the same accuracy numbers. That will be very useful in positioning this work against existing methods." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "## Presentation\nThis paper is well-written with thorough discussion on the related works, the design philosophy and the precise formulations of the proposed modeling. The discussions are usually precise and insightful. The important elements in building the proposed world-models, such as the POMDP formulation, the representation learning (including sufficient details, such as image augmentations, temporal consistency, covariance regularization), the dynamics learning (the conditional independence assumption and the resultant factorization) are all presented clearly, leaving the readers with no doubts on the technical components and the underlying reasoning.\n\nIn many cases, the discussions have involved clear contrast with prior works, for example in Table 1 where the design choices in a good selection of prior works are presented, and compared with the design choices in the proposed SGF.\n\n## Empirical Studies\nThe paper is tested on the standard Atari 100k tasks that is standard for this kind of work. The empirical study seems sounds and demonstrates a few interesting properties of this work that may be useful for researchers in this area. In particular\n- It has presented through ablation studies in Section 4.2. showing the importance of various design components, such as state (image) augmentations, action/frame stacking, temporal consistency and sample contrastive formulation. Additional details are also included in the appendix, such as in Table 3 in Appendix D.\n- The paper has presented detailed comparisons with other methods, as presented in Appendix D Table 2." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper discusses the design of a world model - a parametric model that predicts the transitions probabilities, reward functions and terminal state distributions to serve as a simulation environment for reinforcement learning. The paper proposes to simplify some existing elements in recent world-models, such as sequence models (i.e. RNNs, transformers), to focus on simple ingredients (e.g. frame stacking). It instead proposes to keep the maximum information and temporal consistency formulation as the most essential properties of effective world models. The resultant model is dubbed \"Simple, Good and Fast World Models\" (SGF) which demonstrates somewhat competitive performance against other existing world models in the Atari 100k benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are a few limitations that seem to limit the contributions of this work.\n\n- As discussed thoroughly in the related work section in the paper, world model in training reinforcement learning agents is not a new idea. In such cases, it is useful to establish that this work is addressing a significant weakness in prior works, without sacrificing other important metrics. In this case, the main motivation of SGF seems to be presenting a simple, fast, yet accurate method to train a good world model. While SGF certainly fit the bill for the first two, it seems to fall short significantly in the third. The result in Table 2 seems to suggest that SGF is much weaker than prior works such as IRIS and DreamerV3, which are all based on learning on imagination and hence are arguably fairly related.\n\n- The choice to drop sequence models appears to be a fairly significant limiting factor. This was acknowledged clearly in Line 462-464, which lists the decision to drop sequence models as a potential reasons why this work does not reach SOTA performance on Atari 100k. This seems like a fairly trivial observation - certain games in Atari 100k requires long term reasoning, and it is indeed one of the most challenging aspects of world models for RL agents. It should be expected that removing a components specifically designed to address this important challenge will result in inferior performance. It seems unlikely that researchers in this field will learn much more about how to design better world models if they are simply presented with results comparing methods with and without a sequence model - they probably already know that it is going to be much worse for certain tasks that require long term reasoning." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 do the method perform with long-term dependencies?\n2. How do feature extraction perform if we change the SSL training objective?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. I believe these kinds of works are important. It is easy to just incrementally propose new components to improve the performance of systems while considerably increasing the engineering complexity. This does not give a clear view of the actual importance of the components included in the SOTA of world modeling. Going in a completely different direction is in my opinion a needed move sometimes, and it will help to shape new design choices for world modeling. Hence, the motivation is strong, and the reasoning behind that is coherent.\n2. The results are compelling. I believe that on Atari100K the simple method proposed performs quantitatively well, even compared to baselines that are far more complex, and with considerably lower runtime.\n3. The comparison with existing methods is nontrivial and requires a proper analysis of the literature. Table 1 is also interesting and will be useful for future work.\n4. The paper is well-written and well-motivated, all introduced explanations are useful and the writing is compact enough. The proposed experiments are interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new approach to world modeling for the training of RL-based agents. The current state of literature employs different mechanisms to achieve a correct modeling of long sequences, including reconstruction, recurrent modules, memory, etc. The authors propose to simplify the world modeling and just use self-supervised techniques, inspired by VicReg, and a simple learning strategy for the system dynamics. There is a comparison of existing methods and the proposed one on Atari100k, and appropriate ablations of the components." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I believe that while the proposed method focuses on short-term dependencies, as correctly stated in the limitation, how much performance degrades with an increasingly long-term dependency on actions would be important to quantify. This will allow us to assess the limitations of the proposed method in a more robust manner, for people to build upon.\n2. It is not clear to me why only VICReg is chosen for representation extraction. There are relationships with BYOL and SimSiam as reported in the appendix, but it is unclear how performance would change if these approaches were instead used for representation extraction rather than VICReg." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Although this is partially explained in the Limitations, could you go into more detail about: for more complex tasks than Atari-games, where longer term knowledge needs to be remembered (where orders of magnitude more memory of early events is required), what parts of the models should be scaled up or significantly modified / replaced (e.g. using sequence model: RNN, Transformers, ...) to be optimal in terms of accuracy and training time?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "In this work, authors try to find the most necessary components that are the most optimal in terms of accuracy and training time.\n\n1. The presented SGF approach lies on the Pareto optimality curve on the chart of Accuracy (normalized mean score) and Training time (hours) - Figure 5, where the other points on the Pareto optimality curve are: SPR, DreamerV3, EfficientZero (performs lookahead)\n\n2. The optimal combination of improvements (frame stacking, action stacking, temporal consistency, augmentations, sample-contrastive) has been found to achieve the highest accuracy in five games - Figure 4\n\n3. Optimal sizes of models and training times have been found in Table 6 to achieve the highest possible mean scores\n\n4. The presented experimental results show the necessity of temporal consistency in Figure 7" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors introduce model-based method: SGF (a Simple, Good, and Fast) world model.\nThe simplicity of the model is that: it stacks 4 previous frames and actions to capture short-time dependencies, and it uses data augmentation to enhance robustness. It is trained to provide maximum information and temporal consistency.\nFor simplicity, it does not use: image reconstructions, discretization of representations, sequence models (RNN, transformers, ...), probabilistic predictions for deterministic environments.\n\nThey use several loss functions:\n- MSE-loss for temporal consistency (between next embedding and action-conditioned next embedding)\n- Variance and Covariance regularization loss for Information Maximization (use the current and next batches of embeddings as input)\n- Loss for 3 distributions for Dynamics learning: \n - transition - probability of transition to y'-representation if a-action is applied to y-representation (gradient is only propagated to a-action)\n - reward - probability of r-reward if a-action is applied to y-representation\n - terminal - probability of e-terminal-action if a-action is applied to y-representation\n\nThey train several models using AdamW optimizer: \n- Encoder (f computes representations from current and next input image observations): 4 convolutional layers (kernel size 4, stride 2, pad 1), linear layer dim = 512, norm-layer, SiLU activation\n- Projector (g computes current and next embeddings from current and next representations): MLP with 2 hidden layers of dim = 2048\n- Predictor (h predicts action-conditioned next embedding): MLP with 2 hidden layers of dim = 2048\n- Transition network: MLP with 5 hidden layers of dim = 1024, and a residual connection from the input to the output (suggested to use the sequence model for future work)\n- The networks of the reward distribution, terminal distribution, policy, and value function: MLPs with 2 hidden layers of dim = 1024\n\nThis simple approach achieves shorter training times compared to other world models and good performance on the Atari 100k benchmark." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While optimal sizes of models and training times have been found in Table 6 to achieve the highest possible mean scores, this may mean that either the scalability of the approach is limited, or the approach must scale in many directions simultaneously to achieve even higher mean scores. Although the approach lies on the Pareto optimality curve on the Accuracy vs Training time chart, i.e. it is one of many optimal options, it is not shown how this approach can be scaled or improved to achieve the highest accuracy with increasing Training time. \nOr it requires more serious architectural changes, f.e. as the authors suggest for future work - to use the sequence model for transition network.\n\n2. The proposed method is optimal for tasks that are probably simple enough and do not require remembering very old events, so it is sufficient to have a stack of 4 previous frames and actions, and it is not shown how this approach can be transferred to more complex tasks where orders of magnitude more memory of early events is required." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Considering its simplicity, it would be nice and help the paper strongly if scalability can be demonstrated. Can it be scaled?" }, "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 contains thorough experiments.\n2. The method is simple and fast." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a world model that uses self-supervised representation learning, captures short-time dependencies through frame and action stacking, and enhances robustness against model errors through data augmentation. This paper is based on a partially observable Markov decision process. It stacks the recent observations and actions to capture short-time dependencies. It introduces stochasticity through data augmentation. To build meaningful representations of observations, this paper enforeces information maximization and temporal consistency of the features." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The strategy of stacking observations and actions seem overfitting to the chosen dataset, Atari 100k benchmark, rather than fitting to the real-world, where long-term dependencies are common.\n2. Related to the first weakness, while the games in the dataset are deterministic, real-world can be very stochastic. It is questionable whether the model can be applied in real-world cases." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "What are the essential components of world models? We present a simple, good, and fast world model, and evaluate it on the Atari 100k benchmark." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024simple,\ntitle={Simple, Good, Fast: Self-Supervised World Models Free of Baggage},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yFGR36PLDJ},\nnote={under review}\n}" }, "abstract": { "value": "What are the essential components of world models? How far do we get with world models that are not employing RNNs, transformers, discrete representations, and image reconstructions? This paper introduces SGF, a Simple, Good, and Fast world model that uses self-supervised representation learning, captures short-time dependencies through frame and action stacking, and enhances robustness against model errors through data augmentation. We extensively discuss SGF’s connections to established world models, evaluate the building blocks in ablation studies, and demonstrate good performance through quantitative comparisons on the Atari 100k benchmark. The source code will be made 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": [ "Reinforcement learning", "World models", "Self-supervised learning", "Atari 100k" ] }, "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/ba3b0a1e4449258d1a7a5e67d2c6265328fb8ca9.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": "Simple, Good, Fast: Self-Supervised World Models Free of Baggage" }, "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 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": "* The DPA loss function is an innovative approach that mitigates object hallucinations by targeting phrase-level distinctions, offering a focused solution for multimodal hallucination issues.\n\n* The data augmentation method is straightforward yet effective, generating hallucinated-correct response pairs that enable the model to learn nuanced differences with minimal complexity.\n\n* DPA demonstrates significant performance gains" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Data-augmented Phrase-level Alignment (DPA), a novel loss function designed to reduce object hallucinations in multimodal large language models (MLLMs) while preserving their general vision-language capabilities. By generating pairs of hallucinated and correct responses through data augmentation, DPA trains MLLMs to distinguish hallucinated phrases from correct ones. Experimental results show that MLLMs fine-tuned with DPA achieve significant improvements, reducing hallucination rates and enhancing performance on visual question-answering and image description tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The core idea of the paper is to generate correct-hallucinated data pairs through data augmentation. However, I have three questions about this process.\n1. In hallucination-related datasets, object hallucination does not frequently occur, raising questions about the validity of replacing every possible object and attribute with hallucinations. Since models seldom generate such hallucinations, this augmentation strategy might introduce excessive \"non-realistic\" hallucination cases, leading to a mismatch between training and real-world distributions, potentially impacting the model's generalization.\n2. The method’s effectiveness may be limited by the diversity of data augmentation. Since hallucinated data generation relies on a finite set of replacements, it may not fully cover the types of hallucinations that could appear in practical applications, limiting the model’s ability to handle unseen hallucinations.\n3. The data augmentation strategy itself lacks independent experimental evaluation. The experiments mainly focus on improvements in model performance across different benchmarks, without assessing the augmentation strategy’s generalization effect and impact on model training stability across 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": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How does DPA perform on other types of hallucinations beyond object hallucination, such as attribute or location hallucinations?\n2. The reported results for some baseline methods, such as VCD, differ from those in the original papers. Did you directly test VCD, or were the results extracted from their papers? If the latter, the comparison may not be entirely fair, as it mixes experimental results with reported findings from other sources.\n3. This paper augments the training data with hallucinated responses by substituting terms in both open-set and closed-set cases. However, simple substitution could potentially impact fluency and grammatical accuracy. Could this approach compromise data quality and, in turn, affect model performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper proposes an effective alignment method that successfully mitigates object hallucinations, showing improved scores across hallucination benchmarks in both discriminative and generative tasks.\n2. DPA reduces object hallucinations without impacting the overall performance on VQA tasks, achieving comparable or even higher scores on VQA benchmarks.\n3. The paper provides a variety of quantitative results across multiple benchmarks, including both generative and discriminative tasks. This breadth of evaluation provides some evidence of the DPA’s effectiveness in certain contexts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a Data-Augmented Phrase-Level Alignment (DPA) approach aimed at reducing object hallucinations in Multimodal Large Language Models (MLLMs). The method centers on generating paired “correct” and “hallucinated” responses using data augmentation, which are then used to train a phrase-level alignment loss that reduces the probability of hallucinated tokens. The authors strive to maintain the model’s overall vision-language capabilities while minimizing hallucinations. Experimental results across multiple benchmarks indicate that DPA effectively mitigates hallucinations and may even improve detailed object coverage in generated descriptions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The DPA approach offers limited novelty beyond existing finetuning and data augmentation techniques. While phrase-level alignment is applied in a new way here, it basicly builds on existing concepts and does not significantly advance the field of hallucination mitigation research.\n2. Although the results include some competitive baselines, such as HA-DPO and EOS, several relevant and recent methods are omitted. A more comprehensive comparison would strengthen the evaluation. Additionally, some detailed results for LLaVA-13B and VILA are missing, and the selection of methods across different benchmarks lacks consistency.\n3. While the paper asserts that DPA preserves general vision-language capabilities, the supporting evidence is limited. A broader evaluation across diverse benchmarks would help determine whether this approach impacts overall performance.\n4. The authors highlight the limitations of existing methods, noting they “require massive training data.” However, the proposed DPA also introduces additional training requirements, which suggests a tradeoff between efficiency and effectiveness." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "1. VILA is also based on LLaVA-1.5 SFT. Is DPA equally effective on other architectures, such as Qwen?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The writing of this paper is excellent and very detailed. The experiments are comprehensive, covering most of the popular benchmarks.\n\n2. I really like this paper. Many works that use DPO-like methods to reduce hallucinations experience a decrease in VQA capabilities. The authors identified this issue and proposed a specialized loss to maintain the model’s performance while penalizing hallucinated phrases. Additionally, the authors validated their method separately on non-hallucination benchmarks, such as VQA-v2 and TextVQA, demonstrating its effectiveness. I believe this work makes a significant contribution to reducing hallucinations in MLLMs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel method called Data-augmented Phrase-level Alignment (DPA) to mitigate object hallucinations in multimodal large language models (MLLMs) for vision-language tasks. DPA generates pairs of “hallucinated” and “correct” responses to fine-tune the model, reducing the generation of hallucinated phrases. A KL divergence regularization term is added to retain the model’s general capabilities. Experimental results demonstrate that models trained with DPA exhibit significant improvements in hallucination mitigation and maintain strong performance on general tasks across multiple benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. DPA relies on the quality of the generated “hallucinated-correct” response pairs. If these generated data lack accuracy or diversity, it may affect the model’s training effectiveness and generalization capability.\n2. Although the experimental results demonstrate the effectiveness of DPA, the paper lacks a fine-grained analysis of hallucination types (such as objects, attributes, actions). Such analysis could provide a deeper understanding of the method’s performance across different types of hallucinations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No ethics review needed." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Minors:\n1. The construction of dataset is not introduced with details. Details in appendix should be at least provided via cross-reference.\n2. Figure 5 right is not illustrated in the main text. I find it not easy to understand the information it's trying to convey." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. It is critical to study the issue of hallucination in MLLMs as well as the trade-off phenomenon in the mitigation of hallucination.\n2. The proposed method is simple yet effective.\n3. Extensive experiments verify the effectiveness of the method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose data-augmented phrase-level alignment to mitigate object hallucination in MLLMs. The method mainly involves the generation of negative hallucinated responses and phrase-level fine-tuning. The hallucination issue of models fine-tuned with the proposed method is alleviated and the general performance is maintained." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation to mitigate hallucination at phrase level is not clearly addressed. There is no illustration on why the phrase-level loss can retain the original performance on general multimodal tasks. It seems straightforward that the constraint on the KL divergence can prevent the fine-tuned model from diverging too far. If it is the only reason, the contribution of the method is greatly weakened.\n\n2. The explanation of Figure 2 is likely to be overclaimed. On line 260, it says \"_EOS achieves a slightly lower hallucination rate_\", but the figure shows that the hallucination rate of EOS is around 5.0 and that of HALVA is around 6.5. This difference is noticeable enough to me as the gap of this metric between HALVA and LLaVA-1.5 is similar. Meanwhile, Figure A right demonstrates that HALVA has a much higher F1 score on AMBER, which is natural because neither EOS nor HA-DPO is trained with Yes/No questions.\n\n3. The presentation of results in experiments is inconsistent. The model lists are different in different tables. For instance, EOS-13B is only shown in Table 1, which makes the verification of the effectiveness less convincing.\n\n4. There are other work mitigating hallucinaton with sub-sequence level training [1]. It is recommended to discuss the difference.\n\n[1] Gunjal, Anish et al. “Detecting and Preventing Hallucinations in Large Vision Language Models.” AAAI Conference on Artificial Intelligence (2023)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce phrase-level alignment method that can be applied to off-the-shelf MLLMs for mitigating hallucinations, while preserving their general vision-language capabilities." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024dataaugmented,\ntitle={Data-Augmented Phrase-Level Alignment for Mitigating Object Hallucination},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yG1fW8igzP},\nnote={under review}\n}" }, "abstract": { "value": "Despite their significant advancements, Multimodal Large Language Models\n(MLLMs) often generate factually inaccurate information, referred to as hallucination.\nIn this work, we address object hallucinations in MLLMs, where information\nis generated about an object not present in the input image. We introduce Dataaugmented\nPhrase-level Alignment (DPA), a novel loss which can be applied to\ninstruction-tuned off-the-shelf MLLMs to mitigate hallucinations, while preserving\ntheir general vision-language capabilities. To fine-tune MLLMs with DPA, we first\ngenerate a set of 'hallucinated' and 'correct' response pairs through generative data\naugmentation by selectively altering the ground-truth information of the correct\nresponses at a phrase level. The DPA loss is then used to train MLLMs to reduce\nthe likelihood of hallucinated phrases compared to the correct ones. Our thorough\nevaluation on various benchmarks confirms the effectiveness of DPA in mitigating\nhallucination while retaining the out-of-the-box performance of the MLLMs on\ngeneral tasks. For instance, MLLMs finetuned with DPA, which we refer to as Hallucination\nAttenuated Language and Vision Assistant (HALVA), improve F1 by up\nto 13.4% on hallucination visual question-answering and reduce the hallucination\nrate by up to 4.2% on image description 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": [ "Multimodal LLMs", "Object Hallucination", "Vision-language Models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/101e55289d27641711778b85d109b1c275b35374.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": "Data-Augmented Phrase-Level Alignment for Mitigating Object Hallucination" }, "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 the weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper presents an innovative quantization method (BSQ) that addresses the limitations of existing vector quantization approaches by offering a more efficient and scalable solution.\n2. Extensive experiments on benchmarks such as ImageNet and UCF-101 demonstrate that BSQ-ViT significantly improves reconstruction quality, outperforming prior methods in terms of speed and fidelity.\n3. The methodology is clearly explained with detailed comparisons to related work, and the theoretical basis of BSQ is well-supported with mathematical derivations and experimental validation.\n4. BSQ-ViT's ability to handle both image and video tokenization and perform well on diverse tasks such as compression and generation showcases its generalizability." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a novel image and video tokenizer, BSQ-ViT, based on Binary Spherical Quantization (BSQ) integrated with a transformer architecture. The proposed method achieves state-of-the-art performance in image and video reconstruction with significant improvements in computational efficiency. It introduces a block-wise causal masking for video tokenization, supports variable-length videos, and delivers competitive results in visual compression against standards like JPEG2000 and H.264. BSQ-ViT also shows promising results in image generation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the transformer architecture is explored, the paper does not demonstrate the effectiveness of BSQ within a CNN-based model.\n2. The paper provides limited comparative data in video reconstruction, reducing the robustness of the comparison. Additionally, while block-wise causal attention is noted to impact performance, the study lacks experiments on BSQ without this causal masking. \n3. The reported image and video compression results are better on MS-SSIM, potentially due to the inclusion of perceptual losses like LPIPS and GAN loss in the loss function. However, the lack of subjective evaluation metrics limits the strength of the comparison.\n4. When BSQ and VQ use the same number of indexes, BSQ performs less effectively, and codebook utilization decreases as the bits increase.\n5. In the ablation experiments, the reproduction of the LFQ, which serves as the basis for this work, is too poor in terms of behavior and code usage. And it is premature to conclude that LFQ is less compatible with transformer-based codecs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "See the weakness part" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The Binary Spherical Quantization seems to show more effective training of the qunatization bottleneck. Analysis shows that the proposed method can provide fast speed and good performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper propose a transformer-based image and video tokenizer with Binary Spherical Quantization, which projects the high-dimensional embedding to lower-dimensional hypersphere and applies binary quantization. BSQ is parameter-efficient without an explicit codebook, scalable to token dimensions and compact. The experiments show that the proposed method achieves comparable compression performance with JPEG2000/WebP for images and H.264/H.265 for videos. And it enables masked language models to achieve competitive image synthesis quality to GAN and diffusion methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of comparison at different bitrate range for visual compression results. Table 4 only provides BPP, PSNR and MS-SSIM for one bitrate point. However, visual compression tasks usually require showing a Rate-Distortion curves and compare at different bitrate points. Your can use BD-Rate metric for more reasonable comparison and analyze the results at low bitrate and high bitrate.\n\n- Test settings for ablation study. Please provide more experiment setting details. In Table 5, do VQ, LFQ and BSQ use the same tokenizers? I observed that the metrics vary significantly across different quantization methods. To alleviate the influence of training and focus on the quantization bottleneck, it is better to: 1. use the same tokenizer 2. freeze the other network parts and only train the information bottleneck parts.\n\n- Complexity. It is better to provide the computation complexity and encoding/decoding time for comparison." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. For image compression, why not use arithmetic coding to improve compression performance? Is intra-frame information used for video compression? If not, why not?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The idea of projecting high-dimensional visual embeddings onto a lower-dimensional hypersphere is straightforward yet effective.\n\n2. The motivation is clear, and the overall presentation is coherent and easy to follow. The experiments are comprehensive and provide convincing evidence to support the approach.\n\n3. The BSQ-ViT model achieves competitive performance in diverse tasks such as image/video reconstruction, generation, and compression." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a transformer-based image and video tokenizer that uses Binary Spherical Quantization (BSQ). By projecting high-dimensional visual embeddings onto a lower-dimensional hypersphere and applying binary quantization, BSQ achieves a bounded quantization error. The authors demonstrate that the proposed BSQ outperforms previous methods in image and video reconstruction, image generation, and compression tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This method uses a transformer encoder and decoder, which limit the flexibility in resolution. How do the authors address this issue?\n\n2. For image and video compression results, it would be beneficial to include an LPIPS comparison to assess perceptual performance.\n\n3. There are a few minor issues. 1) In Eq. (7), $\\hat{q}(c|u) = \\frac{\\exp(\\tau c^{T}u)}{\\sum_{c \\in C_{BSQ}} \\exp(\\tau c^{T}u)}$ might need to be revised to $\\hat{q}(c|u) = \\frac{\\exp(2 \\tau c^{T}u)}{\\sum_{c \\in C_{BSQ}} \\exp(2 \\tau c^{T}u)}$ ? 2) On page 16, line 837, $\\frac{e^{\\tau u_d \\hat{c}_d}}{e^{\\tau u_d \\hat{c}_d} + e^{\\tau u_d \\hat{c}_d}}$ should be corrected to $\\frac{e^{\\tau u_d \\hat{c}_d}}{e^{\\tau u_d \\hat{c}_d} + e^{-\\tau u_d \\hat{c}_d}}$. 3) On page 3, line 118, it would be clearer to use different symbols for the downsample factor $q$ and the bottleneck module $q$." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "My major concerns in the weakness part can be summarized as the direct comparison between BSQ, VQ and LFQ on representative image and video tasks are not adequate. As a general tokenizer for image and video, it is important to make direct comparisons with strictly controlled variables." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Visual tokenizers are critical for visual modeling. The proposed Binary Spherical Quantization (BSQ) for unified image and video tokenization is novel and important. \n\nTheoretical contribution is good. It is proved that the proposed BSQ has bounded quantization error. \n\nThe manuscript is well written. BSQ is well placed agains previous VQ and LFQ. \n\nExtensive experiments on visual reconstruction, compression and generation show the better performance of BSQ." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a visual tokenizer based on a Vision Transformer and Binary Spherical Quantization (BSQ). \nThe Transformer-based encoder-decoder leverages a block-wise causal mask and uses only visual tokens from the current or past timestamps for reconstruction. \nBSQ first projects the high-dimensional visual embedding of the encoder to a lower-dimensional hypersphere and then applies binary quantization. \nThe transformer encoder, decoder, and BSQ are seamlessly integrated into the VQ-GAN framework and trained end-to-end. \n\nThe proposed visual tokenizer has several advantages in trading off visual reconstruction quality and computational efficiency and supporting variable-length input. Experiments on visual reconstruction and compression are conducted to verify the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My major concern is the fair comparison with VQ and LFQ. Though the ablation study is provided in tab 5 with image reconstruction task on imagenet-1k, the resolution is only 128, and more tasks should be tested: \n\nIt seems to me the results on image is not as obvious as video. For example, in image reconstruction (tab 1), the proposed method has best image quality, but the parameter number is also larger. It is better to add Ours-VQ for better comparison, as done in tab 2 for video reconstruction. \n\nIn image compression, no VQ or LFQ based method is compared. \n\nFor image generation in tab 3, more steps are used for BSQ, what if we use the same steps as VQ and LFQ? \n\nIn L389, we get a direct comparison between VQ and BSQ, BSQ has more bits (18 vs 14) and the video quality is only comparable with VQ. \n\n \n\n\n\nMinor \nThe derivation for eq. 13 is not given. Please provide a proof or give a reference here. \nThere exist some typos in the derivations. \nIn Line 878, the p shuld be Q. \nIn line 886, the p should be q^hat." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024image,\ntitle={Image and Video Tokenization with Binary Spherical Quantization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yGnsH3gQ6U},\nnote={under review}\n}" }, "abstract": { "value": "We propose a new transformer-based image and video tokenizer with Binary Spherical Quantization (BSQ). BSQ projects the high-dimensional visual embedding to a lower-dimensional hypersphere and then applies binary quantization. BSQ is (1) parameter-efficient without an explicit codebook, (2) scalable to arbitrary token dimensions, and (3) compact: compressing visual data by up to 100×\nwith minimal distortion. Our tokenizer uses a transformer encoder and decoder with simple block-wise causal masking to support variable-length videos as input. The resulting BSQ-ViT achieves state-of-the-art visual reconstruction quality on image and video reconstruction benchmarks with 2.4× throughput compared to the best prior methods. Furthermore, by learning an autoregressive prior for adap-\ntive arithmetic coding, BSQ-ViT achieves comparable visual compression results with commonly used compression standards, e.g. JPEG2000/WebP for images and H.264/H.265 for videos. BSQ-ViT also enables masked language models to achieve competitive image synthesis quality to GAN and diffusion 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": [ "quantization", "visual compression", "visual generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1b455b6708f0254eb7cbde86577b974c8632a9a4.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": "Image and Video Tokenization with Binary Spherical Quantization" }, "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": "Check Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ Combining self-attention for inter-patch dependencies, diffusion mechanisms for intra-patch dependencies, and auto-regressive optimization is an interesting and innovative approach to time series forecasting.\n+ The diffusion-based reverse process for reconstructing the sequence is novel in the context of time series forecasting.\n+ Good writing and organization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces TimeDART (Diffusion Auto-Regressive Transformer for Time Series Forecasting), a novel self-supervised learning framework designed to enhance time series forecasting. TimeDART addresses key challenges in the field, particularly capturing both long-term dependencies and local features within time series data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "+ The combination of Transformer-based attention mechanisms with the denoising diffusion process introduces substantial computational overhead. The paper mentions running experiments on a single NVIDIA RTX 4090 GPU, but it doesn't provide detailed insights into the time and memory consumption required for training. How scalable is TimeDART for larger datasets or real-time applications? \n+ The experiments conducted on noise scheduling, the number of diffusion steps, and the number of layers in the denoising network highlight a significant degree of sensitivity to hyperparameter selection. How the model will perform well in practical settings with minimal tuning?\n+ The cross-domain evaluation shows strong results on energy datasets but weaker performance on datasets like Exchange. Is TimeDART overly sensitive to the type of data it is pre-trained on? For instance, does it struggle with financial or highly volatile datasets because of the lack of shared characteristics between domains (e.g., energy vs. finance)? Could this method benefit from domain adaptation techniques to make the cross-domain transfer more robust?\n+ The model uses a denoising diffusion loss, which may not be the most suitable for every type of forecasting task. How does TimeDART perform with other loss functions (e.g., Quantile Loss, Huber Loss) that are often used in time series forecasting tasks where the goal is to forecast confidence intervals or robustly handle outliers?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "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": "Please see weaknesses." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. We propose a novel generative self-supervised learning framework, TimeDART, which in tegrates diffusion and auto-regressive modeling to effectively learn both global sequence dependencies and local detail features from time series data, addressing the challenges of capturing comprehensive temporal characteristics.\n 2. We design a cross-attention-based denoising decoder within the diffusion mechanism, which enables adjustable optimization difficulty during the self-supervised task. This design significantly enhances the model’s ability to capture localized intra-patch features, improving the effectiveness of pre-training for time series forecasting. Diffusion models and autoregressive attention mechanisms are rare collaborations in temporal tasks, and this field brings new ideas. \n3. The experimental results show that the combination of Mamba and propagation mechanism is very effective, and it also exceeds the predictive performance of supervised learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Recently, effectively capturing both the global sequence dependence and local detail features within time series data remains challenging. To address this, the authors propose a novel generative self-supervised method called TimeDART, denoting Diffusion Auto-regressive Transformer for time series forecasting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The reviewer is concerned about the computational overhead associated with this approach. The reviewer's core concern stems from the introduction of diffusion mechanisms, which can lead to a substantial increase in training and inference overhead for the model as a whole. If the model is expensive, the computational conditions required to solve the real task will be severe. Therefore, the authors need to report the actual time of the training and inference phase of other baseline models in the future and report the GPU and server model.\n2. The motivation for choosing to use a causal mechanism in Transformer requires further explanation. After all, time series data is encoded with more complex patterns. In particular, there are random changes caused by extreme weather events in the meteorological data, and this causal relationship is strong. But many things cause and effect is unclear, so whether such a component is appropriate needs to be used for the specific 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": { "value": "No" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Well-written\n2. The author provides the code\n3. The performance of the model is proved by experiments" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To effectively capture both the global sequence dependence and local detail features within time series data, this paper proposed a novel generative self-supervised method called TimeDART, denoting Diffusion Auto-regressive Transformer for Time series forecasting. Extensive experiments demonstrate that TimeDART achieves state-of-the-art fine-tuning performance compared to the most advanced competitive methods in forecasting tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In the original paper [1], the performance of patchTST seems to be better, and it is suggested that the author should evaluate it more fairly.\n2. Why didn't the author consider evaluating the performance of classification tasks? Currently, aside from forecasting tasks, most self-supervised learning models [2] also focus on the performance of classification tasks. This is because the main purpose of self-supervised learning is to enhance the quality of representations generated by the model and to uncover key semantic features, which is especially important for classification tasks.\n3. A crucial role of self-supervised learning is to improve the performance of backbone models [3], but the author only uses Transformer as the backbone. I am curious whether the proposed method would still be effective if MLP or TCN were used as the backbone.\n4. The author's core motivation remains to capture both global and local dependencies, which is similar to most existing works [1] [4]. In other words, this paper lacks a central challenging problem, making the contribution at the motivational level somewhat limited.\n5. Considering that the core motivation of this paper is to capture global and local dependencies, I suggest the author evaluate the model's performance on datasets with stable patterns, such as PEMS04. This is because datasets like ETT and Exchange have inherent issues with distributional shifts [5].\n\n[1] A time series is worth 64 words: Long-term forecasting with transformers\n\n[2] SimMTM: A Simple Pre-Training Framework for Masked Time-Series Modeling\n \n[3] Cost: Contrastive learning of disentangled seasonal-trend representations for time series forecasting\n\n[4] Segrnn: Segment recurrent neural network for long-term time series forecasting\n\n[5] Exploring Progress in Multivariate Time Series Forecasting: Comprehensive Benchmarking and Heterogeneity Analysis" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "**Question 1:** \n\nThis paper uses an autoregressive prediction paradigm, however autoregressive approaches generally suffer from error accumulation and high inference time overheads. To the best of our knowledge, TimeDART has not designed a special training-inference strategy or model architecture to mitigate these problems. Further discussion of TimeDART's limitations in the autoregressive prediction paradigm is urgent and necessary.\n\n**Question 2:** \n\nAs the authors say ‘we adopted the channel-independence setting’. However, for datasets with a very large number of channels, such as ECL and Traffic, how can a channel-independent design establish connections between multiple channels? Meanwhile, the experimental results of Crossformer[10], SAMformer[11] and iTransformer[12], show that cross-channel modelling is crucial for multivariate time series, and we believe that a discussion on inter-channel modelling for multivariate time series is necessary. However, to the best of our knowledge, TimeDART does not consider these factors in its modelling.\n\n**Question 3:** \n\nSee Weakness 2,3 for our concerns about the experimental results, and we believe that the introduction of competitive and up-to-date baselines is considered necessary. In addition, the authors are expected to explain the phenomena in the ablation experiments, details of which can be referred to Weakness 4.\n\n**Reference:** \n\n10) Zhang, Yunhao and Junchi Yan. “Crossformer: Transformer Utilizing Cross-Dimension Dependency for Multivariate Time Series Forecasting.” ICLR 2023.\n\n11) Ilbert, Romain et al. “SAMformer: Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention.” ICML 2024.\n\n12) Liu, Yong et al. “iTransformer: Inverted Transformers Are Effective for Time Series Forecasting.” ICLR 2024." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors outline the main approaches to self-supervised learning in the time series domain, including mask reconstruction and contrast learning, and analyses the shortcomings of the two existing self-supervised learning paradigms separately, e.g., the mask reconstruction-based approach introduces a huge gap between pre-training and fine-tuning, and the contrast-learning-based approach prioritizes the capture of discriminative features, which lead to a huge discrepancy between pre-training tasks and fine-tuning tasks.\n\nIn addition, the authors raise two issues critical to the self-training paradigm, 1) how to narrow the gap between the pre-training target and the downstream fine-tuning task, and 2) modelling both long-term dependencies and local pattern information in the self-supervised pre-training phase. However, we believe that TimeDART is not the best solution to these problems." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes TimeDART, a novel self-supervised learning paradigm that hopes to learn transferable generic representations from unlabeled data through pre-training and then fine-tune them to different downstream tasks, e.g., forecasting tasks, etc. In this paper, a unified self-supervised learning framework is established by fusing the diffusion-denoising process and autoregressive modeling. By learning global sequence dependencies and local detail features in multi-domain time series data, the model's ability to capture comprehensive time series features and cross-domain generalization ability are improved. Specifically, TimeDART designs a cross-attention based denoising decoder in the diffusion mechanism, which improves the effectiveness of time series pre-training by significantly enhancing the model's capability to capture features within local blocks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Weakness 1:** \n\nAs a Diffusion-based model, we would like to introduce more valuable metrics to fully examine the performance of the proposed TimeDART. Specifically, in order to evaluate the prediction accuracy and generalization capability of TimeDART from both forecasting and generation perspectives, existing studies often include the following four metrics (including Context-FID, Correlational Score, Discriminative Score, and Predictive Score) for comprehensively evaluating the performance of Diffusion methods. \n\nIn addition, the proposed TimeDART is not compared with advanced Diffusion-based approaches, such as Diffusion-TS[1], mr-Diff[2], and MG-TSD[3]. We believe that the introduction of more competitive and up-to-date approaches can demonstrate the effectiveness of the proposed method more objectively.\n\n**Weakness 2:** \n\nIn the in-domain setting (Table 2), there are the following weaknesses: \n\n* The performance improvement of the proposed TimeDART over SOTA methodologies SimMTM and PatchTST is less than 5%, which indicates that the performance improvement of the model is not obvious.\n\n* In addition, we note that the performance of \"Random Init\" shown in Table 1 is slightly worse than Supervised PatchTST. Does this indicate that in the supervised setting of a single domain, the proposed model architecture exhibits worse performance compared to PatchTST? If the modeling capability of the TimeDART is poor in small datasets, the model will often show worse generalization ability in cross-domain pre-training, which leads to the rationality of the model architecture being questioned.\n\n**Weakness 3:** \n\nIn the cross-domain setting (Table 3), there are the following weaknesses that can be improved: \n\n* The model is pre-trained on only two power domain datasets (ETT and Electricity). As a result, only the single domain information is included in the model, which limits the generalization ability of the model under cross-domain challenges. Recent unified time series forecasting models include two paradigms. The first is unified models based on LLM fine-tuning, such as OneFitsAll[4] and TimeLLM[5]. This is followed by pre-training on a multi-domain hybrid Time series dataset followed by fine-tuning on specific downstream tasks, e.g., Timer[6], Moriai[7] and MOMENT[8]. In conclusion, we believe that introducing information from more domains during pretraining can improve the cross-domain generalization of the model, and TimeDART is expected to be pretrained on a wider range of datasets. \n\n* Table 3 only shows the performance comparison of TimeDART under different Settings; however, it lacks the comparison with the latest baseline. In fact, recent UniTime[9] have achieved joint pretraining across multiple domains. Therefore, we expect the authors to introduce more advanced baselines to compare with TimeDART, which will help us get a comprehensive understanding of TimeDART's performance.\n\n**Weakness 4:** \n\nIn ablation experiments (Table 4), there are the following drawbacks: \n\n* There is a lack of detailed descriptions specific to ablation experiments, such as in-domain Settings or cross-domain Settings. When we compare the results in Table 4 and Table 2, we can speculate that the ablation experiment is only carried out in the in-domain setting. However, in the cross-domain setting, the reader is eager to know whether the proposed autoregressive diffusion model is effective. \n\n* In Table 4,\" The \"W/o AR\" model obtained the improved \"W/o AR\" model after introducing the Diffusion-based decoder; However, the performance of the latter was slightly degraded on the ETTh2 and Electricity datasets. This may indicate that the predictions of the model become worse when the diffusion model is introduced. This casts doubt on the rationality of introducing a Diffusion-Denoising process in the Decoder.\n\n**Reference:** \n\n1) Yuan, Xinyu and Yan Qiao. “Diffusion-TS: Interpretable Diffusion for General Time Series Generation.” ICLR 2024.\n\n2) Shen, Lifeng et al. “Multi-Resolution Diffusion Models for Time Series Forecasting.” ICLR 2024.\n\n3) Fan, Xinyao et al. “MG-TSD: Multi-Granularity Time Series Diffusion Models with Guided Learning Process.” ICLR 2024.\n\n4) Zhou, Tian et al. “One Fits All: Power General Time Series Analysis by Pretrained LM.” NIPS 2023.\n\n5) Jin, Ming et al. “Time-LLM: Time Series Forecasting by Reprogramming Large Language Models.” ICLR 2024.\n\n6) Liu, Yong et al. “Timer: Generative Pre-trained Transformers Are Large Time Series Models.” ICML 2024.\n\n7) Woo, Gerald et al. “Unified Training of Universal Time Series Forecasting Transformers.” ICML 2024.\n\n8) Goswami, Mononito et al. “MOMENT: A Family of Open Time-series Foundation Models.” ICML 2024.\n\n9) Liu, Xu et al. “UniTime: A Language-Empowered Unified Model for Cross-Domain Time Series Forecasting.” Proceedings of the ACM on Web Conference 2024." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A novel generative self-supervised learning framework for time series forecasting that simultaneously models global sequence dependencies and captures local detail features effectively." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024diffusion,\ntitle={Diffusion Auto-regressive Transformer for Effective Self-supervised Time Series Forecasting},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yGv5GzlBwr},\nnote={under review}\n}" }, "abstract": { "value": "Self-supervised learning has become an essential and popular approach for enhancing time series forecasting, enabling models to learn universal representations from unlabeled data. However, effectively capturing both the global sequence dependence and local detail features within time series data remains challenging. To address this, we propose a novel generative self-supervised method called TimeDART, denoting Diffusion Auto-regressive Transformer for Time series forecasting. In TimeDART, we treat time series patches as basic modeling units. For one thing, we employ an self-attention based Transformer encoder to model the dependencies of inter-patches. For another, we introduce diffusion and denoising mechanisms to capture the locality features of intra-patch. Notably, we design a cross-attention-based denoising decoder that allows for adjustable optimization difficulty in the self-supervised task, facilitating more effective self-supervised pre-training. Extensive experiments demonstrate that TimeDART achieves state-of-the-art fine-tuning performance compared to the most advanced competitive methods in forecasting tasks. Our code is publicly available at https://anonymous.4open.science/r/TimeDART-2024." }, "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-supervised Learning", "Diffusion Model", "Time Series Forecasting" ] }, "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/30eae5f9d0f0fc3dced313dfde7fd34367baef46.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": null, "title": { "value": "Diffusion Auto-regressive Transformer for Effective Self-supervised Time Series Forecasting" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Questions:\n1. In some numerical results, RPC-BC shows an error floor in the high SNR region but LightBC does not, which is described by \"LightBC behaves somewhat like a linear feedback code\" (Line448). The reviewer wonders the reason of the fact because RPC-BC must be more flexible than LightBC. A possible reason is the hardness of learning RPC-BC. Is there any explanation of the fact? \nAnother question: why LightBC can behaves like linear codes? Does learned LightBC really have some linearlity like a linear code? How to examine it? \n\n2. Is there any possible reason that LightBC outperforms RPC-BC in some settings? Is it related to the hardness of learning models? \n\n3. In Fig.3, why LightBC by federated learning performs close to LightBC by global learning? \n\n4. What is the benefit of fully-learnanable AWGN-BC code? \n\n5. Are the training parameters in Tab. 1 optimized for AWGN-BC? The reviewer is concerned that the use of parameters \"consistent with the training parameters proposed in the single user versions\" (Line 408) may be unsuitable for the AWGN-BC case, resulting in poor performance of RPC-BC.\n\nSuggestions:\n1. The size of trainable parameters should be described. It will be helpful to show how much LightBC is lighter than RPC-BC. \n2. In Line 232, $\\mathbb{R}^{N_s,2}$ should be $\\mathbb{R}^{N_{s,2}}$. \n3. In Line 394, \"The decoder sends the length $|W_\\ell|$−length vector...\" is somewhat confusing.\n4. In Fig. 2(a), the term \"SU\" is confusing because it is denoted by \"TDD\" in the main text." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "* The proposed models are the first fully-learnable architectures for an AWGN-broadcast channel (AWGN-BC). They show excellent performance compared with a conventional concatenated linear code. \n* The authors investigate a federated learning strategy for training weights in the models. In particular, they examine the effect of noise in transmitting gradients in the training process.\n* These issues are of importance in terms of future wireless communications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript proposes two classes of deep-learning assisted encoder and decoders for a broadcast channel. The first is an extension of the RNN-based architecture, and the second is a lighter MLP-based architecture. The authors compare them with conventional concatenated codes in various settings, indicating that the proposed architectures outperform conventional codes. In addition, they consider the federated learning scheme for the models and examine the performance when the gradient of trainable weights is transmitted over noisy channels." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although the reviewer agrees with the importance of the issues, there are several flaws in this manuscript.\n1. **Poor technical contributions** \n\nThe contributions of this manuscript are twofold: the proposal of deep-learning architectures for AWGN-BC and the proposal of federated learning for these models. However, the proposed architectures are simple extensions of existing models for single-user cases. In short, the modifications are only learning multiple decoders and changing the loss function. As for federated learning, the concept is familiar and the modifications of the learning process are straightforward. Including not providing any theoretical analyses, the reviewer cannot avoid judging that the contributions of this manuscript are poor. \n\n2. **Lack of comparison between proposed models**\n\nThe authors proposed two models for AWGN-BC. It is claimed that LightBC is lighter than RPC-BC, but the authors do not explicitly compare the number of trainable weights and/or training costs of these models. This will be important because the comparison of model size is a performance metric other than an error rate. In addition, it will be related to the analysis of the numerical results, e.g., Fig. 3.\n\n3. **Lack of comparison with other models**\n\nAnother issue is that the numerical results do not contain those of other learning-based AWGN-BC codes such as [Li et al. 2022] in the manuscript. Even if Li's model is not fully learnable, as stated in Sec. 1.1, there is no reason to omit the model from comparison in this manuscript. \n\n4. **Possibly insufficient numerical experiments on federated learning**\n\nThe authors state, \"We choose to explore the uncoded downlink versus a quantized method as it has been shown that there is better convergence behavior in federated learning with noisy downlinks Amiri et al. (2021).\" in Line 126-129. However, Fig. 3 only contains the results of uncoded transmission (Line 490). The manuscript seems to be flawed.\n\n5. **No theoretical analysis nor interpretation of numerical results**\n\nThe manuscript contains no theoretical analysis of the proposed models or federated learning process. At the very least, I believe that the interpretation of the numerical results should be described in the manuscript. However, the manuscript only describes the facts from numerical results. See the \"Question\" section for more details." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 the scheme be applied to a block transmission with feedback? I think this scenario is of high importance because cellular systems operate in this regime (hybrid ARQ).\n2. Please provide a more detailed comparison with orthogonal (TDM) schemes when considering noiseless feedback channel. Comparing results of the LightCode paper (arXiv 2403.10751) and Fig. 5 (rate K=3/N=9, and BLER=1e-9 at approximately -1 dB SNR), and results presented in Fig 1. (BLER 1e-6 at the same SNR for K=3/uers and N=18 channel uses).\n3. Please address the problem of error-floor appearing in Figure 1. (RPC, K=3/N=18)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors consider the case of a multiuser broadcast channel with feedback. Feedback is known to improve the capacity region of the multiuser channel without feedback. Moreover, the behavior of this setup in the finite blocklength regime is not addressed in the literature. Thus, the application of machine learning in designing codes for broadcast channels with feedback seems quite reasonable and interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper considers a broadcast channel with feedback and considers a two-user symmetrical AWGN case. Feedback is known to improve the capacity region of this channel, and the authors consider a machine learning-based approach to construct good codes for this setup. The authors consider both noiseless and noisy feedback.\n\nThe proposed training methods are based on the extension of previous works devoted to a single-user channel:\n - Robust Power-Constrained Deep Learning Algorithm (RPC)\n - LightCode\n\nThe authors consider a simulation setup with small block length equal to 9 or 18 channel uses. The authors consider three different scenarios:\n- noiseless feedback,\n- noisy feedback and comparison with an orthogonal (TDD) scheme,\n- noisy feedback and comparison between different coding rates." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper lacks some discussion related to the broadcast channel with feedback. The authors just say that «the use of feedback in the AWGN-BC channel can far exceed the capacity of the AWGN-BC channel without feedback». I think that more discussion on this topic may significantly improve the understanding of the problem and numerical results analysis. The questions that I suggest to address may be as follows:\n- What are existing theoretical results on capacity region?\n- How exactly the capacity region can be improved and in which setups? \n- What about successive interference cancellation that is widely used in communications? Is it applicable to feedback channels? \n- What about finite block length regime? Were there any attempts to address it?\n\nNext, in the final section, the authors mention that \"... numerical studies indicated that there appears to be not much advantage over using the simple extension of LightBC in the broadcast setting versus utilizing TDD with the single user LightCode in the high rate, noisy regime\", which seems a bit confusing. It seems that the improvement can be achieved exactly when a non-orthogonal transmission scheme is used.\n\nThe authors consider numerical results by addressing a finite (N = 18 channel uses) block length, which is too far from any coding scheme used in practice. I suggest adding some discussion on how the block length can be increased and which results one should expect in this case.\n\nThe main numerical results are presented in Fig. 1 (noiseless feedback), and this figure lacks a comparison with any known theoretical bounds (like orthogonal TDM mentioned later when considering noisy feedback channels or a finite block length random coding bound for multiple access channels without feedback) and practical schemes. More references will also improve the understanding of the provided 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": 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": "- Can you please better motivate the FL approach? Is the scenario a distributed communication setting with unknown channels? I believe estimating the channels before training the code would be a better approach in general (as is done in practice). Can you please better motivate the underlying scenario? \n\n- RPC-BC seems to do worse than time-division transmission of point-to-point RPC codes. Then, what is the point of considering these codes? LightBC seems to improve compared to time-division, does that hold for longer codes as well? I expect the training will get harder quickly with the code length, and the gains with respect to TD may quickly diminish. Can you please provide additional simulation for longer code lengths? \n\n- Why consider the code rate of 2/3 in Fig. 1 given that the error probabilities are quite high? \n\n- Can you please provide a comparison of these schemes with the linear coding approaches proposed in the literature? Even though you argue that non-linear coding can help, you do not provide any evidence for that. \n\n- How practical these codes in practice? If I understand correctly, a different neural network decoder is trained for each receiver. In a practical setting, where users roam from cell to cell, does it mean that each user has to have all the decoders as they can be assigned as any user i. Similarly, it seems that a separate encoder/decoder are trained for each SNR, which further increases the memory requirement." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The main strength of the paper is to study a communication scenario that has not received much attention, particularly in terms of the application of recently developed neural network based code designs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies error correction coding over a broadcast channel in the presence of feedback. The goal of the transmitter is to convey independent messages to different receivers using the same channel resources. Assuming that the transmitter can see the channel outputs received by the receivers perfectly or with some noise, its goal is to transmit signals in an iterative fashion to gradually improve the decoding reliability of the receivers of their own messages. The author review the literature in this domain, and emphasize the lack of prior work on deep learning aided code design for this problem, although some works have appeared for the single-user channel with feedback.\n\nThe authors extend two approaches recently proposed for the single-user problem to the broadcast channel setting. They also propose a federated learning based training approach for the code." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The novelty of the paper is very limited. The authors mainly take two existing designs and train these models by considering multiple receivers. This is rather trivial, and as such the paper does not introduce any new concept, architecture or tool. \n\n- The federated learning (FL) approach is not well motivated or explained. Why would a vertical FL approach make sense here? As long as the channel model is available, what prevents the encoder from training all the decoders centrally? Is it a complexity issue? The proposed uncoded transmission of gradients, as also observed by the authors, is limiting, and not well motivated. \n\n- Presentation can be improved. Especially in the numerical results part, there are some confusing sentences. \n\n- Comparison is limited to a single relatively weak code from Li et al. It seems that the state of the art for point-to-point channels with feedback (GBAF codes) is not considered in this scenario. Although complexity is argued against these codes, there is no presentation of complexity for the presented schemes." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Included everything in the weakness section" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Relatively unexplored problem space. While deep learning based feedback schemes are well studied for single user AWGN channels, utility of the same for broadcast channels is not well studied and open for improvements. \n2. The paper is well written overall and easy to follow for someone familiar with the problem space. \n3. The concept of federated training of encoder and decoders is novel and interesting and relevant to practical settings. \n4. Performance gains over existing linear schemes is non-trivial and compelling. \n5. The takeway on LightBC being a better choice in low-noise scenarios and RPC-BC being better in high-noise scenarios is sn interesting observation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a new feedback coding scheme for AWGN broadcast channels (AWGN-BC) by extending two of the exising feedback coding schemes for feedback channels, a RNN based scheme and a lightweight-MLP scheme. Further, a new vertical fedearted learning based training scheme is proposed, which reflects a real world scenario more closely. Finally the authors peroform an emoirial study to demonstrate the benefits of each coding scheme under different channel conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited novelty from an architecture and training perspective, as both RPC and LightCode schemes are being reused directly from existing works. \n2. No system model included in the main body of the paper, making the problem setup hard to grasp.\n3. Throughout the paper, there was the mention of LightCode being simpler than RPC. But there is not study on the complaexity, memory, and run-time comparison between the schemes makaing it harder to understand the difference between the schemes in terms of resource overhead. \n4. Authors mention in introduction that the capacity of AWGN-BC channel increases in presence of feedback. In the experiemnts section, it would be interesting to see some discussion on this and how the feedback is impacting the performance. Specifically, a reference line indicating the capacity of the channel would make it easier to comprehened the efficiency of the proposed scheme compared to the theoretical limits. \n5. Captions to the figures are rather short and very informative. I suggest updating all the captions to reflect the key takeaway. \n6. Experimental section is written poorly. It's not immediately apparent what authors mean by \"Lin.\" in the plots. More discussion should go into explaining the figures well.\n7. Lack of proper baselines. Since the authors propose modifying two schemes that were originally proposed for single user case, there should be atleast one more learning-based baseline that was designed for 2-user case. The reference Li. et al. \"Deep learning-aided coding for the fading broadcast channel with feedback\" would be a good baseline (despite the BPSK modulation).\n\nMisc.\nMinor typos and grammatical errors should be corrected throughout the paper. ex. in line 147 - \"receives\"" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Deep Learning is applied to broadcast channel codes to leverage the performance of non-linear codes. A global and federated model are trained and results are shown for an RNN-based scheme and lighter-weight scheme." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024deep,\ntitle={Deep Learning Aided Broadcast Codes With Feedback},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yHVjncoGSp},\nnote={under review}\n}" }, "abstract": { "value": "Deep learning aided codes have been shown to improve code performance in feedback codes in high noise regimes due to the ability to leverage non-linearity in code design. In the additive white Gaussian broadcast channel (AWGN-BC), the addition of feedback may allow the capacity region to extend far beyond the capacity region of the channel without feedback, enabling higher data rates. On the other hand, there are limited deep-learning aided implementations of broadcast codes. In this work, we extend two classes of deep-learning assisted feedback codes to the AWGN-BC channel; the first being an RNN-based architecture and the second being a lightweight MLP-based architecture. Both codes are trained using a global model, and then they are trained using a more realistic vertical federated learning based framework. We first show that in most cases, using an AWGN-BC code outperforms a linear-based concatenated scheme. Second, we show in some regimes, the lightweight architecture far exceeds the RNN-based code, but in especially unreliable conditions, the RNN-based code dominates. The results show the promise of deep-learning aided broadcast codes in unreliable channels, and future research directions are discussed." }, "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 Learning", "Wireless Communication", "Feedback Coding", "Error Control Coding", "Federated 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/22dcf25f244cb5c843b93abd6f956fa180f17eaf.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/e25fcbd9076322f37c03aa910421b336c26c386d.zip" }, "title": { "value": "Deep Learning Aided Broadcast Codes With Feedback" }, "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 consider the weaknesses section. Essentially, my concerns are regarding the impact of the video encoder and the impact of the choice of tracker. I wish to understand how sensitive the method is to these two components, and whether the improvements are coming from the overall proposed design or certain specific components. I will further discuss this with fellow reviewers before making a final decision on the rating." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The idea is innovative and well motivated.\n2. The paper is well-written and easy to follow.\n3. The method achieves strong performance improvements with respect to the baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces CoSPaL (Contextual Self-Paced Learning), a novel approach to Weakly Supervised Spatio-Temporal Video Grounding (WSTVG), which aims to locate objects in videos based on text descriptions without requiring detailed bounding box annotations during training. CoSPaL proposes three main components to overcome limitations of existing methods: Tubelet Phrase Grounding (TPG) for improved object tracking, Contextual Referral Grounding (CRG) for better query comprehension, and Self-Paced Scene Understanding (SPS) for progressive learning of complex scenarios. Results are reported on common grounding benchmarks such as VidSTG and HC-STVG-v1/v2." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have a few major concerns regarding the proposed method:\n\n1. Regarding the Video Encoder: Existing works such as STCAT and TubeDETR utilize a Resnet-101 backbone to encode each frame. Even recent works such as VGDINO use the Grounding-DINO frozen image encoder. With this method using I3D as a separate video encoder, it could be an unfair comparison with existing works. How does this method perform if the authors utilize the same features as produced by the Grounding-DINO backbone as the video features? To make it consistent with previous works.\n\n2. The method uses a pretrained tracker in the pipeline. I am concerned regarding the impact of this tracker, and how the performance changes if a different tracker is used? There does not seem to be any ablations regarding this." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Please refer to the Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The problem addressed in this paper is of certain significance and is interesting.\n2. The introduction of self-paced learning is reasonably justified.\n3. Experimental results on different datasets indicate that the algorithm has achieved certain improvements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Targeting the issues of inconsistent predictions in time judgment, difficulty in comprehending complex queries, and the complexity of application scenarios faced by weakly supervised Spatio-Temporal Video Grounding (STVG), this paper introduces the self-paced learning method, which has achieved certain performance improvements on two conventional datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation of introducing self-paced learning should be illustrated more clearly. Are there other methods that can address the challenges presented in this paper? It is suggested to provide a more detailed explanation in the related work section.\n2. In Section 3.2.1, the tracking algorithm in the TPG module seems to play a vital role in learning the whole module. It makes a good start with GroundingDINO in the whole model, and it would be beneficial to add an analysis of the ablation of the tracking algorithm.\n3. Also, in section 3.2.1, the part of the Spatial Grounding Module, the formula description seems confused. It does not give key dimension information, such as in the similarity calculations: $\\text{SIM}(f_{w_m},f_{\\tilde{T_k}})=(\\mathrm{MLP_{q}}(f_{w_{m}})^{T}\\mathrm{MLP_{k}}(f_{\\tilde{T_{k}}}))/\\sqrt{d}$, where $f_{w_m}\\in \\mathbb{R}^{1\\times768}$ and $f_{\\tilde{T_{k}}}\\in\\mathbb{R}^{T\\times{256}}$ obtains from the above, so the $MLP_q(f_{w_m})\\in\\mathbb{R}^{1\\times{D}}, MLP_{k}(f_{\\tilde{T_{k}}})\\in\\mathbb{R}^{T\\times{D}}$, where $D$ is the MLP output dimension, but if it is like this, then matrix multiplication is failed. Besides, in $\\mathrm{A_T}(f_{T}, f_{w_{m}})=\\sum_{k=1}^{K}\\operatorname{softmax}\\left(f_{\\tilde{T}_{k}}, f_{w_{m}}\\right) \\mathrm{MLP}_{v}\\left(f_{\\tilde{T}_{k}}\\right)$, what is the mean about $f_T$? Is $\\sum_{k=1}^{K}$ meant to be a matrix addition of all tubelet features? Then it will give $ A_\\in\\mathbb {R}^{1\\times{D}}$; at this point, how to get the distribution of scores between different tubelets? This part of the equation is confusing.\n4. The comparison parameters of GPU memory use and training time in Figure 5 are ambiguous because the fully supervised models compared in the figure are experimented on different resolutions. However, the paper does not list the information about the resolution of the fully supervised models compared at the time of statistics and the resolution of their own models and whether the hardware parameters (e.g., information about CPU, GPU, memory) are unified across the different models, which I think are essential settings." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 above" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThe paper presents a method CoSPaL to address the key issues in WSTVG.\n2.\tThe paper includes comprehensive quantitative analysis and visualization, providing empirical evidence to support the proposed method.\n3.\tThe proposed CoSPaL model shows strong performance gains on multiple datasets, with notable improvements of 3.9% on VidSTG and 7.9% on HCSTVG-v1." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes CoSPaL (Contextual Self-Paced Learning) for Weakly Supervised Spatio-Temporal Video Grounding (WSTVG). The method introduces three key components: Tubelet Phrase Grounding (TPG), Contextual Referral Grounding (CRG), and Self-Paced Scene Understanding (SPS). CoSPaL aims to improve spatio-temporal video grounding by enhancing the model’s ability to understand complex queries and progressively adapt to more difficult scenarios. The effectiveness of CoSPaL is validated on three benchmark datasets, with significant performance improvements over baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe technical contribution and motivation is unclear. For example, TPG module uses cross-modal attention, contrastive learning, and feature reconstruction to facilitate interactions between words and tubelets, these techniques are common in the field. Why this particular implementation contributes to improved spatial and temporal grounding?\n2.\tThe SPS component is not clearly described. It is unclear how sample complexity is determined within the curriculum learning framework, is the object number? But intuitively, videos with more objects, faster changes, and more interactions are more complex. \n3.\tWhat does “Self-Paced” mean? This concept is not clearly explained in the paper.\n4.\tIn Table 5, the introduction of SPS in the TPG module leads to a performance drop in vIoU@0.3 (TPG+SPS vs TPG). This raises concerns about the efficacy of SPS in certain settings. The authors should provide a deeper analysis of why this occurs.\n5.\tIn page 10, the part of Impact on actor localization, the corresponding quantitative results for this claim 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": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please address the weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "Three WSTVG benchmarks are used for performance evaluation.\n\nThe compared methods are state-of-the-art. CG-STVG and VGDINO are published in 2024.\n\nThe figures and tables are clear.\n\nVarious figures are used to present the performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper targets the Weakly Supervised Spatio-Temporal Video Grounding (WSTVG) task, which tries to retrieve specific objects and segment by sentence queries without relying on labeled data. The spatial and temporal grounding modules are used to mine the spatio-temporal information. Also, a contextual referral grounding module is utilized to extracts contextual information from query. Finally, a self-paced curriculum learning strategy is adopted for optimization. Some experiments are conducted." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "In Abstract, authors state that “we first explore the potential of state-of-the-art object detection models for WSTVG”. In fact, it is wrong since many STVG methods use object detection models, e.g., [1].\n\nWhat is the difference between “self-paced scene understanding” and “self-paced curriculum learning”. In Section 3.2.3, authors use “self-paced scene understanding” as the section title, but all the contents describe the self-paced curriculum learning strategy. Besides, the strategy has been used in many works (Wang et al., 2022a; Soviany et al., 2022). Why you treat it as your third contribution and your title?\n\nThe technological novelty is not enough for ICLR. Most modules in the presented network in Figure 3 are popularly used. For example, cross attention in spatial grounding and temporal grounding.\n\nAuthors should conduct the main abalation study in the main paper.\n\nThe start and end timestamps need to be added in Fig 4.\n\nSome grammatical errors. For example, \"three benchmark WSTVG datasets\" in Abstract.\n\n[1] Weakly-Supervised Spatio-Temporal Video Grounding with Variational Cross-Modal Alignment" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024contextual,\ntitle={Contextual Self-paced Learning for Weakly Supervised Spatio-Temporal Video Grounding},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yHj6EunfVQ},\nnote={under review}\n}" }, "abstract": { "value": "In this work, we focus on Weakly Supervised Spatio-Temporal Video Grounding (WSTVG). It is a multimodal task aimed at localizing specific subjects spatio-temporally based on textual queries without bounding box supervision. Motivated by recent advancements in multi-modal foundation models for grounding tasks, we first explore the potential of state-of-the-art object detection models for WSTVG. Despite their robust zero-shot capabilities, our adaptation reveals significant limitations, including inconsistent temporal predictions, inadequate understanding of complex queries, and challenges in adapting to difficult scenarios.\nWe propose CoSPaL (Contextual Self-Paced Learning), a novel approach which is designed to overcome these limitations. CoSPaL integrates three core components: (1) Tubelet Phrase Grounding (TPG), which introduces spatio-temporal prediction by linking textual queries to tubelets; (2) Contextual Referral Grounding (CRG), which improves comprehension of complex queries by extracting contextual information to refine object identification over time; and (3) Self-Paced Scene Understanding (SPS), a training paradigm that progressively increases task difficulty, enabling the model to adapt to complex scenarios by transitioning from coarse to fine-grained understanding." }, "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": [ "spatio-temporal video grounding", "weakly 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/e246594b6969ef5fd36f0a1b3a3804e60d30c17e.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": "Contextual Self-paced Learning for Weakly Supervised Spatio-Temporal Video Grounding" }, "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": { "value": "We present a novel jailbreaking strategy that employs an attacker LLM to generate diversified and obfuscated adversarial prompts, demonstrating significant improvement over past approaches." }, "_bibtex": { "value": "@misc{\nzhao2024diversity,\ntitle={Diversity Helps Jailbreak Large Language Models},\nauthor={Weiliang Zhao and Daniel Ben-Levi and Junfeng Yang and Chengzhi Mao},\nyear={2024},\nurl={https://openreview.net/forum?id=yI60yhMQ7L}\n}" }, "abstract": { "value": "We have uncovered a powerful jailbreak technique that leverages large language models' ability to diverge from prior context, enabling them to bypass safety constraints and generate harmful outputs. By simply instructing the LLM to deviate and obfuscate previous attacks, our method dramatically outperforms existing approaches, achieving up to a 62\\% higher success rate in compromising nine leading chatbots, including GPT-4, Gemini, and Llama, while using only 12\\% of the queries. This revelation exposes a critical flaw in current LLM safety training, suggesting that existing methods may merely mask vulnerabilities rather than eliminate them. Our findings sound an urgent alarm for the need to revolutionize testing methodologies to ensure robust and reliable LLM security." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Weiliang_Zhao2", "~Daniel_Ben-Levi1", "~Junfeng_Yang1", "~Chengzhi_Mao2" ] }, "authors": { "value": [ "Weiliang Zhao", "Daniel Ben-Levi", "Junfeng Yang", "Chengzhi Mao" ] }, "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": [ "Attack", "Large Language Model", "Safety" ] }, "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": "zhao|diversity_helps_jailbreak_large_language_models" }, "pdf": null, "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": "Diversity Helps Jailbreak Large Language 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": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "The benchmark primarily focuses on insurance, specifically insurance laws from the United States. This focus may introduce bias into the evaluation process, posing a risk for models developed in different country contexts." }, "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": "Check Weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Comprehensive Benchmark: The paper presents INS-MMBench, which is the first comprehensive benchmark tailored for evaluating LVLMs in the insurance domain. This benchmark is extensive, covering 8,856 multiple-choice visual questions across 12 meta-tasks and 22 fundamental tasks, providing a robust framework for assessing LVLM capabilities in various insurance scenarios.\n\n2. Systematic Framework: The authors have developed a systematic and hierarchical task definition that ensures the tasks are closely aligned with real-world applications in the insurance industry. This bottom-up approach to task construction enhances the benchmark's relevance and practicality, making it a valuable tool for both research and practical applications.\n\n3. The paper also includes an extensive evaluation of multiple representative LVLMs, offering detailed performance analysis across different insurance types and meta-tasks. This analysis not only validates the effectiveness of the INS-MMBench benchmark but also provides actionable insights into the current capabilities and limitations of LVLMs in the insurance domain, guiding future research and development efforts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces INS-MMBench, a comprehensive benchmark designed to evaluate the performance of LVLMs in the insurance domain. It is the first initiative to systematically review multimodal tasks within the insurance sector and establish a specialized benchmark for it." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Multi-Choice Format Limitations: This benchmark follows a similar style to MMBench and MME in the general multimodal domain, all of which formulate their questions into multiple-choice formats. While this is an effective method for evaluating model performance, it has limitations that prevent generalization to open-ended question answering, which is more representative of real-world applications.\n\n2. Static Benchmark and Data Leakage: The benchmark is static, which does not mitigate the data leakage problem. This will likely render the benchmark less effective in future developments.\n\n3. Focus on US Insurance Law and Potential Bias: The benchmark primarily focuses on insurance, specifically insurance laws from the United States. This focus may introduce bias into the evaluation process, posing a risk for models developed in different country contexts." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Regarding the first limitation, could you share your perspective on how the current selective tasks directly align with the actual stages in the insurance process? For example, specific insurance stages like underwriting or claims processing?\n\nI would consider slightly increasing the score if convinced that the benchmark specifically addresses key insurance stages, rather than being a collection of VQA tasks merely related to the selected insurance categories(auto, property, health, and agriculture)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The motivation behind establishing an insurance benchmark is worthwhile. Evaluating LVLMs' capabilities on core insurance stages like underwriting and claims processing is practical and meaningful.\n2. The benchmark covers a reasonable range of core insurance types relevant to key areas in everyday insurance applications.\n3. The study provides an insightful error analysis, highlighting the current limitations of LVLMs in interpreting insurance-specific visual content." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The potential for Large Vision Language Models (LVLMs) to impact the insurance industry is substantial, yet largely unexplored. This study establishes a benchmark to evaluate LVLM capabilities within the domain, focusing on four main insurance types: auto, property, health, and agriculture. To create the benchmark, the authors gathered multimodal data for each insurance category from public sources and converted it into multiple-choice questions using GPT-4o. They then evaluated popular LVLMs on this benchmark to provide an initial assessment of LVLM performance and reveal current limitations in handling insurance-related content by an error analysis. Finally, the authors try to address gaps in insurance knowledge and reasoning skills by adding insurance-related information to the prompt." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Misalignment between Intent and Implementation**: While the authors claim the benchmark includes 12 meta-tasks and 22 fundamental tasks across stages like underwriting and claims processing in the Introduction section, the tasks illustrated in the paper are only loosely related to these stages. For example, meta-tasks in auto insurance such as “vehicle information extraction” and “vehicle damage detection” focus heavily on general computer vision tasks rather than directly addressing insurance-specific stages. This makes the benchmark feel more like a vision task set than an insurance task set. \n2. **Limited Accessibility for Reproducibility**: Although the authors promise to release the code and dataset, the GitHub repository has not been updated in four months, containing only a readme and a few diagrams. This lack of resources limits my ability to further assess the benchmark’s true rationality and effectiveness.\n3. **Limited Novelty**: Some conclusions, such as “performance of closed-source LVLMs varies by training data size and methods,” are too general and widely understood, offering little new insight. The paper would benefit from focusing on more specific findings directly related to the insurance domain." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Can you provide more details on how the benchmark tasks were selected?\n\nIt would be helpful to understand the criteria used to determine which tasks were included in INS-MMBench. Specifically, how did you ensure that the tasks accurately reflect real-world insurance challenges and not just general visual recognition problems?\n\n2. Have you considered including more complex, multi-step reasoning tasks?\n\nGiven the importance of decision-making in insurance, would it be possible to expand the benchmark to include tasks that require multi-modal integration and reasoning (e.g., verifying a claim using images, text descriptions, and numerical data)? This could better showcase the strengths and weaknesses of LVLMs in handling real-world scenarios.\n\n3. How do you envision improving the interpretability of model evaluations?\n\nSince explainability is critical in insurance, have you considered adding tasks that require LVLMs to provide justifications or rationales for their answers? This could allow for a deeper evaluation of how well models understand and explain their decisions, which is crucial for real-world applications.\n\n4. Do you have insights on the performance gap between open-source and closed-source models?\n\nThe results indicate a narrowing gap between open and closed-source models. Can you elaborate on specific factors contributing to this trend, and how future benchmarks might encourage more competitive open-source solutions?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "(1) Originality: INS-MMBench is the first benchmark tailored to evaluate LVLMs in the insurance domain. The authors' approach to defining tasks using a bottom-up hierarchical methodology is innovative and ensures that the benchmark aligns with real-world insurance scenarios, making it a pioneering effort in applying LVLMs to this new domain.\n\n(2) Quality: The authors systematically identify and organize multimodal tasks across four types of insurance, and their comprehensive evaluation of ten LVLMs provides some insights. The inclusion of detailed error analysis and the exploration of prompt engineering techniques to mitigate common issues further strengthen the paper, offering practical suggestions for improving model performance.\n\n(3) Clarity: The authors explain each step of their methodology in detail. \n\n(4) Significance: The introduction of INS-MMBench contributes to the field, as it enables a more nuanced evaluation of LVLMs in a domain with substantial practical applications. The benchmark could lead to improved automation in insurance-related tasks, such as claims processing and fraud detection, thus enhancing efficiency and accuracy in the industry. Moreover, by highlighting the narrowing performance gap between open-source and closed-source LVLMs, the paper encourages further research and development, potentially driving advancements in accessible and effective AI solutions for the insurance sector." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a benchmark in the systematic evaluation of LVLMs in this field by introducing INS-MMBench, a domain-specific benchmark designed to assess these models across various insurance-related tasks.\n\nKey contributions:\n\n(1) INS-MMBench is the first comprehensive benchmark tailored for the insurance domain. It covers four representative types of insurance: auto, property, health, and agricultural insurance, reflecting real-world insurance scenarios such as underwriting and claims processing.\n\n(2) The authors used a bottom-up hierarchical task definition approach to identify and refine relevant tasks for each insurance type. They collected and processed datasets to create visual question-answer pairs, ensuring that the benchmark aligns with practical applications in the insurance industry.\n\n(3) The paper evaluates different LVLMs using INS-MMBench. The results highlight the challenges these models face and give some insights." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Benchmark Definition Lacks Depth in Insurance Scenarios\n\nWhile INS-MMBench introduces tasks related to insurance, many are more aligned with general, common-sense VQA rather than specialized, nuanced scenarios seen in real-world insurance applications. To better reflect practical needs, the benchmark should include more complex tasks, such as multi-step reasoning or risk assessment based on a mix of visual and contextual data.\n\n2. Overemphasis on Basic Tasks\n\nSome tasks, like license plate recognition, are too basic and can be handled by smaller, specialized models. Evaluating LVLMs on these tasks does not showcase their strengths. Instead, the benchmark should focus on tasks requiring more advanced reasoning, such as verifying claims by cross-referencing multiple data points, to highlight the real capabilities of LVLMs.\n\n3. Limited Emphasis on Reasoning and Higher-Order Tasks \n\nThe benchmark lacks tasks that test higher-order reasoning, which is crucial for insurance scenarios. Tasks involving contextual understanding, complex decision-making, and multi-modal integration would better evaluate how well LVLMs can handle real insurance industry challenges.\n\n4. Lack of Focus on Interpretability \n\nInsurance applications require transparency, yet INS-MMBench primarily uses multiple-choice questions, limiting the ability to assess whether models can explain their decisions. Future benchmarks should include tasks that require LVLMs to provide rationale, enabling evaluation of their interpretability, which is critical for building trust in automated systems.\n\n5. Clarification Needed on Table 1\nIt appears there might be an error in the labeling of the last two rows in Table 1. Currently, \"OmniMedVQA\" is described as domain-specific for math, and \"Mathvista\" as domain-specific for medical. Given the names and typical use cases, it seems like these two may have been accidentally switched.\n\nRecommendations:\n1. Integrate more complex, domain-specific scenarios that mimic real-world tasks.\n2. Replace basic tasks with challenges requiring higher-order reasoning and contextual analysis.\n3. Add tasks that require models to explain their answers, enhancing interpretability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 LVLMs perform on insurance tasks that related to temporal reasoning for example analyzing claim patterns over time? The work evaluates static image understanding, but many insurance tasks require understanding temporal relationships and changes over time, and it seems that those samples are missing from the current benchmark." }, "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 well-written and well-organized. \n- This is the first systematic benchmark specifically designed for the LVLMs evaluation in the insurance domain and fills a gap in the current benchmark that often overlooks domain-specific applications.\n- The experiments are comprehensive, and thorough error analysis categorizing different types of model failures is provided in the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces INS-MMBench, the first comprehensive benchmark designed to evaluate LVLMs in the insurance domain. INS-MMBench includes four insurance types: auto, property, health, and agriculture, and includes 8856 multiple-choice questions across 12 meta-tasks and 22 fundamental tasks. It is designed to evaluate LVLMs in practical insurance tasks, such as vehicle damage detection and health risk monitoring, combining real-world visual information with insurance-specific questions. Through the experiments, the authors show the current limitations of LVLMs in insurance domain and suggests targeted data and domain knowledge for improving the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The human baseline experiments only involve 3 graduate students specialized in insurance, which is a small sample size. This might not accurately represent the range of expertise and variability in the real-world insurance evaluations. I would suggest bringing in more experts from the industry to help perform the human evaluation. \n- This work does not discuss potential biases in the data sources or methods for mitigating them, which means that there is a risk that the benchmark may favor certain model behaviors or fail to generalize to different insurance scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024insmmbench,\ntitle={{INS}-{MMB}ench: A Comprehensive Benchmark for Evaluating {LVLM}s' Performance in Insurance},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yIN4yDCcmo},\nnote={under review}\n}" }, "abstract": { "value": "Large Vision-Language Models (LVLMs) have demonstrated outstanding performance in various general multimodal applications such as image recognition and visual reasoning, and have also shown promising potential in specialized domains. However, the application potential of LVLMs in the insurance domain—characterized by rich application scenarios and abundant multimodal data—has not been effectively explored. There is no systematic review of multimodal tasks in the insurance domain, nor a benchmark specifically designed to evaluate the capabilities of LVLMs in insurance. This gap hinders the development of LVLMs within the insurance domain. In this paper, we systematically review and distill multimodal tasks for four representative types of insurance: auto insurance, property insurance, health insurance, and agricultural insurance. We propose INS-MMBench, the first comprehensive LVLMs benchmark tailored for the insurance domain. INS-MMBench comprises a total of 2.2K thoroughly designed multiple-choice questions, covering 12 meta-tasks and 22 fundamental tasks. Furthermore, we evaluate multiple representative LVLMs, including closed-source models such as GPT-4o and open-source models like BLIP-2. This evaluation not only validates the effectiveness of our benchmark but also provides an in-depth performance analysis of current LVLMs on various multimodal tasks in the insurance domain. We hope that INS-MMBench will facilitate the further application of LVLMs in the insurance domain and inspire interdisciplinary development." }, "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 vision-language model", "insurance", "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/a8564e8547b762c3d4ddd5f30ba9d184a07ff782.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/ce43e2c17282ed6b8f8737b19b9bc6be05415252.pdf" }, "title": { "value": "INS-MMBench: A Comprehensive Benchmark for Evaluating LVLMs' Performance in Insurance" }, "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": 4 }, "primary_area": null, "questions": { "value": "- How much does performance change if attention maps in self-attention layers are learned as opposed to derived from AF2 predicted contacts?\n\n- Based on my understanding of Fig. 1, the hyperparameter $d$ should be 20, corresponding to the number of amino acids. Why do you not embed amino acids in a higher dimensional space as you get deeper into the VAE?\n\n- Instead of using a dimension-wise FC layer, do you think it would be valuable to have an attention pooling layer to reduce the number of parameters (# of parameters in attention pooling does not scale with the length of the sequence)? It also automatically allows you to handle sequences of different lengths. \n\n- For matENC-DMS, are variants at the same position either all in the training set or the testing set? Training on some varaints and testing on other variants at the same position has been shown to be a form of data leakage that inflates performance.\n\n- Does difference in performance between matENC-DMS and matVAE-MSA depend on the number of variants assayed in the DMS? Could you include that In Fig. A8?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The authors explore novel architectural innovations to the DeepSequence/EVE family of models. In particular, they use self-attention layers where the attention map is determined from predicted contacts in AF2 structures. They also place more expressive priors on the latent space in matVAE-MSA. These are innovative ideas that have not yet been considered in the field. \n\n- The authors clearly benchmark their method to other state-of-the-art methods and clearly show the impact of their architectural modifications on model performance. This is one of the clearest papers I have read." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces two new methods for variant effect prediction: (1) matVAE-MSA which is a VAE trained on sequences from the protein family of interest and (2) matENC-DMS which is trained on DMS data from a specific protein. matVAE-MSA is similar to DeepSequence and EVE but the authors introduce architectural changes (i.e. self-attention layers) and experiment with more complex priors: mixture of diagonal Gaussians and VAMP. On pharmacogenes in ProteinGym, matVAE-ESM underperforms DeepSequence and ESM models. However, matENC-DMS, whose architecture is exactly the encoder portion of the DMS data and is trained on functional activity scores from a DMS assay, outperforms all other methods. Given that models trained on DMS data do significantly better than models solely trained on sequences from a MSA, the authors conclude that DMS data is valuable to train variant effect predictors on pharmacogenes." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The primary weakness of this paper is that their VAE model does not outperform existing unsupervised variant effect predictors (like ESM or DeepSequence). They find that using a more complicated prior does not improve performance and that self-attention layers with attention maps defined using AF2 contacts does not help. \n- The authors do go on to say that their encoder-only model trained on DMS data does outperform unsupervised variant effect predictors, but they do not compare to unsupervised variant effect predictors fine-tuned on DMS data. Their are many ways this fine-tuning has been proposed in the past and the authors should benchmark matENC-DMS to those methods: https://pubmed.ncbi.nlm.nih.gov/35039677/, https://www.nature.com/articles/s41467-024-51844-2, and https://arxiv.org/abs/2405.06729. \n- The authors should try pre-training a VAE on MSA data and then fine-tuning the encoder of the MSA on DMS data. \n- A central goal of the paper seems to be to identify the settings in which DMS data is useful for improving variant effect predictions. In Fig. A8, the authors are unable to find any correlations between metadata of the protein and performance difference between the DMS-trained and MSA-trained models. However, they don't consider structural features of the protein itself. Analysis along those lines would be interesting." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "See \"Weaknesses\"" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "* **Novelty**: To the best of my knowledge, the authors' proposed framework is novel.\n* **Impact**: The authors' motivation (i.e., assessing how well evolutionary pressure corresponds to fitness and the corresponding impact on variant effect prediction) is solid, and such studies would likely be of interest to the machine learning for proteins community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Here the authors propose a framework based on representation learning via VAEs for use on downstream deep mutational scanning tasks. The authors discuss their design choices and then proceed to benchmark their method against standard baseline methods for these tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Despite the potential impact of the authors' work, I believe that the authors' submission has significant issues that prevent me from recommending acceptance at this time. I provide details on the major issues below:\n\n* **Unclear motivation for model design choices**: The authors spend a significant amount of time experimenting certain modeling/architecture choices (e.g. using a mixture of gaussians prior rather than a unimodal prior), which in the end don't have an impact on model performance. Indeed, this is listed as one of the authors' main contributions in the introduction. Could the authors comment on their rationale for exploring these modeling choices? Did previous results for this task find that more expressive priors led to improved performance? Or was there a more principled reason to assume that these specific priors would lead to better performance? Without more context it's hard for the reader to understand why these results are being presented. On a related note, it would be great to see an ablation study assessing the impact of training a predictor on DMS datasets using representations from previous methods (e.g. DeepSequence) compared to the same task with the authors' proposed architecture. Without this information, it's difficult for the reader to understand whether any boosts in performance for the models trained on DMS data can be attributed to the authors' proposed encoder network or if the results are solely due to training on DMS data.\n* **Unclear significance of experimental results**: Perhaps most importantly, it's not clear to me that any meaningful conclusions can be drawn from the experimental results (e.g. those presented in Table 3). In particular, given the large error bars it's difficult for the reader to assess if the provided results are statistically significant. Could the authors provide results from e.g. a t test? Moreover, it's not clear to me how the authors selected their final model hyperparameters (e.g. learning rates). The authors mentioned that their choices \"preserv[ed] stability and convergence\", but without more details it's hard to tell if these values were cherry-picked. Were these parameters e.g. chosen via cross-validation/performance on a held-out validation set? Given these issues, it's thus unclear whether the authors' claims are supported by their experimental results.\n* **Not self-contained/writing issues**: Given that ICLR is a general machine learning conference (as opposed to a more biology-focused venue), it would greatly improve the manuscript for the authors to spend more time in the introduction describing the problem setup and significance. Indeed, the introduction section to the manuscript feels extremely rushed, with little time spent on introducing the problem setting tackled by the authors. For example, providing a gentler introduction to domain-specific terms like deep mutational scanning, On the other hand, a significant amount of space is spent describing hyperparameters (e.g. sections 3.1/3.2) or details of individual datasets (e.g. section 2.3 + Table 1), which could be relegated to the Appendix. I would thus recommend that the authors restructure the manuscript so that the main text is self-contained for a general ICLR reader, with ancillary experimental details moved to the appendix to make space as needed. On a related note that could save some space, it's unclear to me why the authors spend a significant amount of time introducing certain mathematical/machine learning concepts which are subsequently not used in the method (e.g. introducing matrix decomposition before stating that transformer layers are used)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. If pharmacogenes experience low evolutionary pressure, why is MSA included? (line 18)\n2. What is the significance of the statement \"sets a new benchmark for 2 out of 26 proteins\"? (lines 21-22)\n3. Regarding the \"5-fold cross-validation framework,\" what is the rationale for training a model to fit a DMS dataset rather than conducting zero-shot predictions? For a new protein, if a DMS dataset is already available, it is likely that users have already identified favorable mutants.\n4. Are there any differences between the MSA sequences used and those provided by the ProteinGym benchmarks? If so, why not utilize ProteinGym's versions?\n5. Why is there a manual selection of a threshold to binarize the DMS target scores? What are the specific criteria for this selection and what are the resulting values?\n6. Could you clarify the statement \"For instance ESM2 (150M), ESM2 (15B) and ESM-1v (ensemble) are all flavors of ESM\"?\n7. How should we interpret the statement \"We compared with those models in ProteinGym which perform the best on at least one pharmacogene-related protein DMS dataset according to SpearmanR\"? Does \"perform the best\" refer to zero-shot prediction or supervised learning tasks? What are the candidate models and their respective scores? Have all models on the ProteinGym leaderboard been considered in this comparison?\n8. While the algorithm does not apply any specific designs to pharmacogene, it seems like a general framework that can be applied to any mutation effect prediction tasks. In this case, it is suggested to test the model on the complete ProteinGym benchmarks of all 217 assays." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- A clear presentation on the new transformer-based module for the encoder and decoder.\n- A comprehensive investigation and analysis on the impact of different designed modules to the prediction task." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an encoder-decoder language model similar to DeepSequence for the task of mutation effect prediction and tests it on 33 drug-related DMS datasets from ProteinGym." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The presentation of the motivation is unclear (Q1, 3).\n- The justification for the experimental design and significance of the results is not clearly articulated (Q2, 3, 4, 6).\n- The design of the prediction tasks appears to be questionable (Q3, 5, 8).\n- The comparison with baseline methods is incomplete (Q7)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Can the authors more clearly explain what they feel are the contributions of the paper?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The experimental results are cleanly laid out and the authors do a nice job of not over-selling their results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper first proposes an unsupervised variant effect predictor (VEP), matVAE-MSA. The model is a VAE with transformer layers for the encoder and decoder with a sparse attention mask derived from the WT AlphaFold-predicted structure. Additionally the authors propose a supervised model, matENC-DMS which uses the same encoding architecture as matVAE-MSA, but maps the encoded vector to a scalar to predict the fitness label. matVAE-MSA does not consistently outperform existing unsupervised approaches, but does provide the best results for two proteins, MK01 and RAF1. matENC-DMS, which is trained on family-specific DMS data, consistently improves performance over unsupervised approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Unfortunately I see a number of weaknesses with the current paper. First, the use of a transformer as a VAE for modeling protein families is not new and exists in ProT-VAE. Nonetheless, given that this new architecture does not outperform existing simple VAE architectures it is not clear what contribution this paper is making with this? While the matENC-DMS results are interesting, there is no benchmarking with other supervised approaches. In particular, the paper \"Learning protein fitness models from evolutionary and assay-labeled data\" addresses combining evolutionary data with DMS data. At a minimum the authors should compare to this approach. Overall, I wasn't able to see what contributions the authors are making. The new VAE model doesn't seem to add any new insights into how to design better unsupervised generative models and it does not outperform existing approaches. The supervised model is not benchmarked against any similar approaches and does not provide any new insights." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Matrix Variational Auto-Encoder for Variant Effect Prediction in Pharmacogenes},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=yIRtu2FJvY},\nnote={under review}\n}" }, "abstract": { "value": "Variant effect predictors (VEPs) are designed to predict the impact of protein variants on cellular function, traditionally using data from multiple sequence alignments (MSAs). This assumes that natural variants are fit, a premise challenged by pharmacogenomics, where some pharmacogenes have low evolutionary pressure. In this context, deep mutational scanning (DMS) datasets are of particular interest since they provide quantitative fitness scores for variants. In this work, we propose a transformer-based matrix variational auto-encoder architecture and evaluate its performances on $33$ DMS datasets corresponding to $26$ drug target and absorption-distribution-metabolism-excretion (ADME) proteins available in the ProteinGym benchmark. Our model trained on MSAs (matVAE-MSA) outperforms a model similar to the widely used VEPs in pharmacogenomics, and sets a new benchmark for $2$ out of $26$ proteins. We compare matVAE-MSA with matENC-DMS, a model with similar capacity, but trained and evaluated on DMS data in a 5-fold cross-validation framework. matENC-DMS outperforms both the best available model for $15$ out of $33$ DMS datasets and matVAE-MSA for all ADME, and certain drug target proteins. Our results shed new light on the role of evolutionary pressure for the validity of the premise of VEP design. In turn motivating the development of DMS datasets to improve VEPs on pharmacogene-related proteins." }, "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": [ "variant effect prediction", "variational auto-encoder", "transformer", "deep 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/1f7b99f9edc86487ebb5ec005085bfc7fd2e9931.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": "A Matrix Variational Auto-Encoder for Variant Effect Prediction in Pharmacogenes" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]