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": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please see the weaknesses 2,3" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The research problem is important and timely, as large language models are rapidly advancing and achieving near-human capabilities, and allowing them to surpass human performance with potentially only weak human supervision is a critical issue.\n2. The paper is well-written and clearly structured, with a good introduction and related work section, clearly described methodology, and well-organized experiments and results.\n3. The experimental results are promising. The scale of the experiments, diversity of evaluation, and the rich comparison are appreciated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper explores the alignment problem when the LLM outperforms humans and human supervision is therefore weak. The authors propose a multiagent contrastive preference optimization (MACPO) framework to for weak-to-strong alignment by iteratively reinforcing unfamiliar positive behaviors while penalizing familiar negative ones. The authors evaluate their method on HH-RLHF and PKU-SafeRLHF datasets and show that MACPO improves alignment performance of strong students and weak teachers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the description of the MACPO framework in section 4.1, 4.2 is clear, the introduction of the framework in line 053 and line 196 is confusing. This is mainly due to the use of term 'behavior' and 'familiar' before they are defined. Moreover, the paragraph in line 196 is almost the same as the paragraph in line 061, which does not provide improved clarity.\n2. In Section 3.1, the formulation of the problem, the authors adopt an analogy setting where weak teachers are fine-tuned small models and strong students are big models 'initialized on weak labels generated by weak teachers for the held-out question set(line 261)'. In this case, it is not clear whether the strong student is truly strong, and it is not clear how this situation relates to the motivation problem in line 013. \n3. Computation efficiency discussion and comparison would be helpful as the proposed method requires multiple iterations of optimizing multiple models. Would the Strong-to-weak alignment and Self-alignment benefit from the longer training time (with the same effective computation as MACPO)?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "In Section 3.1, the paper assumes that using weak teachers can incrementally improve alignment by reinforcing positive behaviors. Could the authors clarify what criteria define a \"weak teacher\" and how its effectiveness in improving alignment is measured compared to using strong agents from the beginning?\n\nThe paper describes using multiple agents for reinforcement but lacks details on whether adjustments are made dynamically based on each agent's performance. Are there mechanisms in place to adjust reinforcement strategies depending on agent success rates, and if so, how are these adjustments implemented to ensure optimal alignment?" }, "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 using a multi-agent contrastive preference optimization approach to achieve weak-to-strong alignment is innovative. Most existing work focuses on direct fine-tuning or reinforcement learning methods to improve alignment, but these approaches cannot leverage the incremental learning power of weak teachers to iteratively refine alignment, limiting the model’s ability to generalize across various behaviors. This paper introduces a method that enhances alignment by reinforcing positive unfamiliar behaviors and penalizing negative familiar ones, as well as significantly improves alignment performance as more weak teachers contribute to training.\n\n2. The workflow is well-structured, as it combines contrastive preference optimization with a multi-agent setup to make the alignment process adaptable and iterative. This approach encourages behavior diversity among agents and further enhances the robustness and effectiveness of the alignment process.\n\n3. The experiments are extensive, with detailed analysis of the results. These experiments validate the effectiveness of the MACPO framework and demonstrate the scalability and adaptability of the method across models with different sizes and complexities​." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a multi-agent contrastive preference optimization (MACPO) approach for weak-to-strong alignment in LLMs. The authors utilize a multi-agent framework that iteratively improves both weak teachers and strong students by reinforcing unfamiliar positive behaviors and penalizing familiar negative ones. The experimental results show that MACPO achieves enhanced alignment performance over traditional methods, particularly as the number of weak teachers increases.\n\nThe main contributions of the paper include: 1) introducing the MACPO framework for weak-to-strong alignment; 2) incorporating mutual positive behavior augmentation and hard negative behavior construction to support iterative improvements; and 3) validating the proposed method's effectiveness through experiments on helpfulness and harmlessness alignment datasets​." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experiments use only two alignment datasets, expanding the evaluation to include additional alignment tasks like toxicity detection or complex reasoning would provide a more comprehensive assessment of MACPO's generalizability.\n\n2. In Section 4.2 HARD NEGATIVE BEHAVIOR CONSTRUCTION, the paper does not clearly explain how agent interactions are managed or how behaviors are tracked across weak and strong agents throughout the alignment 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": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the weakness, I'm happy to modify my rate based on the response of the authors and refer to other reviewers' comments." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors clearly define the problem, making it easy to understand the motivations behind their approach. This clarity in the problem statement helps to set a strong foundation for the rest of the work. And this paper opens up several avenues for future research, providing a solid foundation for follow-up studies. The authors discuss limitations and possible extensions, showing an awareness of the field's current needs and future directions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces MACPO, a framework for weak-to-strong alignment that enables strong language models to learn from weaker teachers. The authors claim that MACPO enhances positive behavior exchange and alignment refinement through iterative strategies, improving both teacher and student alignment, and results on benchmark datasets show MACPO’s effectiveness, with stronger alignment as the number of weak teachers grows." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The most concerning aspect is the statement in line 224: *“Since a high perplexity ppl of the positive strong student indicates weak labels may contain negative noises,”* which appears to form the foundational assumption of the entire framework. I have not encountered a statement like this before. In my opinion, alignment and preference learning should prioritize evaluation metrics as the basis for setting up metrics. Providing substantial evidence, such as a reference paper or experiments to demonstrate this metric’s validity and necessity, would be essential for this work. However, I could not find corresponding evidence, which may represent a fundamental weakness. At least the author needs to show a positive correlation between perplexity and answer quality (although this positive correlation is also very weak evidence to me, considering that correlation does not mean causation). However, the author did not even show the most basic evidence, which is very limited.\n\nI outline other concerns:\n\n1. I find the pipeline somewhat confusing. According to Algorithm 1, the proposed method first derives the “Strong Student” before proceeding with “Weak Teacher” training. Typically, in teacher-student learning paradigms, the teacher model is trained first and then provides supervision for the student model. If my understanding of the pipeline is correct, could the authors clarify the rationale behind this approach?\n\n2. It appears that the model sizes listed in Table 1 are inconsistent. For example, MACPO results are reported using both the 8B and 70B models, while other approaches only rely on the 8B model. This discrepancy may undermine the fairness of the comparisons.\n\n3. In Table 1, how are the final evaluation results derived based on a third-party reward model? With multiple gold models available in RewardBench, it would be helpful for the authors to explain their choice of this particular model over others.\n\n4. For HH-RLHF training, it’s unclear why the authors opted to use only 10K samples, which represent less than 10% of the original dataset. Do the final results depend heavily on this specific subset? This choice should be validated, as it raises concerns about the generalizability of the experimental results.\n\n5. Are the experiments restricted solely to alignment for helpfulness and harmlessness? It would be beneficial to extend these experiments to other tasks, such as Reddit TL; DR. Even if the approach does not perform well on such tasks, presenting these results could offer valuable insights." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "If I understand correctly, when constructing positive behaviors, one selection criteria is that the weak labels should have low perplexity when evaluated on the strong student. However, doesn’t this low perplexity mean that this weak label is familiar to the strong student? Because low perplexity indicates that this weak label is likely to be generated by the strong student as well. This seems to contradict the goal of generating unfamiliar positive behavior." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed approach is intuitive and partially solves the problem of collapsing by learning through self-generated data. \n\n2. The ablation study is comprehensive, validating the claims. The authors clearly illustrate the benefits brought up by unfamiliar positive behavior and familiar negative behavior." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes Multi-Agent Contrastive Preference Optimization (MACPO), which aims at letting strong students and weak teachers to learn from each other by encouraging unfamiliar positive behaviors and penalizing familiar negative behaviors. The proposed algorithm achieves better performance compared with other weak-to-strong alignment methods on helpfulness and harmlessness benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing can be significantly improved, especially for section 4. The notation often comes with 4 to 5 super and subscripts, which is very difficult to follow. I highly recommend the authors clean this part up.\n\n2. The proposal is intuitive, but the concept of “familiar” and “unfamiliar” is not well defined or discussed. Does familiar mean self-generated content and unfamiliar mean content generated by other models? What are the measures you can use for determining familiarity?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a multi-agent contrastive preference optimization (MACPO) framework to facilitate weak teachers and strong students learn from each other to improve weak-to-strong alignment performance." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024macpo,\ntitle={{MACPO}: Weak-to-Strong Alignment via Multi-Agent Contrastive Preference Optimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x1Okv4kbVR},\nnote={under review}\n}" }, "abstract": { "value": "As large language models (LLMs) are rapidly advancing and achieving near-human capabilities, aligning them with human values is becoming more urgent. In scenarios where LLMs outperform humans, we face a weak-to-strong alignment problem where we need to effectively align strong student LLMs through weak supervision generated by weak teachers. Existing alignment methods mainly focus on strong-to-weak alignment and self-alignment settings, and it is impractical to adapt them to the much harder weak-to-strong alignment setting. To fill this gap, we propose a multi-agent contrastive preference optimization (MACPO) framework. MACPO facilitates weak teachers and strong students to learn from each other by iteratively reinforcing unfamiliar positive behaviors while penalizing familiar negative ones. To get this, we devise a mutual positive behavior augmentation strategy to encourage weak teachers and strong students to learn from each other’s positive behavior and further provide higher quality positive behavior for the next iteration. Additionally, we propose a hard negative behavior construction strategy to induce weak teachers and strong students to generate familiar negative behavior by fine-tuning on negative behavioral data. Experimental results on the HH-RLHF and PKU-SafeRLHF datasets, evaluated using both automatic metrics and human judgments, demonstrate that MACPO simultaneously improves alignment performance of strong students and weak teachers. Moreover, as the number of weak teachers increases, MACPO achieves better weak-to-strong alignment performance through more iteration optimization rounds." }, "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": [ "weak-to-strong alignment", "preference 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/68485b2c007c09bae088cc313edc0be469f1b8e7.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": "MACPO: Weak-to-Strong Alignment via Multi-Agent Contrastive Preference Optimization" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. The authors need to include Decision Diffuser as the baseline.\n\n2. Why there are only part of the environments in the ablation study in Table 3." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The presentation of this paper is great and easy to follow.\n2. The idea of this paper is simple and clear. \n3. The experimental result looks good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new diffusion-based RL planning method named DMEMM. Unlike previous methods that use fixed isotropic variance and disregard the rewards which may lead to a mismatch between generated trajectories and those desirable for RLs, DMEMM explicitly incorporates transition-based, reward-based, and reward-aware diffusion modulation loss. By doing so, DMEMM enhances both the coherence and quality of generated trajectories, achieving state-of-the-art performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The tested diffusion-based offline RL planning algorithm missed the Decision Diffuser[1], which performance is much better than the original Diffuser. Also, it's important that the Decision Diffuser doesn't have the issue of inaccurate transition as it only predicts the state sequence and calculates the actions by inverse dynamics.\n\n2. The motivation of this paper is ambiguous. I assume that the initial motivation is to improve the consistency of the dynamic by explicitly incorporating the loss of the difference between the real states after transition and the generated states. However, the initial idea of diffusion models is to measure the distribution of the trajectories, which has already included the dynamics. I'm wondering if adding such a loss can really improve the consistency of dynamics. The authors need to seriously discuss this problem.\n\n3. Same as previous papers, this paper also uses the paradigm of autoregressive generation, where each time only the first action will be executed. In this case, generating accurate transitions seems to be less important. \n\n4. The paper is lack of novelty, the idea of adding a loss on the transitions is an incremental work.\n\n\n[1] Is Conditional Generative Modelling All You Need for Decision-Making? Anurag A., et al. ICLR 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": 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": "- Will the authors publish the source code?\n\nI do not have further questions since the paper was written in a very straightforward way and easy to understand. To improve the work, please see the comments above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper was written in a very straightforward way, making it easy to follow and understand\n2. The being discussed problem is important and the paper proposed a practical method to solve it." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focused on offline reinforcement learning (offline RL) with diffusion models as planner. Following Diffuser (Janner et al. 2022), a series of studies have been proposed the improved the performance. This paper discussed the problem of generating trajectories with consistency between transitions to ensure coherence. The being proposed methods includes a transition-based diffusion modulation loss, and dual reward guidance. Using the mixed loss function, the being proposed method was tested on D4RL locomotion, showing SOTA performance compared to previous diffusion planning algorithms; and on Maze2D, showing similar performance with previous best." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The being tested benchmarks are too narrow, recent diffusion planning studies usually included other D4RL environments AntMaze, Franka Kitchen, etc. But this paper only examined D4RL locomotion, thus it remains unclear whether the being proposed method work well in general tasks.\n\n2. While the being proposed method performs well in D4RL locomotion compared with previous diffusion planning methods (avg. normed. return = 87.9, Table 1), it should be discussed that the previous best diffusion model-based method on this benchmark was from diffusion policy [1] (avg. normed. return = 88.0 [1], and 89.0 in [2]).\n\n3. When compared with HD-DA, the performance on the Maze2D (Table 2) is worse when the size becomes large. Since I do not see other experimental results on planning, I am skeptical about the scalability of the being proposed method when task becomes more challenging.\n\nOverall, the paper provides limited insights. There lack both theoretical and empirical insights about the problems being touched, due to limited variety of benchmarks and limited analysis (with only a very formulaic ablation study), making this paper not qualified for the ICLR bar. \n\n## Minor problems\n\n- In Tables 1 and 2, error bar should be explained whether it is STD or SEM or confidence interval.\n- Many related works on diffusion planning on offline RL are not discussed, see [3] for reference\n- No reproducibility statement\n\n### Ref\n\n[1] Wang, Zhendong, Jonathan J. Hunt, and Mingyuan Zhou. \"Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learning.\", ICLR 2023.\n\n[2] Dong, Zibin, et al. \"CleanDiffuser: An Easy-to-use Modularized Library for Diffusion Models in Decision Making.\", NeurIPS 2024 Track on Datasets and Benchmarks.\n\n[3] Zhu, Zhengbang, et al. \"Diffusion models for reinforcement learning: A survey.\" arXiv preprint arXiv:2311.01223 (2023)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. The paper needs to explain how diffusion models can be applied within imitation learning [1, 2], a subset of reinforcement learning.\n\n2. The experiments, limited to locomotion and Maze2D, are insufficient. Adding manipulation and image-based tasks would strengthen credibility, especially given diffusion models’ origins in image domains.\n\n3. How does DMEMM handle environments with highly stochastic transitions or non-stationary reward functions, and what are the main challenges it faces in such contexts?\n\n4. What are the primary computational trade-offs introduced by incorporating dual-guided sampling in DMEMM, and how does it perform compared to standard model-based and model-free RL approaches in terms of efficiency?\n\n5. To what extent is DMEMM sensitive to the weighting of transition- and reward-based modulation losses, and what practical guidelines can be provided for setting these parameters across various RL environments?\n\n6. Can the proposed transition-guided sampling in DMEMM be further adapted to handle hierarchical or multi-step decision-making tasks, and if so, what modifications might enhance its applicability?\n\nI would be inclined to raise the score if you clear up my cnofusion and discuss diffusion model's applicability to imitation learning.\n\n[1] Tim Pearce, Tabish Rashid, Anssi Kanervisto, David Bignell, Mingfei Sun, Raluca Georgescu, Sergio Valcarcel Macua, Shan Zheng Tan, Ida Momennejad, Katja Hofmann, and Sam Devlin. Imitating human behaviour with diffusion models. In International Conference on Learning Representations, 2023.\n\n[2] Shang-Fu Chen, Hsiang-Chun Wang, Ming-Hao Hsu, Chun-Mao Lai, and Shao-Hua Sun. Diffusion model-augmented behavioral cloning. In International Conference on Machine Learning, 2024." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. **Motivation and intuition**: The motivation for improving trajectory generation in RL through transition-consistent diffusion models is convincing.\n\n2. **Novelty**: The way of how to utilize transition dynamics and cumulative reward modulation in the diffusion model is intuitive and novel. This paper presents an effective way to implement this idea.\n\n3. **Technical contribution**: The transition-based modulation loss for enforcing transition consistency seems particularly effective for producing coherent trajectories in complex environments.\n\n4. **Ablation study**: The ablation studies are comprehensive, breaking down each framework component and showing its role in improving trajectory quality." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the challenge of improving diffusion model-based trajectory generation for offline reinforcement learning (RL) by incorporating transition dynamics and reward functions through a new framework, Diffusion Modulation via Environment Mechanism Modeling (DMEMM). Experiments show that DMEMM achieves state-of-the-art results on D4RL and Maze2D tasks, with ablation studies highlighting the effectiveness of each framework component. This work addresses a key gap by enhancing trajectory coherence with real-world constraints in diffusion-based RL planning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Clarity**: The meaning of some notations, such as the subscripts and superscripts on τ, is not clearly explained. While the intended meaning can be inferred after careful review, a clearer definition of each notation would improve the paper's readability.\n\n2. **Completeness**: The paper lacks a discussion on the application of diffusion models within imitation learning [1, 2], a subset of reinforcement learning, which is necessary to clarify the advantages of using diffusion models in planning.\n\n3. **Experiment details**: The experiments are not sufficient, as they focus only on locomotion and Maze2D environments. Adding other environments, such as manipulation and image-based tasks, would enhance the credibility of the results. Given that diffusion models originated in image domains, testing on image-based environments would be valuable.\n\n4. **Reproducibility**: The authors did not provide code nor commit to releasing it upon acceptance, which limits the reproducibility of the results." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weaknesses" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well-written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a model-based offline diffusion RL method. By adding auxiliary loss about state transitions and cumulative rewards to the conventional diffusion loss, the authors claim to solve the problem of generating trajectory discontinuity when diffusion is applied to reinforcement learning. The authors validate the proposed method on D4RL and claim that their method achieves SOTA." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The PRELIMINARIES are not convincing enough. The author claims in the PRELIMINARIES that reinforcement learning is modeled as a quadruple $(S,A,T,R)$ in (Sutton & Barto, 2018), whereas in general, we model reinforcement learning as a quintuple $(S, A,T,R,\\gamma)$. The objective of reinforcement learning policy optimization is generally to maximize the cumulative \\textbf{discount} reward, and this paper does not mention anything about reward discount $\\gamma$. Given that this is significantly different from the normal research paradigm of reinforcement learning, if this is by design, the author should explain the purpose of doing so. If this is an oversight, I hope the author will correct it in future research.\n\n2. The motive is unclear. the motivation presented by the authors is that \"the use of fixed isotropic variance and the disregard for rewards may lead to a mismatch between generated trajectories and those desirable for RL \". I don't understand how this problem relates to the method proposed by the author. The authors claim that this problem comes from (Wu et al., 2019). However, after browsing the paper, I did not find that the author of the original article claimed this problem. If this is proposed for the first time, please provide more detailed analysis and examples to prove that it does exist and will have an impact on offline reinforcement learning, and it is necessary to prove in the experiment that the method proposed in this paper will indeed solve or link this problem.\n\n3. The novelty of the method is limited. Since the reverse process of the diffusion model is to predict the denoising error, I do not quite understand how the formula (9) in the paper is attached to the loss of the diffusion model. From the formula (10) in the paper, I believe that the reward model fits the Q function in reinforcement learning (of course, there is no discount term $\\gamma$). Formula (11) is the Q-weighted fitting error. Overall, whether it is learning state transition or using the Q function, it can not impress me.\n\n4. The results are not convincing. The authors' experimental results on D4RL did not live up to their claims of SOTA. Many studies using diffusion models for offline reinforcement learning have provided more competitive experimental results, such as [1, 2]. In addition, \"Notably, both DMEMM-w/o-λtr and DMEMM-w/o-tr-guide exhibit significant performance drops,\" the authors said in the ablation experiment. emphasizing the crucial role of incorporating transition dynamics in our method.\" However, from the experimental results in Table 3, The difference between dMEMm-w /o-λtr and dMEMm-w /o-tr-guide and DMEMM is insignificant. If the author sticks to his point, I hope the author can come up with more convincing evidence, such as significance testing.\n\nReference:\n\n[1] Chen H, Lu C, Ying C, et al. Offline Reinforcement Learning via High-Fidelity Generative Behavior Modeling[C]//The Eleventh International Conference on Learning Representations.\n[2] Wang Z, Hunt J J, Zhou M. Diffusion Policies as an Expressive Policy Class for Offline Reinforcement Learning[C]//The Eleventh International Conference on Learning" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024diffusion,\ntitle={Diffusion Modulation via Environment Mechanism Modeling for Planning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x1SfON9HvT},\nnote={under review}\n}" }, "abstract": { "value": "Diffusion models have shown promising capabilities in trajectory generation for planning in offline reinforcement learning (RL). However, conventional diffusion-based planning methods often fail to account for the fact that generating trajectories in RL requires unique consistency between transitions to ensure coherence in real environments. This oversight can result in considerable discrepancies between the generated trajectories and the underlying mechanisms of a real environment. To address this problem, we propose a novel diffusion-based planning method, termed as Diffusion Modulation via Environment Mechanism Modeling (DMEMM). DMEMM modulates diffusion model training by incorporating key RL environment mechanisms, particularly transition dynamics and reward functions. Experimental results demonstrate that DMEMM achieves state-of-the-art performance for planning with offline reinforcement learning." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Reinforcement Learning", "Offline Reinforcement Learning", "Planning", "Diffusion 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/ebf4a2e27b8e3a83646101628c76704bf83ff4fe.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": "Diffusion Modulation via Environment Mechanism Modeling for Planning" }, "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": 2 }, "primary_area": null, "questions": { "value": "See above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This work includes a CogMath framework that could be helpful in evaluating the math reasoning abilities of LLMs more robustly, and covers several dimensions that may introduce perturbation to the stability of reasoning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces a CogMath framework that consists of nine agents to evaluate the mathematical reasoning ability of large language models from the perspective of comprehension, problem solving and solution summarization. \n\nSpecifically, in the comprehension stage, the agents attempt to rephrase, disrupt (permute word ordering), remove condition and add condition of the original question. In the problem solving stage, the agents attempt to conduct analogical reasoning, numerical transformation and knowledge refinement (reshape the semantics of 'half') of the original question. In the solution summarization stage, agents attempt to question the information in the intermediate steps of the solution, and conduct backward reasoning of the question. The experimental results demonstrate that the abilities of current strong LLMs on GSM8K and MATH are overestimated by 30-40% by the calibration of those agents. Besides, CogMath may not serve as an effective prompt-based reasoning enhancement and the problem difficulty and lengths in MATH are negatively correlated with the pass rates in CogMath." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It is not easy to imagine how a handful of agents included in CogMath can be generalized to more challenging questions. For example, how does the backward reasoning being feasible in mathematical proofs. The knowledge redefinition only limits its scope to 'half' and the questions contain the word, where works like FRoG (Li et al. 2024) includes richer quantifier-based variants of GSM8K. It is not surprised to see that the current figures of LLMs in reasoning is not a stable display, but attempts like one-time numerical transformation might make it more robust, but only marginally. Besides, I didn't find enough evidence regarding efforts to make sure the agents faithfully finish their jobs.\n\nThis work also collects MExam. However, I know nothing about it from the contents.\n\n\nReference\n\n[1] FRoG: Evaluating Fuzzy Reasoning of Generalized Quantifiers in Large Language 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": 3 }, "primary_area": null, "questions": { "value": "How can the inquiry agents ensure that they generate good questions that meet the dimension requirements? Is there any filtering process involved?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Comprehensive evaluation across nine dimensions can enhance the current math benchmarks.\n- Extensive experiments with multiple representative LLMs demonstrate the limitations of current LLMs on math reasoning capabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a multi-agent framework, CogMath, to evaluate the mathematical abilities of LLMs from a cognitive perspective. CogMath breaks down mathematical problem-solving into 3 stages: problem comprehension, problem solving, and solution summarization with nine evaluation dimensions. CogMath generates test samples cross multiple cognitive perspectives using a multi-agent system and reveals that current LLMs' math capabilities are overestimated, which demonstrates the strengths and weaknesses of different models at different evaluation dimensions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- CogMath uses LLMs to construct test samples and evaluate the model-generated answers across multiple dimensions. However, the correctness of generated test cases and the evaluation quality can be a major concern. It would be helpful to add human evaluation on the generated test samples and the judging process.\n- The performance degradation can be a regular case instead of an interpretation of overestimation, as the CogMath test questions can be harder than the original questions after processing across multiple dimensions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses section" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tA comprehensive and scientific benchmark that deeply investigates the flexible reasoning of LLMs is essential for the community.\n2.\tThe authors consider nine dimensions across problem comprehension, solving, and solution summarization, which aids in identifying the main challenges faced by current models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a benchmark for comprehensively evaluating the mathematical abilities of LLMs by examining three cognitive reasoning stages: problem comprehension, problem solving, and solution summarization. Experiments indicate that we may overestimate the capabilities of current LLMs, primarily due to their excessive imitation of superficial reasoning patterns." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tWhile this work addresses cognitive mathematical dimensions comprehensively, I have a question regarding the motivation. Why do the authors believe that previous works introducing perturbations into existing benchmarks are task-specific?\n2.\tMore details are needed about the dataset construction procedure, including how the judge agent is used to ensure the quality of $q_i$, how multiple reference agents are negotiated to finalize the answer, and which foundation models are utilized behind these agents.\n3.\tFor Figure 2, it would be clearer to replace the dimension index with the dimension name. Additionally, for Figure 3, it would be more straightforward if each group of bars represents the same model.\n4.\tIn Section 4.6, the current experiment uses a one-shot setting. Have the authors considered a nine-shot setting, where each demonstration represents one dimension?\n5.\tIn Section 4.7, how are the five tiers of difficulty defined?\n6. In Table 1: Should $21 be 21?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024cogmath,\ntitle={CogMath: Evaluating {LLM}s' Authentic Mathematical Ability from a Cognitive Perspective},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x1nlO1d1iG},\nnote={under review}\n}" }, "abstract": { "value": "As large language models (LLMs) exhibit potential in solving complex mathematical tasks, increasing attention has been directed toward constructing benchmarks to evaluate their mathematical capabilities. However, existing benchmarks are either limited to specific task types (e.g., long-text problem understanding) or rely solely on a coarse measure of answer accuracy, making them insufficient for assessing a model's authentic mathematical proficiency. In this paper, we propose CogMath, which provides a comprehensive assessment of LLMs' mathematical abilities based on human cognitive processes. Specifically, inspired by cognitive theories, CogMath formalizes the reasoning process into 3 stages that align with human cognition: problem comprehension, problem solving, and solution summarization, and encompasses 9 fine-grained evaluation dimensions from perspectives such as numerical calculation, knowledge, and counterfactuals. In each dimension, to carry out a scientific evaluation, we develop an ``Inquiry-Judge-Reference'' multi-agent system, where the Inquiry agent generates inquiries that assess LLMs' mastery from this dimension, the Judge agent ensures the inquiry quality, and the Reference agent provides correct responses for comparison with the LLMs' actual performances. A LLM is considered to truly master a problem only when excelling in all inquiries from the 9 dimensions. In experiments, we evaluate 7 mainstream LLMs by applying CogMath to three benchmarks, which cover the full K-12 mathematical curriculum. The results reveal that the authentic mathematical capabilities of current LLMs are overestimated by 30-40%. Moreover, we locate their strengths and weaknesses across different stages/dimensions, offering constructive insights to further enhance their reasoning abilities." }, "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": [ "Mathematical Reasoning", "Large Language Models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/00361965dee521cf43a8528825eac1c5beed233c.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/ef81be3b6c1780541e7d8b562f78f344829894f8.zip" }, "title": { "value": "CogMath: Evaluating LLMs' Authentic Mathematical Ability from a Cognitive Perspective" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "See 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": "- The idea is simple and neat. Directly optimizing the noise towards samples of maximum reward value is technically sound. The paper presents the idea in a clear way.\n\n- The paper considers both differentiable and non-differentiable reward settings, making direct noise optimization a quite practical method to use.\n\n- The paper provides some interesting theoretical insights to justify the drawback of direct noise optimization. Moreover, it provides a solution to avoid this drawback with a probability regularization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a method called direct noise optimization (DNO) for maximizing some reward function for the generated samples. Different from finetuning and RL methods, DNO operates as a form of testing-time optimization. DNO is also extended to the non-differentiable reward setting by leveraging zero-order optimization techniques. Some theoretical analyses are provided to account for its empirical effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- As the paper mentioned, directly optimizing the input noise can result in degenerated solutions, which the authors describe as OOD reward hacking. I think the problem, despite being alleviated by some additional regularization, will fundamentally limit the generation quality. This is conceptually similar to the problem of adversarial examples. Moreover, regularizing the noise to lie in a high-probability region may alleviate the reward hacking problem, but it also limits the ability to maximize the reward function. How can the effectiveness of this probility regularization be measured in practice? How to set the hyperparameter $\\gamma$ in practice?\n\n- Zero-order optimization is applied to the non-differentiable reward setting, but I am highly skeptical about its empirical effectiveness, as the zero-order optimization is known to be difficult to converge, not to mention that the noise optimization could be very challenging and non-convex/non-smooth.\n\n- The performance gain of DNO over the other baselines also seems be quite marginal from Table 1. For the other baseline methods, there are many hyperparameters that can be tuned. I think to better illustrate the performance gain, it will be good to compare the performance with different finetuning steps (or different inference-optimization steps)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to Weakness part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ This study considers aligning diffusion models in the inference time, alleviating the computational cost of aligning the model.\n+ The authors identify and address the problem with reward hacking in optimization.\n+ The proposed method is applicable for both differentiable and non-differentiable reward functions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the alignment of diffusion models by optimizing noises in the sampling process. The authors propose to optimize noises at each sampling step to improve the reward value, rather than optimize the model parameters. To this end, they discuss the effectiveness of the proposed DNO method, and address the reward hacking problem in optimization via a regularization term. Experiments show that DNO improves the reward value of generations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It seems that the idea of optimizing noises has been proposed by (Wallace et al., 2023b; Ben-Hamu et al., 2024; Novack et al., 2024; Karunratanakul et al., 2023). What is the essential difference between DNO and these prior works, except that DNO optimizes noises at each step?\n- The assumption of Theorem 1 is not convincing. I am not sure how to understand the smoothness from Figure 4 (Tang et al., 2024a). Although the noise-to-sample mapping is smooth, the reward function is usually complicated, especially for reward functions related to human preference. Is there any direct justification for the smoothness of $r\\odot M$?\n- The faster convergence speed of DNO is not well supported by their theoretical analysis. Equation (5) only demonstrates the final performance of DNO with SDE may be better than DNO with ODE, but does not justify the optimization speed.\n- In equation (10), there should be a $1/\\mu$ term.\n- There are several concerns about the experiment in Table 1. First, it is unclear what prompts and how many generations are used for evaluation. Second, for DNO, the annotated time is confusing. Does ``1 min’’ mean that generating one image using DNO costs 1 minute? Third, how about the performance of DNO with ZO-SGD, hybrid-1, and hybrid-2?\n- The evaluation for reward hacking based on $P(z)$ is limited. $P(z)$ only indicates the distribution of initial noise $z$ but cannot describe the distribution of the final generated image. I’d like to see more comparisons between images generated by different methods, either quantitatively or qualitatively.\n- Which experiment supports the claim in Line 452 `` while it prevents the test metrics, i.e., the HPS score and Pick Score, from decreasing throughout the process'’?\n- Is the proposed DNO also effective for more complex prompts? Have the authors compared images generated by different methods on both simple and complex text prompts?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 fact, the model is repeatedly applied to intermediate noises too. Why there is no regularization of out-of-distribution intermediate noise inputs to the model?\n2. Why there is no comparison to LGD and other baseline in Section 5.1? The performance of LGD is surprisingly bad in Section 5.2, even worse than SD 1.5. The authors argue that the reason is the complex reward used in 5.2, and I wonder whether LGD has a better performance in similar reward functions like 5.1." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper provides a thorough treatment of DNO, including theoretical analysis, practical implementation details, and solutions to key challenges (OOD reward hacking, non-differentiable rewards). The probability regularization technique using concentration inequalities is particularly novel and well-motivated since it addresses important practical concerns without no fine-tuning, works with arbitrary reward functions, can handle non-differentiable rewards, etc.\n2. The experimental evaluation is extensive and convincing, showing that DNO can match or exceed the performance of fine-tuning methods while requiring only inference-time optimization. The authors test on multiple reward functions and provide detailed ablation studies." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Direct Noise Optimization (DNO), a novel approach for aligning diffusion models with continuous reward functions at inference time. The key idea is to optimize the injected noise during the sampling process to maximize desired reward functions, without needing to fine-tune the model parameters. The authors develop theoretical foundations for DNO, introduce probability regularization to prevent out-of-distribution samples, and extend the method to handle non-differentiable rewards. Extensive experiments demonstrate that DNO can achieve state-of-the-art reward scores within reasonable inference time budgets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the paper proposes a regularization technique to provide overfit, the resulting images still seem to overfit. Besides, the regularization only focuses on constraining the initial noise distribution to the high probability area. Also see question 1\n2. Lack of baseline results: There is no inference time cost comparison with LGD since it's an important inference-time method as the baseline. Besides, a more relevant training method is DRAFT [1], which also uses direct backpropagation to optimize the reward function. It also faces a similar issue of reward hacking as the proposed method.\n\n\n[1] Directly Fine-Tuning Diffusion Models on Differentiable Rewards\nKevin Clark, Paul Vicol, Kevin Swersky, David J Fleet" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. How does DNO perform when used with other diffusion sampling methods besides DDIM? From my perspective of view, I think DDIM is the easiest and most direct way to model the mapping from noise to sample, i.e. $M_\\theta(z)$. If using other schedulers other than DDIM, I wonder how DNP will work to simulate $M_\\theta$. Moreover, if the algorithm leverages a differentiable reward function for optimization, the gradients are passed backward through all diffusion steps. I assume that would bring significant computational costs comparable to direct reward propagation. Can the authors comment on this?\n\n2. I don't see how eq (2) brings a \"prompt-agnostic\" method. For example, when aligning stable diffusion, is the optimized initial noise a \"universal\" optimal value for all prompts?\n\n3. Currently, there are already many works aimed at aligning diffusion models. Roughly, they can be classified into two categories,\n (1) directly fine-tune the diffusion models. See: https://arxiv.org/abs/2407.13734, https://arxiv.org/abs/2410.08315. While DNO includes comparisons with alignprop and ddpo in Table 1, comparisons with many later RL-based finetuning methods that alleviate reward hacking issues are missing. I wonder how DNO would work if compared to them. \n (2) purely inference time techniques. While DNO does not fine-tune diffusion models, DNO still needs to optimize through the diffusion pipeline to get optimized noise. The computational costs are still heavy. Here are some works that are purely inference-time and prompt-agnostic: https://arxiv.org/abs/2410.08134, https://arxiv.org/abs/2408.08252.\n\nDiscussing and comparing these works quantitatively (or qualitatively) is important for positioning DNO. However, it might also be true that these works are concurrent with this submission. Finally, at least qualitative discussions are needed.\n\n4. I assume the numbers in Table 1 do not consider reward hacking which is problematic for a method that claims to alleviate reward hacking. This is because alignprop would significantly suffer from mode collapse when aesthetic score > 8.5 from my experience. But Table 1 reports 8.9 for alignprop. Can the authors comment on this?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. DNO's main feature is training-free. Current researchers have been more interested in inference-time techniques that do not require costly fine-tuning. Therefore, this work stands for an important and attentive direction of aligning generative models.\n2. The integration of probability regularization to mitigate out-of-distribution reward hacking is interesting.\n3. The authors applied DNO across various reward functions, highlighting its adaptability.\n4. DNO is compared rather comprehensively with some baselines on a variety of reward functions, which is good." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the alignment of diffusion models to downstream tasks with continuous reward functions. It proposes Direct Noise Optimization (DNO) as a tuning-free, inference-time method for adjusting generated samples to maximize target rewards. Moreover, in the face of reward hacking (aka over-optimization), the paper introduces a probability regularization for alleviation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While theoretical parts of the probability regularization are comprehensive, I find those parts hard to parse. Could the authors highlight the key message of such regularization, and why would it work from a high-level perspective?\n2. Reward hacking is known to be a key concern in aligning LLM/diffusion models. The authors attempt to alleviate hacking through a regularize, which is good. However, the evaluation way remains problematic. In Fig 2, the authors presented that adding regularize sacrifices little performance for retraining fidelity. However the authors did not compare regularized DPO with other methods. In the end, the observation of Fig 2 is not a surprise to me because adding regularization almost surely leads to a tradeoff of performance and fidelity. There exist some works showing that simply adding small regularization to align-prop would greatly alleviate reward hacking. Therefore, Fig. 2 does not provide any further impressive information.\n3. One primary concern is that this work lacks many comparisons with existing fine-tuning-based and inference-based techniques that appear earlier or concurrent to this work. See more details below." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This work studies the problem of inference-time alignment of diffusion generative models with downstream objectives" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024inferencetime,\ntitle={Inference-Time Alignment of Diffusion Models with Direct Noise Optimization},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x1uv2gdjKV},\nnote={under review}\n}" }, "abstract": { "value": "In this work, we focus on the alignment problem of diffusion models with a continuous reward function, which represents specific objectives for downstream tasks, such as increasing darkness or improving the aesthetics of images. The central goal of the alignment problem is to adjust the distribution learned by diffusion models such that the generated samples maximize the target reward function. We propose a novel alignment approach, named Direct Noise Optimization (DNO), that optimizes the injected noise during the sampling process of diffusion models. By design, DNO operates at inference-time, and thus is tuning-free and prompt-agnostic, with the alignment occurring in an online fashion during generation. We rigorously study the theoretical properties of DNO and also propose variants to deal with non-differentiable reward functions. Furthermore, we identify that naive implementation of DNO occasionally suffers from the out-of-distribution reward hacking problem, where optimized samples have high rewards but are no longer in the support of the pretrained distribution. To remedy this issue, we leverage classical high-dimensional statistics theory to an effective probability regularization technique. We conduct extensive experiments on several important reward functions and demonstrate that the proposed DNO approach can achieve state-of-the-art reward scores within a reasonable time budget for generation." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Diffusion Models", "Inference-Time Alignment", "Optimization", "RLHF" ] }, "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/403ffe02031b1810b750bb8bf3267a63f7c512da.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/862145dca24fcbc59a503e3b4521fe03c3274513.zip" }, "title": { "value": "Inference-Time Alignment of Diffusion Models with Direct Noise Optimization" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 check weakness for details." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The proposed dataset could be useful for the community.\n2. Both close and open-sourced models are evaluated.\n3. Metrics are studied with human verification." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces SPORTU, a new benchmark designed to evaluate the capabilities of Multimodal Large Language Models (MLLMs) in sports understanding and reasoning. SPORTU consists of two components: SPORTU-text, focusing on text-based reasoning, and SPORTU-video, focusing on video-based reasoning. The authors evaluate various LLMs on both components, revealing limitations in complex reasoning tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The reviewer is concerned about the lack of diversity and coverage of the dataset because of the limited prompt templates and number of samples. \n\n2. Implementation could be possibly flawed. \n- The error in Figure 6 looks suspicious and makes the reviewer wonder whether the model is called correctly or not. \n- The reasoning prompt asks the model to first generate answer and then reasoning, which is not optimal since the model's final answer cannot benefit from the reasoning process.\n- It is known that LLM usually prefers its own answer so it is important to understand G-eval' quality with different LLMs as the rater.\n\nMinor:\nL821 typos of \"Section ??\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "n/a" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The video quality for the datasets\n- For each sports type, the video are biased to certain views or events?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- A multimodal new dataset for sports domain (with multiple sports) and well annotated by experts; the dataset should be helpful for the research communities\n\n- A well prompting capabilities to show the limitation of current LLM capabilities on the dataset. \n\n- Evaluating several reasonable public or private LLM models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides a multimodal dataset (text and slow-motion video) for evaluating (multimodal) LLM capabilities in the sports domain." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Yet another vertical dataset for LLM\n- It's helpful but marginal to expand the technical depth for the community\n- not clearly identified what current models failed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. How do you split the SPORTU-text questions into rules-related, strategy-related, and scenario-related? What is your basis?\n\n2. What are the results on rule/strategy/scenario, respectively, on sportu-text?\n\n3. How is the error analysis in 5.1 conducted? Is there a definition for each error type?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. It proposes SPORTU-text and SPORTU-video to boost understanding more sports with rules understanding in text and video domains.\n2. It analyzes the views, reasoning prompts, sport types, the error types, which are comprehensive.\n3. The writing is clear." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In the AI+Sports area, existing works are limited to restricted kinds of sports, absence of explanations, or lack of reasoning on rules, and it proposes SPORTU consisting of SPORTU-text and SPORTU-video to boost understanding more sports with rules understanding. SPORTU-text evaluates models on rule comprehension and strategy understanding in the pure text domain and SPORTU-video evaluates models on understanding both video details and rules in the video domain.\nIt evaluates LLMs and MLLMs on SPORTU-text and SPORTU-video, revealing their limitations in complicated sports questions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. A benchmark aims to evaluate certain abilities and give some insights. The paper does not deeply discuss why the models have different performances and does not give advice on how to resolve the problem of understanding videos and reasoning on rules.\n\n2. Prompt strategy in LLM can also be tested on MLLM when evaluating on SPORTU video benchmark to see how the reasoning process influences MLLM.\n\n3. It's not very clear if the questions in this dataset can comprehensively detect the models' abilities to understand sports.\n\n4. The Pearson correlation between humans and the other metrics is low. Many are near 0." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 refer to the weakness part." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. As a sport domain understanding benchmark, the proposed SPORTU combines text-based and video-based tasks to assess models' sports reasoning and knowledge application abilities.\n\n2. The evaluation setting is comprehensive including the direct prompting, chain-of-thought (CoT) prompting. In addition, few-shot promoting is also applied in SPORT-text evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper prensents SPORTU, a comprehensive Sports Understanding Benchmark that integrates both text-based and video-based tasks to evaluate models’ sports reasoning and knowledge application capabilities. Based on this benchmark, this paper tests the capability of existing open-source or close-source models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Unclear motivation. The authors should clarify the differences between the proposed SPORTU and existing sport domain understanding benchmarks. Although discussions have been made in introduction and related work section together with Table 1, it is still unclear why the introduced features , for example, slow motion, multi-camera angles are important. More discussions and visualizations are needed.\n\n2. Missing details in dataset construction. There exist some unclear details in the dataset construction. For example, how to guarantee the multi-camera setting? Is it achieved simply by human annotator check? In addiation, the proposed SPORTU contains both the multi-choice and open-ended question, how are these two categories divided?\n\n3. More advanced evaluation methods should be applied. For example, ST-LLM [1], qwen-vl [2]\n\n4. The paper writing should be polished. Some references are missing, for example \"Section ??\" in Line 821. The quotation mark error in '”Why is it a foul in the video?”' in Linee 482.\n\n[1] ST-LLM: Large Language Models Are Effective Temporal Learners\n[2] Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024sportu,\ntitle={SportU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x1yOHtFfDh},\nnote={under review}\n}" }, "abstract": { "value": "Multimodal Large Language Models (MLLMs) are advancing the ability to reason about complex sports scenarios by integrating textual and visual information. To comprehensively evaluate their capabilities, we introduce SPORTU, a benchmark designed to assess MLLMs across multi-level sports reasoning tasks. SPORTU comprises two key components: SPORTU-text, featuring 900 multiple-choice questions with human-annotated explanations for rule comprehension and strategy understanding. This component focuses on testing models' ability to reason about sports solely through question-answering (QA), without requiring visual inputs; SPORTU-video, consisting of 1,701 slow-motion video clips across 7 different sports and 12,048 QA pairs, designed to assess multi-level reasoning, from simple sports recognition to complex tasks like foul detection and rule application. We evaluate four prevalent LLMs mainly utilizing few-shot learning paradigms supplemented by chain-of-thought (CoT) prompting on the SPORTU-text part. We evaluate four LLMs using few-shot learning and chain-of-thought (CoT) prompting on SPORTU-text. GPT-4o achieves the highest accuracy of 71\\%, but still falls short of human-level performance, highlighting room for improvement in rule comprehension and reasoning. The evaluation for the SPORTU-video part includes 7 proprietary and 6 open-source MLLMs. Experiments show that models fall short on hard tasks that require deep reasoning and rule-based understanding. Claude-3.5-Sonnet performs the best with only 52.6\\% accuracy on the hard task, showing large room for improvement. We hope that SPORTU will serve as a critical step toward evaluating models' capabilities in sports understanding and reasoning. The dataset is available at \\url{https://anonymous.4open.science/r/ICLR_01-42D5/}" }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multimodal Large Language Models", "Sports Understanding", "Benchmark" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/a59f8040eefb9787243f17a62b36fb2f0b8313e1.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": "SportU: A Comprehensive Sports Understanding Benchmark for Multimodal Large Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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. Would it be possible to incorporate more details of the graph learning into the main paper?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. The problem has multiple applications and is very sensible. Image compression has direct implications, and it is sensible to think image compression in terms of its final use (object detection, classification, etc) as opposed to only pixel reconstruction.\n\n2. The proposed approach has a strong mathematical foundation and makes intuitive sense (conditional entropy should help resolve redundancy between certain tasks). \n\n3. Quantitative results are thorough and show promise." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors focus on learning methods for image compression. To this end, they propose a method that optimizes image compression for a number of different downstream visual tasks. They consider the potential redundancy of encodings for similar groups of visual tasks, and they utilize directed acyclic graphics to learn causal relationships between tasks. This approach leads to a better multi-task representation. They evaluate their approach on a number of different visual tasks and demonstrate convincing quantitative results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I believe the description of the DAG learning should have more detail. Specifically, some of the details of the DAG based algorithm (described in the appendix as Algorithms 1 and 2) should be incorporated into the main paper. \n\n2. There are some typos and grammatical errors." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. Could you state the number of parameter sets used and explain how they are allocated across groups and individual tasks? The paper seems to suggest it's per group, but it also needs a set of task parameters per task for computing the pairwise gradient coherence.\n2. How is the forecast loss in Eq.3 used? \n3. Line 287, could you clarify if task order impacts the cost calculation and, if so, how do you address this potential variability in their method?\n4. Is there any constraint on the subsets used for set cover?\n4. Is it better or worse when the same task is not allowed to appear in different clusters? \n5. How is bitrate controlled in the proposed framework? How do you determine $\\lambda_i$ for each task?\n6. Is the causal discovery step required? Could the minimum description length principle be used for causal discovery and, thus, unifying step 2 and 3 as finding a DAG structure that minimizes the bit-rate?\n7. Could you provide a complexity analysis showing how computational requirements change as the number of tasks increases?\n8. A simple baseline would be multi-task learning naively combined with end-to-end compression, i.e. without the grouping, just use them auxiliary tasks with well-tuned weights. Why would you not include such a baseline? How would this baseline compare theoretically to the proposed method?\n9. Figure 3, could you add a legend for what the color shade represent? Could you define the \"Anchor\" baseline in the figure caption or main text?\n10. How does the method perform compare to more familiar baselines such as VQGAN?\n11. Github link is not provided?\n12. L137 typo \"Hu\"?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper discusses an important but less studied aspect of representation learning - how to leverage different supervised signals to learn a better representation, where \"better\" is defined as lower bitrate with higher downstream performance. This principle is sound. The experimental results demonstrate that the proposed framework achieves better performance comparing to end-to-end compression methods using the same bitrate." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework, Taskonomy-Aware Multiple Task Compression (TAMC), for lossy compression where the distortion is defined by multiple downstream tasks. First, TAMC groups tasks into clusters where tasks within a cluster are mutually supportive and a share representation is learned for each cluster. Then, it leverages causal discovery to identify dependencies between groups. This results in a directed acyclic graph (DAG), which can be used for further compression of the representation. Experiments on the Taskonomy dataset demonstrate that TAMC achieves superior bitrate reduction and task performance compared to baseline compression methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Clarity**: the description of the framework lacks clarity. Given that the paper proposes a fairly complex system, clarity is even more important for the readers to understand. For example, although there is a space constraint, the description of step 3 (taskonomy-based compression) is too brief. Other more specific questions are raised below.\n\n**Lack of ablations**: there are at least 2 ablations the authors could provide. 1. an end-to-end compression with multiple supervised tasks as auxiliary tasks; 2. single task groups, where instead of grouping tasks together in phase 1, simply treat each task as a group and carry out the rest of the 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": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- Although the paper states that tasks were grouped based on gradient coherence, the justification for task grouping is not clearly stated in LN 470. Could the task grouping process be further explained?\n- In Appendix A.5.2, the authors state that they select the parent node for a chosen child based on the node with the highest mutual information and claim that this is equivalent to the optimal choice for conditional entropy. While this claim seems correct, in the partial ordering of Algorithm 5, the child node is not explicitly defined, so the claim does not appear to be directly related to the implementation of the proposed method. It seems there might be an implicit assumption that the task with the larger entropy in its representation distribution is assigned the parent node. Would this be an accurate interpretation?\n- It is mentioned that the total bit-rate of task-wise representations was used as a regularization for model training, but there is not enough explanation regarding the bit-rate for each task in the trained models. It would be helpful to include information on the observed bit-rates for each task and to explain how these bit-rates may be related to the tasks.\n- Although theoretical approaches are mentioned in Appendix A.5.4 and A.5.5, they do not seem to be fully utilized or proven within the logic of the paper. Is there a plan to elaborate on these? In addition, could further explanations be provided in relation to the above Weaknesses and Questions?\n- (minor) Uniform upper and lower case conventions in subsection titles (especially in Section 5)\n- (minor) Typo in Equation 3 (\\theta_s -> \\theta_s^t)\n- (minor) Typos in Algorithm 1, 2 in Appendix (end for\"\\n\"return)" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Hierarchical structure of multiple tasks: The authors propose a methodology that learns a universal representation for computer vision tasks by utilizing techniques from multi-task learning and uses these representations to form a hierarchical structure of the tasks. I believe this methodology could aid research on general representations in areas beyond data compression.\n- Robustness of the trained representation: At lower bit-rates, the performance on downstream tasks consistently outperforms trained compression methods (such as MLIC++, ELIC).\n- Ablation with random graphs: The authors compared the performance of the proposed causal DAG structure, which exploits causal relationships, with that of a random graph structure to verify its effectiveness. This demonstrates the existence of inter-task relationships and shows that efficient learning and inference, which take these relationships into account, have a considerable impact on performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a data compression technique using representations that can be applied to multi-tasking in computer vision. They propose a methodology for task-aware representation learning by using multi-task learning approaches and grouping similar tasks based on the alignment of gradient vectors during the training phase. To efficiently utilize the learned representations, they calculate the causal relationships between task groups using conditional entropy and construct a directed acyclic graph (DAG). The proposed method was validated using the Taskonomy dataset, and the baselines included traditional data compression methods (e.g., JPEG, WEBP) as well as some recent methods (e.g., ELIC, MLIC++, etc.). The authors confirmed the robustness of the proposed method at relatively lower bit-rates compared to the baselines across multiple tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of clarity, not well-organized components: The figures in the paper are generally not well-organized, and the font size within the figures is excessively small, causing visibility issues. In particular, in Figure 2, the font size within the figure is less than half the line height. Even if it results in a reduction of the amount of information in each figure, it seems necessary to increase visibility by focusing on the most important information.\n- Relatively low data decoding accuracy: As mentioned by the authors, the results in subsection 4.3.2 show that the proposed method performs similarly or relatively worse than existing methods in terms of image compression. In order to show that this difference is not a significant problem for the application of the method, it seems necessary to qualitatively compare the decoded images, at least with some examples.\n- No results with time and space complexity: One of the critical aspects of a data compression algorithm is minimizing the time and space cost of its execution. Since this work focuses on data compression techniques using multi-task learning methods rather than representation learning methods, I believe it is necessary to analyze such costs to justify the applicability of the proposed method.\n- Lack of baselines for comparison: While I think it seems meaningful that the paper uses traditional data compression techniques (e.g., JPEG, WEBP) as baselines, there is still a lack of comparison with more recent trainable methods. Even if the compared baselines are considered state-of-the-art models, I believe that additional comparisons are necessary to clarify the consistent robustness of the proposed method compared to other approaches.\n- To summarize, the paper lacks completeness in explaining the methodology and convincing the readers through the results, and the experiments are not sufficiently thorough. Regarding the experiments, it may be helpful for the authors to refer to the analytical techniques used in the MLIC++ paper [1], which they have cited as a key reference. Through a comprehensive revision of the content and additional analyses, the paper’s clarity can be improved, and its novelty can be further emphasized.\n\n[1] Wei Jiang et al. MLIC++: Linear Complexity Multi-Reference Entropy Modeling for Learned Image Compression. https://arxiv.org/abs/2307.15421" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024which,\ntitle={Which Tasks Should Be Compressed Together? A Causal Discovery Approach for Efficient Multi-Task Representation Compression},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x33vSZUg0A},\nnote={under review}\n}" }, "abstract": { "value": "Traditional image compression methods often overlook the intricate interdependencies in multi-task learning, resulting in inefficiency and redundancy. In this paper, we propose a novel compression framework that leverages causal graph models to uncover conditional relationships between mutually beneficial task clusters. By constructing directed acyclic graphs (DAGs) based on conditional entropy, we capture the causal links among tasks, enabling progressive, context-aware compression. Parent representations act as hyperpriors for their dependents, reducing redundancy, enhancing scalability, and boosting compression efficiency. Extensive experiments across key computer vision tasks, including segmentation, depth zbuffer, and autoencoding demonstrate superior bitrate reduction and task performance. Our findings underscore the importance of disentangling task representations and modelling causal relationships for efficient multi-task compression, offering a new perspective on compact representation learning for advanced intelligent systems. Code will be available at: https://github.com." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Video Coding for Machine", "Image Compression", "Multi-task Learning", "Causal Discovery" ] }, "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/23222ae8a10530aef73cfb4e6a91d4f9681f741a.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": "Which Tasks Should Be Compressed Together? A Causal Discovery Approach for Efficient Multi-Task Representation Compression" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "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": "- Could the authors explain why Cond-FiP performs similar to some baselines in noise prediction and sample generation, especially when the node scale is the same or smaller than used in training? How is the FiP model implemented? In the original paper, it seems that the task is the recovery of the topological ordering. Is the FiP baseline here aware of the causal structure of datasets?\n- How does the alternating application of transformer blocks E work? Is this just an alternating optimization method where you optimize for samples when nodes are fixed and optimize for nodes when samples are fixed?\n- For zero-shot inference of the SCM distribution, what is the level of distribution shift that a new dataset can have for this method to be able to extrapolate well?\n- The main capabilities of the proposed framework are noise prediction and observational/interventional sample generation. However, individual counterfactual sample generation is also important in many applications. Can this framework enable counterfactual inference?\n- In the Adaptive Transformer Block, which iteratively updates the noise conditioned on the dataset embedding $z_{emb}$, can we interpret this as sort of a noise abduction typically performed in counterfactual inference?\n- How exactly does one perform interventions in the Cond-FiP? It would be beneficial if the authors elaborate on this mechanism in the paper. From my understanding, we just feed in an intervened SCM causal graph with the mutilations and use the corresponding dataset embedding for conditional generation. However, this is not made very clear in the paper." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper is one of the first to consider generalizing the learning of functional mechanisms of structural causal models from arbitrary datasets and causal graphs and is a significant step toward building causally-aware foundation models.\n- The paper is written well with clear intuitions and explanations as to how it relates to similar work (e.g., FiP).\n- Although the assumptions are a bit strong (additive noise model, causal graphs known, noise variables known), the general idea of using an amortized procedure to approximate SCM distributions in a zero-shot manner is quite interesting.\n- The empirical results are convincing and show interesting observations, especially the performance of sample generation under distribution shifts and as the causal graphs scale up. It is certainly impressive that Cond-FiP can approximate the SCM distribution of 50 and 100 graph nodes quite well given that it was trained on only datasets with 20-node graph size." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Learning the causal generative process from observational data is a challenging problem bottlenecked by the necessity of learning a separate causal model for each dataset. This paper studies a unifying framework to enable zero-shot inference of causal generative processes of arbitrary datasets by training a single model. The authors adapt a recent advancement in causal generative modeling (FiP) to infer generative SCMs conditional on empirical dataset representations in a supervised setup, where the SCM is reformulated as a fixed-point problem. They propose an amortized procedure that takes in a dataset and its causal graph and learns a dataset representation. Then, the authors train a model conditioned on dataset embeddings to learn the functional mechanisms of the generative SCM. This framework enables both observational and interventional sample generation in a zero-shot manner. Empirical results show that the method performs competitively with baseline models for in-distribution and out-of-distribution settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- In a synthetic data scenario, assuming access to the noise samples is a feasible assumption, but for real-world datasets, this will not typically hold. Using the noise samples as the supervision for the dataset embedding model may easily become unrealistic. The authors have an evaluation on a real-world benchmark (Sachs) in the appendix where they fit a Gaussian model. However, interventional sample results are not provided.\n- I believe the idea to directly infer the SCM distribution under the additive noise model assumption is interesting. However, again, the feasibility of this assumption may not always hold. It is true that we often parameterize causal models as linear/nonlinear additive noise models, but this can be violated in practice. It seems that this approach would only hold under the strict assumption of additive noise models.\n- Knowledge of the causal graph for several datasets can potentially be a strong assumption. In real-world datasets, the causal graph may be unknown and must be discovered. However, for the sake of this work, the synthetic scenarios are good for proof of concept." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. I have a hard time understanding the Generation with Cond-FiP paragraph. The \\mathcal{T} symbol seems to have a different meaning in the Projection paragraph than in the Generation with Cond-FiP. Is it a notational error? Please clarify. Also, what is the relation between eq. 5 and the equation in the Adaptive Transformer Block paragraph? They seem oddly similar yet, if I understand correctly, the \\mathcal{T} denotes the whole generation process described in the Adaptive Transformer Block paragraph.\n2. The Interventional Predictions explanation is unclear to me. Could you elaborate more on how you “modify in place the SCM obtained by Cond-FiP”?\n3. Could you provide more details on the generation of the datasets? What is the density of the graphs used in training and evaluation? What are the differences in the parametrization of functional relationships between P_in and P_out?\n4. What kind of regressor was used as a causal mechanism in DoWhy? What architecture was used in DECI?\n5. What is the role of DAG attention in dataset embedding? Can you elaborate on what would happen if you were to use standard dot product attention in the dataset encoding phase?\n6. It seems that the performance of baselines on RFF OUT dataset degrades significantly compared to RFF IN. Could you explain why?" }, "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 is well motivated by recent advances in the area, novel, and of interest to the community.\n2. The method’s description is detailed and well-structured." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a novel method and task of inferring functional relationships of SCMs from observations and their associated causal structures in a zero-shot manner. The problem is well motivated by the literature. The main technical contributions of the work are the definition of Cond-FIP (modification to the FIP architecture), and the construction of the dataset embeddings (used to condition FIP). The latter is done by training the transformer-based encoder to predict the evaluations of the functional relationships over the distribution of SCMs. The former is done using the FIP architecture that takes as input sample embeddings conditioned on the SCM and is trained with the MSE objective on samples from multiple SCMs.\n\nThe method is evaluated in the synthetic setup on varying graph structures, functional relationships, and noises. It performs comparably to the selected set of baselines on noise prediction, sample generation, and interventional sample generation tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Experimental evaluation seems limited. The authors claim that their approach matches the performance of SoTA, but they only compare against a small set of methods. It seems natural that authors would compare against cited work (Javaloy et al., 2023; Khemakhem et al., 2021) and architectures used in causal discovery literature [1, 2],\n2. In various fragments of the text the authors claim that “this is the first time that SCMs are inferred in a zero-shot manner from observations“. While it is clear to me that this work combined with previous literature technically allows SCM inference (graph and functional relationships), this ability is not demonstrated nor evaluated in this work. \n3. The method does not generalize well to larger datasets (as stated in Appendix C). This is an important limitation and should be stated explicitly in the main text.\n4. The evaluation would be hard to reproduce as the details of train and test dataset generation are not provided.\n5. Some parts of the method’s explanation are unclear to me. (see Questions)\n\n[1] Annadani at al., BayesDAG: Gradient-Based Posterior Inference for Causal Discovery, NeurIPS 2023\n[2] Brouillard et al., Differentiable Causal Discovery from Interventional Data, NeurIPS 2020" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1.The theoretical foundation of \"ZERO-SHOT LEARNING OF CAUSAL MODELS\" indeed needs to be improved, given its complexity and the challenges it presents in the field of machine learning. Because the cross domain SCM difference seems to be significant. I still feel puzzled why zero shot learning theoretically works. Is there problems such as incapability of the observation-based learning with conditional FiP?\n\n2.In the experimental section, I think more comparisons are needed. For example in DoWHY dataset, you may vary the node number and more generative mechanisms. Also, some ablations study of the important modules of FiP (e.g., encoder) should be presented since this seems to be an important factor for the final results.\n\n3.I still feel that the interventional and sample generation can be more comprehensive. Is there any demo example to show that how your method differs from others? I guess that which node you manipulate still has a huge impact on the result. Am I right or not?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.The idea of zero-shot learning with conditional FiP is interesting. \n\n2.The problem studied is important in the sense that zero-shot learners are commonly needed to finish the tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes to amortize the learning of a conditional version of FiP to infer directly the generative SCMs from observations and causal structures in a zero-shot manner." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The theoretical justifications need to be improved.\n2. Some experiments 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": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No ethics concerns." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see weakness above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Zero-shot learning of a causal model is an important task, potentially due to the limitation of the training dataset when training the model. So I believe the task proposed by the paper is highly motivated. Additionally, the paper is written in a structured way." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method to learn a single model capable of inferring the causal generative processes of datasets in a zero-shot manner, rather than learning a specific model for each dataset. The general idea is to amortize the learning of a conditional version of the FiP architecture, which can infer the generative SCMs directly from observations and causal structures on synthetically generated datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I would suggest author to either use a figure or introduce a bit more background of the idea of amortized causal learning in related works.\n\n2. In background on causal learning, I still don't know how to compute H. and what $Jac_x$ is. Also, $d_n$ and M are not defined.\n\n3. In training setting, I don't understand how to use noise samples. In equation (1), they seem to be residual terms but in Section 3.1, they are said to play the rule of the target variable.\n\n4. Do we need specific requirements on training data? Are they supposed to cover as many as domains?\n\n5. Why modeling noise will help zero-shot ability of SCM? I wish the paper gives more explicit explanations.\n\n6. In Table 2 and Table 3, it seems that sometimes Cond-FiP has the worse performance than FiP. This needs to be explained." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose an approach to amortize the learning of causal models, thus enabling zero shot inference." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024zeroshot,\ntitle={Zero-Shot Learning of Causal Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x3F8oPxKV2},\nnote={under review}\n}" }, "abstract": { "value": "With the increasing acquisition of datasets over time, we now have access to precise and varied descriptions of the world, capturing all sorts of phenomena.\nThese datasets can be seen as empirical observations of unknown causal generative processes, or Structural Causal Models (SCMs).\nRecovering these causal generative processes from observations poses formidable challenges, and often require to learn a specific generative model for each dataset.\nIn this work, we propose to learn a \\emph{single} model capable of inferring in a zero-shot manner the causal generative processes of datasets. \nRather than learning a specific SCM for each dataset, we enable FiP, the architecture proposed in~\\cite{scetbon2024fip}, to infer the generative SCMs conditionally on their empirical representations.\nMore specifically, we propose to amortize the learning of a conditional version of FiP to infer directly the generative SCMs from observations and causal structures on synthetically generated datasets.\nWe show that our model is capable of predicting in zero-shot the true generative SCMs, and as a by-product, of (i) generating new dataset samples, and (ii) inferring intervened ones.\nOur experiments demonstrate that our amortized procedure achieves performances on par with SoTA methods trained specifically for each dataset on both in and out-of-distribution problems. \nTo the best of our knowledge, this is the first time that SCMs are inferred in a zero-shot manner from observations, paving the way for a paradigmatic shift towards the assimilation of causal knowledge across 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": [ "Causality", "Transformers", "Generative Models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/52bd709d10a2fe6cbcb0f53b1d4609f5e2fcd544.pdf" }, "presentation": null, "primary_area": { "value": "causal reasoning" }, "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": "Zero-Shot Learning of Causal Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- Please summarily explain why concrete distributions help compared to $\\epsilon$-greedy?\n- How does your new trick compare to [1]?\n\n[1] Simple and Effective Transfer Learning for Neuro-Symbolic Integration. Daniele et al. 2024" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The general idea of having membership functions for symbols is interesting, and is potentially an elegant abstraction of ideas from DSL. \n- The idea of learning rules from perception in general is very interesting, and is of significant importance in NeSy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Summary:\nThe authors propose a Neuro-Symbolic(NeSy) method that can learn rules from perception in an end-to-end fashion. The paper aims to improve upon recent lines of work in this direction, and proposes new designs and symbol selection strategies, with different levels of noise and distributions. The paper also proposes a gradient shortcut strategy to improve the training and convergence. The paper also show experiments on synthetic MNIST based NeSy benchmarks --- widely used to test NeSy frameworks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Clarity: I think the paper is very unclear. This can be significantly improved by adding a running example, and potentially moving the comparison (3.1) and parts of section 4 to the Appendix. However, the general writing of related works and introduction is not quite clear. I would suggest the authors to arrive at the main message of the paper at earlier stage in the introduction.\n\n- Section 2.1 is well written, and well-motivated. However, 2.3 and 2.3 (in my understanding the main contributions of the paper) are not clearly presented. It is not clear why Concrete distributions help?\n\n- Figure 3's explanation is quite unclear and imprecise.\n\n- I am not sure, how novel the idea of gradient shortcuts is --- 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": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How complex can this approach effectively scale to? \n2. In the experimental section, the random seed is fixed at 42, which is unexplained." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Originality: 2.5/5\n\nCompared to the existing work DSL, this approach offers the ability to explicitly learn the logic equation. It introduces a Concrete distribution alongside the Godel t-(co)norm proposed in DSL, which adds noise to the soft symbolic values.\n\nQuality: 2/5\n\nPros: The context and mathematical formulations are well presented.\n\nCons: Scalability is a significant concern, as the number of neurons correlates with the number of internal states. While DSL has demonstrated scalability up to a 1000-digit sum, this work does not address its own scalability. How complex can this approach effectively scale? In the experimental section, the random seed is fixed at 42, which is unexplained.\n\nClarity: 2/5\n\nPros: The introduction provides a clear and convincing narrative.\n\nCons: The figures do not effectively support comprehension. For example, while Figure 1 illustrates the different outcomes of EQL and AFL, fully understanding it requires a detailed study of both paradigms. Figure 3 is even more challenging, as the components are not well explained in the caption, and there are many variants with minor changes. Figure 4 lacks sufficient details for understanding the experiments and data fully. \n\nSignificance: 2/5\n\nScalability is the primary limitation impacting the influence of this work on the neurosymbolic community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work proposes a logic formula learner framework, which comprises of three components: LFL-Type1 learns arbitrary logical formula, LFL-Type2 learns a look-up table, and LFL-Type3 has combinatorial search freedom between them. This frame work is end-to-end differentiable, can converge in a single run, and can learn arbitrary logic formula with the symbolic module." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "See strengths 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": 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": "It is not completely clear to me to what degree other approaches rely on the same amount of practical/engineering tricks to achieve proper convergence. I'm under the impression that e.g. ILP does not need this as heavily, but I also have to admit to not being an expert on logic formula learning." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Very well written, clearly presented.\n\nThe problem is interesting, as learning logical formulae opens many doors for NeSy learning.\n\nThe approach seems to be mostly original: The foundations (DSL and dNL) were clearly acknowledged and delineated from original contributions by the authors." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the LFL (Logical Formula Learner), an architecture that entails and expands prior work, in particular DSL (Deep Symbolic Learning) and dNL (differentiable Logic Network) . Both prior works feature learnable logical representations. LFL proposes an architecture that includes both DSL and dNL, but within LFL the user can choose more or less representational freedom. This architecture also includes MLPs as loss shortcuts and several loss terms to control convergence.\n\nEvaluation is done on multi-digit MNist addition, MNist-addition and on 3-layer logical formula." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Figure 4(d) seems a bit dishonest. Apparently DSL's poor performance is down to solely training mode settings. This should be presented more clearly and transparently, e.g. by showing test performance.\n\nThe paper could use a comparison to some of the competing branches of related work, e.g. differentiable/neural ILP.\n\nVarious typos:\nTypo achieved / archived bottom of page 1.\nFig 3's caption: references to \"such as [9] and [16]\" should probably read \"such as Equations [9] and [16]\".\n\"In these networks we also constrain**t**\" -> \"In these networks we also constrain\"." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "- The paper is missing several existing works on learning end-to-end learning: [1-5], of which [1-4] also learn proper logic rules. Furthermore, [4-6] are papers that also use an MLP loss / 'gradient shortcut' to (pre-)train the classifiers, which is critical to getting these systems to work. This is therefore not a new trick. \n- I understand the relation to EQL, and it's good to cite it, but outside of inspiration, these two methods are not that directly related? It seems like LFL is more like differentiable inductive logic programming in the sense of [7] than symbolic regression. \n- The footnote 2 is not correct: Scaling these weights does more than scaling the learning rate as it changes how close to 0.5 the average values of g(w_i) are. This means the derivatives of other parts of your system will change.\n- The 'LFL' framework is not clearly and formally defined. What are the constraints on the different choices of the functions?\n- The motivation for the linear reconstruction model is not clear to me. Is this purely for debugging? \n- The data in MNIST Addition is not binary, so why is a binary cross-entropy loss used? \n- What is the motivation for the label loss? \n- Figure 3 can be trimmed significantly: A single one of these 4 images should suffice by just denoting what is optional. \n- The architecture for multi-digit MNIST Addition in DSL (which I think is copied here) is highly specific to this task - How would this extend to different tasks?\n- An experiment on data different than MNIST would be nice - MNIST is easily clusterable, making the MLP shortcut a useful signal, but this isn't the case for all data. \n- The dataset setup for MNIST Addition is not correct and uses infinite data, while the canonical setup in the community is to take the training images, and randomly create a partition of them for different sums, giving finite data. The dataset is then about 60.000/2N in size - making the problem harder to train for larger multi-digit problems. This should be fixed.\n- The relevance of experiment 3.1.1 could be clearer\n- Experiment 3.2: I don't understand this result. There is nothing wrong with different training vs inference behaviour and this is standard in many DL layers. The label on Figure 4.d is unclear: This is _test_ accuracy under _training_ mode, right? \n\nMinor\n- The writing contains some typos. Eg: 052: Archived by - achieved by, 519: Conclution -> conclusion\n- Be careful with quotes '', they should be different when opening\n\n\n[1] Li, Zenan, et al. \"Neuro-symbolic learning yielding logical constraints.\" Advances in Neural Information Processing Systems 36 (2024).\n\n[2] Wang, Po-Wei, et al. \"Satnet: Bridging deep learning and logical reasoning using a differentiable satisfiability solver.\" International Conference on Machine Learning. PMLR, 2019.\n\n[3] Cunnington, Daniel, et al. \"The role of foundation models in neuro-symbolic learning and reasoning.\" International Conference on Neural-Symbolic Learning and Reasoning. Cham: Springer Nature Switzerland, 2024.\n\n[4] Aspis, Yaniv, et al. \"Embed2Rule Scalable Neuro-Symbolic Learning via Latent Space Weak-Labelling.\" International Conference on Neural-Symbolic Learning and Reasoning. Cham: Springer Nature Switzerland, 2024.\n\n[5] Daniele, Alessandro, et al. \"Simple and Effective Transfer Learning for Neuro-Symbolic Integration.\" International Conference on Neural-Symbolic Learning and Reasoning. Cham: Springer Nature Switzerland, 2024.\n\n[6] Aspis, Yaniv, et al. \"Embed2Sym: scalable neuro-symbolic reasoning via clustered embeddings.\" Kern-Isberner G, Lakemeyer G, Meyer T, editors. Proceedings of the 19th International Conference on Principles of Knowledge Representation and Reasoning (KR2022); 2022 Jul 31-Aug 5; Haifa, Israel.[California]: IJCAI Organization; 2022. p. 421-31.. International Joint Conferences on Artificial Intelligence Organization, 2022.\n\n[7] Evans, Richard, and Edward Grefenstette. \"Learning explanatory rules from noisy data.\" Journal of Artificial Intelligence Research 61 (2018): 1-64." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The task of learning both rules and classifiers is quite challenging, and the addition of noise via the Concrete distribution could be a good solution to this." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors study three variants of differentiable inductive logic programming to both learn end-to-end neural classifiers and rules in a NeSy predictor setup. The architecture is based on a form of fuzzy operators with added stochasticity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Currently, the paper is hard to follow, with a lot of abbrevations (it is difficult to remember what LFL-Type1-3 are!). The problem setup is not well-defined. I could infer it from my background, but this should be clear. I found it difficult to follow exactly how the 5 models (dNL, DSL and LFL-Type-i) fit in. The different LFL types are poorly motivated, and just are defined separately. Finally, the paper misses some critical related work, limiting the novelty of the paper. \n\nThe paper is limited in evaluation. It is only evaluated on a single task (MNIST Addition) and has a major problem in dataset setup which I will discuss below. Some different datasets, ideally also with different visual datasets than MNIST, would help with convincing that this model can properly learn." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. There is no formal definition of ϵ-greedy policy in line 171, page 4. \n2. There is no formal definition or citation for Gödel t-(co)norm in line 215 page 4. \n3. The input of the softmax function is a vector. However, the input of the softmax function defined in Equation (9) line 173 page 4 is a real number, which is hard to be understood by readers." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors conduct experiments with multiple settings to learn rules from a CNN model. The authors open their source code for the reference." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a general framework of network modules that explicitly equate a logical formula after the convergence of another neural network. The model is end-to-end differentiable, and training all modules from scratch achieving joint convergence in a single run, and explicitly learning arbitrary logical formulas (within limited complexity) with its symbolic module.\n\nHowever, I think this work is borderline rejected because of the following reasons:\n1. Some notations are not defined clearly and are used wrong. \n2. The structure of the manuscript is too messy to determine which part is the novelty and which is the preliminaries." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The structure of the paper is not clear to present which part is the novel contribution and which part is the preliminary work. For example, Sections 2.2.1 and 2.2.2 describe the Differentiable Neural Logic Network (dNL) Deep Symbolic Learning (DSL), these statements should be listed together in Preliminaries, not the contribution section. Furthermore, the terminology in Sections 2.2.1 and 2.2.2 should have citations. \n2. Some formats do not fit the standard such as ‘define in 6’ in line 156 and ‘16 is also’ in line 241. \n3. Some definitions are not very clear such as the ‘binary data’ in 103. \n4. The authors only compare performance with Deep Symbolic Learning (DSL) proposed by Daniele et al., 2022. I think there should be more benchmarks to prove the performance of the proposed methods experimentally." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose the Logical Formula Learner framework, a general framework of network modules that explicitly equate a logical formula after convergence." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning Arbitrary Logical Formula as a Sparse Neural Network Module},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x3cFAoorct},\nnote={under review}\n}" }, "abstract": { "value": "NeSy (Neuro-Symbolic) predictors are hybrid models composed of symbolic predictive models chained after neural networks. Most existing NeSy predictors require either given symbolic knowledge or iterative training. DSL (Deep Symbolic Learning) is the first NeSy predictor that supports fully end-to-end training from scratch, but it learns a look-up table rather than arbitrary programs or formulas. We propose the Logical Formula Learner framework, a general framework of network modules that explicitly equate a logical formula after convergence. We then propose 3 novel designs within the LFL framework with different levels of combinatorial search freedom: LFL-Type1 learns arbitrary logical formula, LFL-Type2 learns a look-up table, and LFL-Type3 has combinatorial search freedom between them. LFL-Type1 and LFL-Type2 show improvements over previous designs, and all three types can be wrapped into NeSy predictors. To our knowledge, LFL-Type1-based NeSy predictor is the first NeSy predictor that supports fully end-to-end training from scratch and explicitly learns arbitrary logical formulas." }, "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 AI; System 2 intelligence; Deep Symbolic Learning (DSL); Equation Learner (EQL); differentiable Neural Logic Networks (dNL)" ] }, "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/0b59910fdeafd65685fbd5f16da7e0539851908f.pdf" }, "presentation": null, "primary_area": { "value": "neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/28416bbec966182b62f055df578b49d68848cb63.zip" }, "title": { "value": "Learning Arbitrary Logical Formula as a Sparse Neural Network Module" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "What is the location parameter m in the caption of table 3?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors propose an upper bound for the variance of loss gradient estimates during training of flow-based generative models.\n2. With the VR sampling strategy, the number of training iterations required to achieve similar performance is significantly reduced compared to the baseline." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors theoretically identify the high variance in loss gradient estimates at intermediate training timesteps in flow-based generative models, and this variance affects influences the convergence of the optimization process during training. To address this issue, they construct an upper bound for the average total gradient variance using a function related to the signal-to-noise ratio (SNR) and propose a Variance-Reduction Sampling (VR-sampling) strategy. This strategy prioritizes sampling timesteps from high-variance regions more frequently to improve training efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. There is no comparison with the state-of-the-art methods, and the results in Table 2 are significantly worse than the current state-of-the-art performance.\n2. The paper lacks a detailed introduction to the baseline used for comparison in the results tables.\n3. The paper does not provide an analysis of how the VR reduction strategy enhances the qualitative results.\n4. In Section 3.3, the sampling process in the proposed strategy is not clearly explained. The current description mentions calculating normalization, followed by the PDF and the inverse CDF, but it is unclear how the probability density function π(t) is derived. Is π(t) the result of the inverse CDF? Additionally, there is no information on the final distribution from which the samples are generated." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. Based the discussion and method in the paper, during training, we will sample more high-variance regions for training. Will this cause overfitting on high-variance regions?" }, "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 provides the theoretical analysis for the training of the flow model.\n\n2. Based on their findings and theoretical analysis, they propose a simple but quite effective method to accelerate the training.\n\n3. I think this method can also be applied to other models easily and the conclusion will still be valid." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1. This paper first analyzes the reason for training issues in Flow models and finds the important role of high-variance regions.\n\n2. Based on the theoretical analysis, this paper proposes a variance reduction sampling strategy to sample more timesteps with high variance to accelerate model convergence.\n\n3. The experimental results show that the acceleration is significant." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The complexity and overhead of Monte Carlo simulations. This work heavily depends on Monte Carlo simulation, which is complex and time-consuming. Although in the paper, the time is much faster than full training, it may still cause a large overhead. I think this paper should involve more analysis of several aspects that influence the speed and performance of simulation such as the higher number of samples and higher data dimension." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* Could the authors explain more detailedly about the VR-sampling strategy (line 315~323)? \n* Could the authors offer a figure like FIgure.5(a) in [1] to show the results on how the VR-sampling strategy improve the gradient variance?\n\n[1] Tero Karras, et al. Elucidating the design space of diffusion- based generative models." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* This papers provides theoretical analysis on gradient variance for conditional flow matching loss and proves the convergence rate during training.\n* The experimental results are clear and well verified the effectiveness of the VR-reduction sampling strategy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a Variance Reduction-based diffusion timesteps sampling strategy to accelerate the training of DMs. The authors identify that the variance of gradient estimates in the training process is higher at intermediate timesteps, which affects training stability. They propose VR-sampling to prioritize sampling from high-variance regions, thereby accelerating training. The method is shown to significantly speed up training across various noise schedulers and data dimensions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper provides a theoretical analysis of the variance of the DSM loss and the convergence speed of DMs training in SGD. Based on this analysis, the paper proposes a method to accelerate training of DMs. While the experimental results appear promising, there is a lack of analysis explaining why the VR-Sampling strategy, specifically \"sampling timesteps in high-variance regions more frequently,\" leads to improved convergence speed.\n\n2. Building on point 1, it seems that the proposed method is anagolous to some empirical findings, such as the Log-normal and logit-normal techniques mentioned in [1] and [2]. The relationship between the theoretical insights and the proposed algorithm could be made more explicit.\n\n3. Section 4.2 is not entirely convincing. The results on the baseline models are not as strong as those reported in the original paper. For instance, in [3], the FID score (cfg=1.5) is 2.06 and 2.62 respectively on ImageNet 256 and ImageNet 512. In contrast, the best reported FID score for the baseline in this paper is 8.34, as shown in Table 2. While I understand that the experiments have controlled the number of training iterations, I believe a comparison of the convergence points is also necessary to provide a more comprehensive evaluation.\n\n4. On line 290, the Monte Carlo approximation is somewhat confusing. Are you using samples $x_1\\sim q(x_1)$ to compute both the outer expectation and inner expectation? Could the authors please provide a more detailed explanation to clarify this concept?\n\n**minor points** \n* Equation at line 290 is not labeled.\n\n[1] Tero Karras, et al. Elucidating the design space of diffusion- based generative models.\n\n[2] Patrick Esser, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML 2024.\n\n[3] Nanye Ma, et al. Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 it right that the proposed VR-sampling is only performed in the training phase, not in the inference phase?\n- VR-sampling can support different choice of Noise Scheduler (Diffusion, Linear, Cosine), but it seems currently only tested on deterministic ODE formulation. Is it possible to extend it to SDE formulation (as a future work)?\n- Does Figure 2 show that S is highly correlated with the resolution, as both CIFAR and ImageNet 256 are using 64x64 resolution?\n- Could the author provide some insights how the cfg value can affect the results? I can see with higher cfg in Figure 3, 4, 5, the FID curve for VR-sampling can converge more similar to the baseline (Fig. 3). That said, what would you expect to happen if cfg is higher like 4.0, would the method still clearly outperform the baseline?\n- In line 324-331, the author mentioned VR-sampling takes extra time to find the probability density function (PDF) and construct $x_t$. To my understanding, this process requires evaluating the equation (missing equation number?) in line 290 and the term $p_t(x|x_1,i)$ using the neural network prediction. But I am not clear which state/checkpoint of the neural network is used to compute this term?\n- It is briefly mentioned \"Perception prioritized training of diffusion models\" use heuristics to sample time steps based on signal-to-noise ratio (SNR). But I am wondering how it compares to the proposed VR-sampling method?\n- The improvement seems more visible in Figure 13, than the improvement shown in Figure 14, 15. Does it give a hint that VR-sampling might work better with higher reoslution images or latents?\n\nSome minor comments:\n- I would suggest Figure 13-15 can be larger to better see more details and utilize the space.\n- In table 1 \"Ho et al., 2020b\" is the same as \"Ho et al., 2020a\". It is cited twice.\n- In table 6, how CFG scale is used in training? According to \"Classifier-free diffusion guidance\" [Ho and Salimans 2022], the training does not use the value directly, but with a random discard of the conditioning to train unconditionally." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- How to sample time steps is indeed an important problem in traininig diffusion models, the method is with a clear motivation.\n- The paper is well-written and easy to follow.\n- The results including FID curves and tables clearly and consistently show the effectiveness of the proposed method on different variants of flow models and model architectures, which are widely used in generative modeling recently.\n- The proposed method is also supported with necessary proof of bounds and convergence analysis.\n- It is appreciated that the paper includes an anonymous code link for reproducibility." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a Variance-Reduction Sampling (VR-sampling) strategy that samples the time steps in high-variance region more frequently to enhance training efficiency in diffusion models. The method first identifies the root cause of the high variance and proposes an effective solution for importance sampling. The proposed strategy is also supported with proof of bounds and convergence analysis. The results show that the proposed method outperforms various baselines, experimented on ImageNet and CIFAR datasets and different model architectures including U-Net, DiT and SiT." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper might lack some dicussions to related method in diffusion models. For example, in line 196, Iterative α-(de)Blending [1] also shares the same formulation as rectified flow. In Monte Carlo sampling, variance reduction can be achieved by using not just importance sampling, but also control variates and correlated sampling, which are initially explored [2,3] for diffusion models. Also, there exists similar line of work considering variance reduction in deep learning such as Deep Learning with Importance Sampling [4] (and there are more), which might also be worth mentioning.\n- The visual improvements of VR-sampling shown in Figure 14, 15 seem not significant to me. It will be great to have results on one of the Animal/Human faces (e.g., AFHQ, CelebA, FFHQ), or one of the LSUN datasets. As observed in Figure 13 and 14, the method seems to be working better on face images. Also, results on LSUN bedroom or church might be interesting as the images are with specific textures and patterns.\n\n\n[1] Iterative 𝛼-(de)Blending: a Minimalist Deterministic Diffusion Model, Heitz et al. 2023\n\n[2] Variance reduction of diffusion model's gradients with Taylor approximation-based control variate, Jeha et al. 2024\n\n[3] Blue noise for diffusion models, Huang et al. 2024\n\n[4] Not All Samples Are Created Equal: Deep Learning with Importance Sampling, Katharopoulos and Fleuret 2018" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024vrsampling,\ntitle={{VR}-Sampling: Accelerating Flow Generative Model Training with Variance Reduction Sampling},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x3jRzVAltZ},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in text-to-image and text-to-video models, such as Stable Diffusion 3 (SD3), Flux and OpenSora, have adopted rectified flow over traditional diffusion models to enhance training and inference efficiency. SD3 notes increased difficulty in learning at intermediate timesteps but does not clarify the underlying cause. In this paper, we theoretically identify the root cause as a higher variance in the loss gradient estimates at these timesteps, which hinders training efficiency. Furthermore, this high-variance region is significantly influenced by the noise schedulers (i.e., how we add noises to clean images) and data (or latent space) dimensions. Building on this theoretical insight, we propose a Variance-Reduction Sampling (VR-sampling) strategy that samples the timesteps in high-variance region more frequently to enhance training efficiency in flow models. VR-sampling constructs sampling distributions based on Monte Carlo estimates of the loss gradient variance, allowing it to easily extend to different noise schedulers and data dimensions. Experiments demonstrate that VR sampling accelerates training by up to 33\\% on ImageNet 256 and 50\\% on ImageNet 512 datasets in rectified flow models. Furthermore, VR-sampling could simplify the hyperparameter tuning of logit-normal sampling introduced in SD3. \nThe code is available anonymously in~\\url{https://github.com/AnonymousProjects/VR_sampling.git}." }, "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": [ "Flow Generative Models", "Training Acceleration", "Diffusion 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/d1e1398988d672b59bbdeb4eda67284fa0e9da4c.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "VR-Sampling: Accelerating Flow Generative Model Training with Variance Reduction Sampling" }, "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": "Please see weaknesses" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-written and easy to follow.\n2. The experiments in Table 1 show the efficacy of the method on the given dataset.\n3. Fast-SQN solves the computational efficiency issue by cleverly adding a spline interpolation layer." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a normalisation scheme called Structural Quantile Normalisation (SQN). Unlike many other methods, the proposed transformation is differentiable. The authors blend in the ideas of quantile normalisation, kernel density estimation, and PCHIP to propose their method. The method can be slow in its true form, but the authors provide a modification called Fast-SQN to counter it." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Table 1 shows better performance of the proposed method but lacks the error bounds, it might be helpful to see the fluctuation if different random seeds are employed for the same model.\n2. Authors cited that Gaussianisation techniques can mean losing intrinsic patterns in the data and are non-differentiable. However, I would argue that methods like Normalising Flows are differentiable, and a few such layers in the beginning might help with this. It would be interesting to see if NF-transformed data lags in performance compared to the methods reported in Table 1. \n3. Only one dataset is discussed, it is hard to say if the results might hold for any kind of data.\n4. Can this transformation also help with Image data? or use of KDE limits applicability to high dimensionality? \n5. Can using these transformation layers with methods that aim at transforming the distribution to standard normal can we reduce the computational complexity of those methods?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See weakness" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The proposed methods are well-motivated and explained well. It covers existing methods as special cases and can be computed efficiently with the proposed FAST-SQN version." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a feature scaling method that transforms a feature such that the feature is close to normally distributed after the transformation. A kernel density estimation is fitted to model the pdf of the feature, then the inverse Gaussian cdf is applied to the cdf estimate of the feature. The kernel density estimation allows the method to keep the local structure of the original distribution. Experiments in one data set demonstrate that the proposed methods outperform other scaling methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Empirical validation is not enough\n\n1. Only one real data experiment is provided\n2. It is hard to tell if the performance improvements are significant, e.g. in table 1, model 3, Fast-SQN only improves RMSE by 0.001 over YJN, mean and standard deviation over multiple runs should be reported.\n3. The proposed method should be a general scaling method for any feature, but it seems only the target feature is normalized in the experiments (line 390 in the paper). I thought rescale input features would be more useful in model training.\n4. One advantage that is emphasized for the proposed method is its differentiability, but if it is only used in the experiment setting in the paper, differentiability doesn't seem to have any advantage.\n\nOverall, I think the empirical validation is far from convincing. The proposed methods should be able to apply to a wide variety of data. Much more experiments are needed to really demonstrate that it is superior to existing methods. If it can indeed improve model performance across a lot of data sets and especially if it can be applied similarly to batch-norm or layer-norm layer, this could be a major contribution to the community, but the current results don't support the claim." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The proposed method seemingly only works in one dimension. In multi-dimensions, should one use Sklar's Theorem, if so, what do you do with Copulas? If not, how does this apply to general probability measures on $\\mathbb{R}^d$?\n\n- What is the regularity (continuous, uniformly continuous, etc... for some reasonable metrization of the weak topology on a reasonable subset of $\\mathcal{P}(\\mathbb{R})$) of this operation in general? Why should it be invertible in high-dimensions? \n\nThe most major question I have is: *Why use this method at all?* Specifically:\n\n- Is there any *provable* mathematical theoretical guarantee that it *must* improve downstream learning? \n\n- If you cannot provide a guarantee, then there must at least be convincing experimental evidence (what is offered certainly is not).\n\n- What is the take-way of Propostion 1, why is it important, and what is the concrete implication on the proposed method? \n\nEven if the above questions could be answered, I must ask: how does this work in high dimensions? This is not obvious to me, I'm assuming it is not to the authors either or the method would have been presented in greater generality." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The authors provide a reasonable literature review and the graphs are very elegant." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a new data-standardization method *applicable only to one dimensional data* which uses a specific procedure to monotonically perturb a KDE of a *one dimensional* applied to empirical measure, mapping it to a centered normal law. \n\nSome extremely limited numerics on very a small dataset (used in basic textbooks and old Kaggle competitions) is used to evalued at the normalization procedure." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are no real justification of support of the method whatsoever. \n\n- There is no theoretical support (there is a proposition in the appendix, which is formulated non-rigorously. What time of convergence is this? The KDE is a function of a random quantity, the empirical (random) measure associated with some one-dimensional law, then should I interpret this as $\\omega$-wise convergence? \n\nTherefore, this is no theoretical contribution.\n\n- The numerics are not convincing, only a toy dataset (the California Housing Market) dataset is considered, which is something on sees in introductory text books. Thus, there is also no experimental support for the proposed method.\n\n\n---\nBtw, the convergence in the proposition in the appendix is in the point-open topology, perhaps that should be mentioned (as the limit is difficult to interpret without reading the first line of the proof). I think this should be clearly said in the statement." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "With respect to weakness 2 above, I would suggest combining with either linear methods (which still have some value due to interpretability) or modern neural network approaches like TabPFN. Or perhaps the authors could evaluate their approach on neural network methods -- eg CSDI-T (Zheng & Charoenphakdee, TRL @ NeurIPS 2022) or TabDDPM (Kotelnikov et al, ICML 2023) -- for tabular data generation, which also require feature preprocessing." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Feature preprocessing is an important task, as even deep learning methods such as TabPFN require preprocessing to deal with highly-skewed features.\n\n2. Using PCHIP splines to achieve linear scalability while maintaining smoothness and monotonicity is an interesting proposal." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "SQN is a feature preprocessing method that interpolates between z-score scaling and quantile normalization via the Gaussian KDE, utilizing PCHIP splines to achieve linear scalability while maintaining smoothness and monotonicity. This method is evaluated on the California housing dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The experimental section is quite weak, only evaluating on the California housing dataset. The results would be much more convincing if the authors showed benefits on a benchmark of tabular datasets. And, while the authors claim that differentiability is valuable, but don't show that this provides any empirical benefits.\n\n2. The authors evaluate their method in combination with feedforward neural networks, which are not even close to SotA for tabular data. \n\n3. The originality of the proposal and the completeness of the related work section are seriously affected by the fact that the overall proposed approach is the same as KDIT (The Kernel Density Integral Transformation, McCarter, TMLR 2023). In both cases, one uses the KDE to smooth the pdf of a feature, then applies the inverse cdf of a reference distribution.\n\nGranted, there are some differences, but these are minor:\n\n(A) While the KDIT software package implements using the Gaussian as the reference output distribution, the KDIT paper only evaluates the uniform distribution as reference, thus interpolating between min-max scaling and quantile normalization, while SQN interpolates between z-score scaling and quantile normalization.\n\n(B) KDIT uses the polynomial-exponential kernel (Fast exact evaluation of univariate kernel sums, Hofmeyr, TPAMI 2019) to obtain linear complexity, while SQN uses PCHIP splines for this. PCHIP splines are superior in that they enforce not just monotonicity, but also smoothness; whereas the KDIT is only almost-everywhere smooth. It's not clear that everywhere smoothness is really valuable, given that the authors combine their approach with neural networks with not-everywhere-smooth ReLU activations." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a differentiable feature scaling technique that balances Gaussian approximation and structural preservation, outperforming existing methods in multiple error metrics on real-world data." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024structural,\ntitle={Structural Quantile Normalization: a general, differentiable feature scaling technique balancing gaussian approximation and structural preservation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x3l0fQubOn},\nnote={under review}\n}" }, "abstract": { "value": "Feature scaling is an essential practice in modern machine learning, both as a preprocessing step and as an integral part of model architectures, such as batch and layer normalization in artificial neural networks. Its primary goal is to align feature scales, preventing larger-valued features from dominating model learning—especially in algorithms utilizing distance metrics, gradient-based optimization, and regularization. Additionally, many algorithms benefit from or require input data approximating a standard Gaussian distribution, establishing \"Gaussianization\" as an additional objective. Lastly, an ideal scaling method should be general, as in applicable to any input distribution, and differentiable to facilitate seamless integration into gradient-optimized models. Although differentiable and general, traditional linear methods, such as standardization and min-max scaling, cannot reshape distributions relative to scale and offset. On the other hand, existing nonlinear methods, although more effective at Gaussianizing data, either lack general applicability (e.g., power transformations) or introduce excessive distortions that can obscure intrinsic data patterns (e.g., quantile normalization). Present non-linear methods are also not differentiable. We introduce Structural Quantile Normalization (SQN), a general and differentiable scaling method, that enables balancing Gaussian approximation with structural preservation. We also introduce Fast-SQN; a more performance-efficient variant with the same properties. We show that SQN is a generalized augmentation of standardization and quantile normalization. Using the real-world \"California Housing\" dataset, we demonstrate that Fast-SQN outperforms state-of-the-art methods—including classical and ordered quantile normalization, and Box-Cox, and Yeo-Johnson transformations—across key metrics (i.e., RMSE, MAE, MdAE) when used for preprocessing.\nFinally, we show our approach transformation differentiability and compatibility with gradient-based optimization using the real-world \"Gas Turbine Emission\" dataset and propose a methodology for integration into deep networks." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "feature scaling", "preprocessing", "normal distribution", "differentiable transformation", "quantile normalization", "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/35f7066fa6d2282022689c85f8cd67eb9d852675.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/2f6348ab81083738649af1bc39c1483a76123269.zip" }, "title": { "value": "Structural Quantile Normalization: a general, differentiable feature scaling technique balancing gaussian approximation and structural preservation" }, "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": "(1) Please explain how you consider the role of early stopping, and what regime does the analysis and SANER experiments consider. \n\n(2) Does SANER improve over SAM even accounting for early stopping, i.e. is the best early-stopped checkpoint of SANER better than best early-stopped checkpoint with SAM?\n\n(3) I am sorry if I missed this, but how do you set the SAM perturbation radius hyperparameter. How does it interact with \\alpha? Do the trends of gradients across different groups hold for different radii as well?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper builds up and contributes to the literature on understanding the benefits of SAM. The paper is generally well written and easy to read. Understanding SAM is broadly conceptually interesting, and the application to label noise is practically well motivated. \n\nIt builds up ideas in a simple fashion, and lays out clear hypotheses being tested and implications of these hypotheses. The study of which parameters have their gradients upweighted, down-weighted or reversed is novel to the best of my knowledge. The broader implications of this analysis are a little unclear (detailed below), but the analysis is clean and clearly presented. \n\nThe paper provides a new and simple modification to SAM based on the analysis and performs a large-scale evaluation over many datasets and architectures. While I have some reservations about the soundness of some aspects, overall this was a clear effort for a thorough investigation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper builds on the observation that SAM improves robustness to label noise over SGD. The paper analyzes which parameters' gradients are affected in different ways, between the SAM and SGD update. This provides some insight into the importance of further decreasing the norms of those parameters that have a lower norm under the SAM perturbation. Empirical results seem to show gains when performing this modification of SAM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One glaring gap in the analysis of the paper is the effect of early-stopping. Early-stopping seems to improve clean test accuracy in both SAM and SGD. The paper does not make it clear whether the goal is to compare SANER, SAM, and SGD at the final checkpoint or best early stopped checkpoint. Even with early stopping, SAM outperforms SGD. However, from Fig 1(c) it seems like SAM vs SANER shows no gains if we do early stopping. In all the experiments reported later, it is unclear which checkpoints are compared. \n\nFrom an analysis perspective, the authors acknowledge the weakness of focusing on noisy train accuracy (rather than final test accuracy): intuitively fitting noise at training is bad, but for different algos, this could manifest differnetly in terms of affecting test accuracy. Furthermore, clean training might also be affected which can affect test accuracy. I agree with the authors on this weakness. \n\nFrom a practical perspective, if the gains do not hold with early stopping, the benefit is less clear. Furthermore, SANER introduces a new hyperparameter \\alpha and adds complexity in practice. \n\nThe role of the perturbation radius of SAM is not discussed - in practice, this is another important parameter. How does this radius interact with the new hyperparameter \\alpha introduced? Is it strictly better to tune \\alpha rather than the radius, or should we tune both in parallel, or do they interact in ways such that we need to tune over both in combination. \n\nOverall, the analysis is potentially clean and interesting, but for reasons above, the soundness / comprehensive analysis is missing some key aspects. There is no clear generalizable insights from this work, and gains in practice (if they hold up under further experiments) come at the cost of additional hyperparameters." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the comments on weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. In the context of increasingly large training datasets, it is practically significant to effectively prevent models from overfitting on noisy label data. Relying solely on algorithms and rules to clean the data is often too costly, so optimizing algorithms to reduce the impact of noisy label data is more valuable.\n2. The author's approach of comparing the relative sizes of the gradients of each parameter in the SAM and SGD algorithms during model updates, and using this as a criterion to assess whether the parameters are overfitting to the model due to noisy label data, is straightforward and exhibits a certain level of innovation. The experimental design is also very reasonable.\n3. The author's experimental design is quite rigorous, and they conducted a substantial number of experiments and tests on several public datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The author studied the relationship between the gradients of the SAM and SGD algorithms, and through experimental validation, found a method more suitable for enhancing the model's resistance to overfitting on noisy label data. The algorithm modifies the weights of different parameter gradients during model updates to improve the model's robustness. In terms of experiments, the author conducted validation on several classic public datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.I noticed that Figure 2 in Section 3 is based on training a ResNet-18 model on CIFAR-10. Shouldn't testing be conducted on more models and datasets to confirm that the three types of parameters, Group A, B, and C, indeed exhibit the proportions mentioned in the paper?Perhaps conducting experiments on a larger ResNet model (ResNet-50) and a larger dataset like CIFAR-100 or ImageNet (if conditions permit) would provide more convincing results.\n\n2.Based on the author's analysis, in fact, at epoch 100, there was a noticeable increase in the parameter ratio of Group C, which maintained around 10% during the early training stages. Would it be more meaningful to directly use the reverse gradients for noisy label data? Also, I noticed that the ratios of Group B and Group C are quite similar in the early stages of training, and the values for Group A are mostly around 1.0 while those for Group B are around 0.99. If this is the case, I don't believe that SGD and SAM have different perspectives on the parameters in these two groups.As before, I believe it is necessary to validate this distribution pattern on larger models and datasets. Additionally, I think it would be meaningful to further investigate whether there are significant changes in the parameters within the three groups as training progresses, as well as the specific distribution of values." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the weaknesses above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper reveals an interesting phenomena, showing that there are dimensions with lower magnitudes and the same signs as the corresponding components in SGD, that are responsible for superior robustness of SAM. This is interesting and novel to my knowledge.\n\nThe paper is clear and easy to follow. While I found the observation interesting, I believe the paper requires to back up the results with a more in-depth study (either some theoretical analysis or more in-depth ablation study and a wider range of experiments) to be considered significant." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper conducts an empirical study which reveals that during training, some gradient dimensions contribute to the superior robustness of SAM over SGD against noisy labels. These dimensions have lower magnitudes and the same signs as the corresponding components in SGD. Reducing the weight of such dimensions during training improves robustness of SGD and SAM variants against label noise." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the proposed idea of reweighting dimensions with lower magnitudes and same signs as SGD is simple (simplicity is good), it requires an appropriate linear scheduler that gradually decreases the reweighting factor from 1 to a predetermined value over k epochs. Both the reweighting factor and k needs to be tuned and I didn't find an ablation study or any insight on how to set these parameters for different datasets/architectures. Table 5, 6 in the appendix only compares using or not using this scheduler (k=0 and k=50).\n\nConsidering that this is an empirical study, I'd expect much more in-depth empirical study of the observed phenomena. For example, what is the effect of datasets and architectures on the observed phenomena? What's the effect of the label noise level on the observed phenomena? How should one set the parameters of the proposed method? The paper includes experiments on Cifar10-100 (and their subsampled versions) with label noise level of 25% and 50%, and Webvision. On Cifar-10, the improvement is not significant, while on Cifar-100 the proposed method works much better. Nevertheless, there is no explanation or deeper analysis provided in the paper. Existing methods for robust training against noisy labels usually perform much better on Cifar10. Considering that the pattern is different for the proposed method, digging deeper into why things work differently would be an interesting addition to the paper. The authors can also analyze the effect of the proposed reweighting methods combined with existing robust training methods, as is done in the original SAM paper.\n\nFinally, I believe adding a theoretical study of the observed phenomena using a simple data model and network architecture, or even a simple toy example explaining the effect of reweighting would significantly strengthen 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": 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）Regarding the division of gradient components into Group A, B, and C, what specific criteria or methodology did you use for this grouping? Could you provide pseudocode for this process to make the grouping mechanism clearer? Additionally, could you explain why these particular components were chosen for reweighting, and include systematic mathematical derivations to show how these components effectively suppress noisy label fitting?\n2）The paper includes comparisons with VaSSO; however, the results under different noise rates show considerable gaps. Could you explain why such significant differences occur under identical experimental settings? Moreover, could you conduct additional experiments covering more diverse noise ratios to provide a balanced comparison, and clarify whether these differences are due to inherent limitations of the methods or specific experimental setups?\n3）In the appendix, only ablation studies for the presence or absence of the α parameter were conducted, but there is a lack of detailed experimental analysis on the selection of α values (e.g., from 0 to 1 or values greater than 1). How do different values of α affect performance across varying noise ratios and network architectures? Providing such detailed experimental support would help in understanding the optimal choice of α under different scenarios." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1）SANER introduces a novel reweighting mechanism for specific gradient components, effectively extending SAM to better handle noisy labels. This innovation is well-motivated and provides a meaningful contribution to the field of noise-resistant optimization.\n2）The experimental validation convincingly demonstrates that SANER outperforms SAM and other optimizers in different noise scenarios. The visualizations provided further reinforce the claims regarding the noise robustness of SANER.\n3）The use of a linear schedule for adjusting parameters during training is practical and helps stabilize the learning process, especially in the early stages when noisy labels are more problematic." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents SANER, a novel approach designed to improve the robustness of Sharpness-Aware Minimization (SAM) in the presence of noisy labels. The proposed method introduces a reweighting mechanism for gradient components to suppress the fitting to noisy labels, thereby enhancing generalization performance. Extensive experiments are conducted on various datasets, demonstrating the effectiveness of SANER compared to SAM and other optimizers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1）The paper lacks a detailed explanation of how the gradient components are divided into Group A, B, and C. It is unclear what specific criteria were used for this division, and including pseudocode would enhance clarity and reproducibility. Additionally, while visual evidence supports the reweighting mechanism, systematic mathematical derivations justifying the effectiveness of the approach in suppressing noisy label fitting would be valuable.\n2）The manuscript includes comparisons with VaSSO, but there are notable gaps in performance under different noise rates, even with identical experimental settings. Clarifying why these significant differences occur would improve the discussion. Moreover, additional experiments with varied noise ratios would help provide a more balanced comparison. Furthermore, comparing SANER with methods such as AdaSam and Adan could better help demonstrate its broader generalization capabilities or effectiveness across different tasks and domains. AdaSam is primarily used in natural language processing tasks, while Adan is a recent adaptive Nesterov momentum algorithm showing promise in vision, language, and reinforcement learning tasks. \n3）The analysis of the sensitivity of the parameter α is limited. A more comprehensive study covering different noise ratios, network architectures, and training conditions would be beneficial. For instance, exploring α’s behavior in noisy segmentation tasks using Vision Transformers (e.g., ViT-based noisy segmentation) as well as standard classification tasks with different architectures (e.g., ResNet versus VGG) would provide a clearer picture of the parameter's impact across varied 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": 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. Regarding Figure 3, since Group A also contributes to SAM’s resistance to noisy fitting, why not reweight Group A as well?\n2. Since SAM enhances model performance in clean label scenarios, would SANER perform worse than SAM in such cases?\n3. Are there any improvements when combined with other label noise learning algorithms?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The writing quality is good.\n2. This manuscript provides a detailed experimental analysis of SAM and its robustness to label noise.\n3. The proposed method is tested across different model architectures, layer widths, and data sizes, demonstrating its generalization capability across different settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript explores the potential of SAM in the context of noisy label learning. Specifically, it provides an empirical study of the component-wise gradients in SAM, identifying the gradient components most crucial for handling noisy labels. Building on these insights, this manuscript proposes SANER, a method that adjusts SAM gradients to improve model robustness against label noise." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The theoretical analysis is lacking. Specifically, there is no justification for why downweighting the gradient component in Group B helps the model resist label noise, or why gradients from noisy samples dominate Group B as the model begins to overfit. A theoretical analysis could suggest a more effective gradient adjustment method than simple downweighting.\n2. Some experimental results are missing.\n 1. Experimental results for high label noise settings (e.g., 80% symmetric label noise) are absent. Given that this manuscript primarily addresses label noise learning, evaluating the proposed method under high label noise conditions is essential, as it is a common practice in current label noise learning literature.\n 2. The clean accuracy results for different values of $\\alpha$ are missing. Recall that $p_\\text{clean}$ still accounts for 1 / 3 of the total components in Group B, so downweighting Group B may potentially harm the accuracy of clean samples." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024improving,\ntitle={Improving Resistance to Noisy Label Fitting by Reweighting Gradient in {SAM}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x3lE88YkUl},\nnote={under review}\n}" }, "abstract": { "value": "Noisy labels pose a substantial challenge in machine learning, often resulting in overfitting and poor generalization. Sharpness-Aware Minimization (SAM), as demonstrated in Foret et al. (2021), improves generalization over traditional Stochastic Gradient Descent (SGD) in classification tasks with noisy labels by implicitly slowing noisy learning. While SAM’s ability to generalize in noisy environments has been studied in several simplified settings, its full potential in more realistic training settings remains underexplored. In this work, we analyze SAM’s behavior at each iteration, identifying specific components of the gradient vector that contribute significantly to its robustness against noisy labels. Based on these insights, we propose SANER (Sharpness-Aware Noise-Explicit Reweighting), an effective variant that enhances SAM’s ability to manage noisy fitting rate. Our experiments on CIFAR-10, CIFAR-100, and Mini-WebVision demonstrate that SANER consistently outperforms SAM, achieving up to an 8% increase on CIFAR-100 with 50% label noise." }, "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": [ "label noise", "sharpness-aware minimization", "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/cb4ded673972927aaabd709350bb6b7269d13f00.pdf" }, "presentation": null, "primary_area": { "value": "optimization" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/419eb160b3e6b48a4c20d11100a0abf5c3f350eb.zip" }, "title": { "value": "Improving Resistance to Noisy Label Fitting by Reweighting Gradient in SAM" }, "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": "* Could you provide more details about the datasets that you’re using? For example, how many datapoints are in each and how “diverse” are the different prompts\n* In 3.2 in the definition of PM(T), should it be lowercase pi?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The work points out an important shortcoming of language models, namely, an inability to accurately reflect probabilistic information given in a context. They further show how recent fine-tuning approaches affect this calibration\n* Many different models are evaluated, including the base and chat versions of several popular architectures, showing consistency and their observations and allowing for broader conclusions \n* The paper is well-written and the authors provide good motivation for the problem they’re exploring" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The study investigates the alignment of language models’ output probabilities with the numeric or probabilistic information in the contexts they’re given. They explore models that have undergone different fine-tuning techniques (instruction tuning, preference alignment) and see how this affects the model’s explicit reasoning in comparison to base models. They look at whether biases in token probabilities (e.g, a first-mentioned bias) can be identities. They find that across model architectures, language models are generally not well calibrated in this respect. Instruction-tuning seems to exacerbate the issue, often leading to mode-collapse. They also observe some interesting systematic biases for different model families/fine-tuning strategies. These findings highlight an important limitation of language models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The contributions of the work are rather limited. A very specific question is being asked and its unclear how relevant this question is to the broader community. While the work reveals an undesirable model behavior, it doesn’t propose methods for fixing these behaviors. There is a short discussion of potential reasons for the behaviors, but no empirical evidence for or against these hypotheses are given\n* Many of the results are missing standard errors or significance tests.\n* The enormous set of results in the appendix is difficult to navigate" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "> Title: Are Language Model Logits Calibrated\n\nThe word logits is not used in the paper at the moment and logits (the pre-softmax activations) are not analysed---only the probabilities output by LMs (post-softmax) are analysed. I thus believe the word \"logits\" doesn’t belong in the paper’s title. Further, the title—in my opinion—suggests the paper is about a more traditional notion of model calibration (e.g., Expected Calibration Error; ECE). Changing the title to highlight this different “view” of model calibration could be helpful. E.g., “Are Language Model Outputs Calibrated to Explicit Probabilistic Prompts” or something analogous.\n\n> Most Plots\n\nThe Figures in this paper are not readable when printed in black and white." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper investigates an interesting question: whether language models produce calibrated outputs when prompted with explicit probabilistic information.\n\nThe paper also proposes two related experiments to investigate model behaviour in this setting, analysing results in these two settings in detail." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper analyses how calibrated language model (LM) outputs are when prompted with explicit probabilistic information (e.g., prompt: “an urn has 5 blue and 3 red balls, I pick one, this ball is”, LM output: blue/red”).\nThey perform two main experiments:\n* “Distributions”. In this first experiment, they prompt the model with information about a uniform distribution (e.g., “Sampling from a uniform distribution with support from 2 inclusive to 5 exclusive, I sampled the number”) and evaluate the probability a LM places on tokens $\\{0, 1, 2, 3, …, 9\\}$.\n* “Probabilities”. In this second experiment, they prompt the model with an implicit distribution (e.g., “From 17 red marbles and 99 blue marbles, Billy reached blindly into the bag and grabbed a marble with the colour”) and see whether the model places calibrated probability on the tokens of interest (in this case, $\\{red, blue\\}$).\n\nIn the first experiment, the paper finds that instruction-tuned models are *less* calibrated than base models. It also finds that models have systematic preferences for some tokens (e.g., systematically preferring token 3 to others). In the second experiment, the paper finds that instruction-tuned models are *more* calibrated than base models (the opposite from the first experiment). \n\nThey also perform a human study, where they compare LM to human behaviour." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the paper performs two interesting experiments to investigate the role of explicit probabilistic calibration on language models, I believe that in its current state, the insights that can be drawn from it are somewhat limited. The paper performs a behavioural analysis on language models with two relatively simple and similar settings. I’d expect an ICLR paper to perform a more thorough analysis. Some suggestions below.\n\nI think the paper would be much stronger if it performed one of the following analysis (or both):\n* *Training data analysis.* While this may not be possible with the models currently analysed in the paper, the authors could extend their analysis to, e.g., Pythia or OlMO, and evaluate how the analysed LM behaviour relates to different statistics of the training data. The paper speculates about this in section 6, but this could be analysed.\n* *Mechanistic analysis.* The paper currently evaluates model behaviour purely as a black box. Performing mechanistic/causal analysis of how this behaviour relates to model internal activations could make the results more insightful. E.g., the paper could use distributed alignment search (Geiger et al. 2023) to find subspaces in the model which control the model’s behaviour on these tasks.\n\n\nGeiger et al. 2023. Finding Alignments Between Interpretable Causal Variables and Distributed Neural Representations" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "n/a" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- Well defined problem statement; the authors don't make any claims that they can't justify.\n- Thorough set of experiments to validate the existence of this problem, with very detailed methodology.\n- A human study that shows, interestingly, that humans are also not well-calibrated and that by some measures, the models are actually better calibrated than humans." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors find that LLM predictions are poorly calibrated for outputs that should ideally recover some aleatoric uncertainty, like the probability of drawing marbles of one color of out of a bag. Different models behave in different systematically uncalibrated ways, and sota models often allocate all of the probability to just one token." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There seems to be a lot of relevant work on LLMs expressing uncertainty that could be cited here (e.g., see https://arxiv.org/abs/2401.06730 and the related work discussed therein).\n- There is no discussion of conformal prediction and how it might be used to address this problem, it it hasn't been done so already.\n- In summary, the paper identifies a problem that is already well known in some capacity---albeit maybe not in the particular formulation proposed here. Although the paper is well-written and the problem is well-articulated, no real solution is offered. Because of this, I lean towards 'borderline reject' for an ICLR paper, though I could seeing it faring better at an NLP or CL conference." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. What would the result be like if you search over different temperature choices? \n\n2. Have you checked the actual text output of the models if they directly give the valid token in the next token position?" }, "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 is clearly written and easy to read. \n2. The idea of using context as the implicit likelihood for calibration is interesting and provides a good starting point for analysing the calibration level of language models.\n3. The human study is a good addition to show that both human and language models are biased, motivating the research for improving the calibration level of language models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the calibration of language models. It defines calibration as the degree to which the output distribution probabilities of the candidate tokens are aligned to the likelihood inferred from the context. \nThe author evaluated the case of drawing marbles on open and close-source models and showed that LMs are poorly calibrated in this case. \nThe author also conducted further analysis and a human study to compare the calibration level of the model to humans." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main and critical issue of this paper is that they didn't discuss the impact of the temperature of the softmax operation for getting the probabilities. The choice of temperature can greatly influence the entropy and final calibration level of the model, which will greatly affect the conclusion drawn from the paper. As shown in Table 1, changing temperature from 0.1 to 1.0 greatly changed the performance of the random baseline. \n\n2. The overall contribution of the paper is poor. The issue of low entropy of the Chat models and their token bias has been shown in different papers. The novelty of this paper is primarily on the definition of the calibration by using context information. \n\n3. The author only focuses on the next token position. It is also possible that the model may not give you the answer in the text token. As discussed in the paper, the PM can be low, meaning all valid tokens have low probabilities. If the model outputs 'of red' rather than 'red', probabilities of the second position should also be considered." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024are,\ntitle={Are Language Model Logits Calibrated?},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x418ZpazsR},\nnote={under review}\n}" }, "abstract": { "value": "Some information is factual (e.g., \"Paris is in France\"), whereas other information is probabilistic (e.g., \"the coin flip will be a [Heads/Tails].\"). We believe that good Language Models (LMs) should understand and reflect this nuance. Our work investigates this by testing if LMs' output probabilities are *calibrated* to their textual contexts. We define model \"calibration\" as the degree to which the output probabilities of candidate tokens are aligned with the relative likelihood that should be inferred from the given context. For example, if the context concerns two equally likely options (e.g., heads or tails for a fair coin), the output probabilities should reflect this. Likewise, context that concerns non-uniformly likely events (e.g., rolling a six with a die) should also be appropriately captured with proportionate output probabilities. We find that even in simple settings the best LMs (1) are poorly calibrated, and (2) have systematic biases (e.g., preferred colors and sensitivities to word orderings). For example, gpt-4o-mini often picks the first of two options presented in the prompt regardless of the options' implied likelihood, whereas Llama-3.1-8B picks the second. Our other consistent finding is mode-collapse: Instruction-tuned models often over-allocate probability mass on a single option. These systematic biases introduce non-intuitive model behavior, making models harder for users to understand." }, "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": [ "language modeling", "calibration", "model understanding" ] }, "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/939f6159ab4ea076f1f53422adcd4a7a4146cd12.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": "Are Language Model Logits Calibrated?" }, "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": "Can the authors present a clear comparison against vanilla SGD (without momentum) and how their results compare against existing results based on Algorithmic stability (for instance, Hardt et al 2016 that the paper cites)?" }, "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 the results appear to be fairly general and novel to my knowledge, though I am not an expert in recent developments in this area." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents high probability convergence rates and generalization bounds for SGD with momentum under a heavy tailed noise assumption for both general non-convex functions and under PL conditions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The empirical section is left somewhat underbaked - it will be worthwhile presenting additional results on more realistic tasks/datasets/neural networks (including transformer style architectures)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. What do you think are the key challenges in obtaining these theoretical guarantees for SGDM? Since there are some existing results as indicated in Table 1 that are worse than the results in the current paper, I'm curious about what are the key technical improvements over their analysis.\n\n2. By considering heavy-tailed gradient noise (large $\\theta$), what additional algorithmic insights can we obtain for the analysis? e.g. how do you compare SGDM with other first-order stochastic optimization methods?\n\n3. I'm not familar with the generalization literature; can you explain why you are considering $T=n/d$ in the general non-convex case and $T=n^2$ in the PL case?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors provide novel convergence results for SCDM, which is a popular algorithm in practice but is much harder to analyze than SGD.\n\n2. All the assumptions that this paper makes are followed by detailed discussions and comparisons to existing literature." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper derives high-probability convergence rates and generalization bounds for stochastic gradient descent with momentum (SGDM), a popular algorithm in practice. The bounds are first established for general non-convex functions, assuming sub-Weibull gradient noise. Then the authors move on to consider functions that satisfy the PL condition and derive improved bounds, with generalization bound independent of the dimension." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although I'm not so familiar with related literature, it seems that there are many existing high-probability and generalization bounds established for various optimization methods. It is unclear what are the key differences (see also 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": 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": "Figures 1 and 2 exclusively present variations of SGDM. Including additional baselines would enhance the comparative analysis, providing insight into the performance of the proposed algorithm relative to established methods. These baselines could incorporate algorithms from previous literature on non-convex optimization." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "For general convex case, the authors prove that the convergence bounds of SGDM are sharper than that of related works and present the first generalization bounds of SGDM. With the additional Polyak-Łojasiewicz condition, the convergence bounds of SGDM achieve a faster $O(1/T)$ rate." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies Stochastic Gradient Descent with momentum (SGDM) and introduces theoretical convergence bounds and generalization bounds for SGDM. These bounds are tighter and faster than the theoretical results of related works in different settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Missing baselines for comparison in numerical experiments. (details in questions)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "- $\\tilde{O} (1/T^{1/2})$ is the same rate as Li & Orabona (2020)? What do you mean by “the convergence bounds are tighter than those of the related work”? (I’m aware that Li & Orabnoa assumes sub-gaussian noise).\n- Assumption 2.8 needs to define quantities like $j_t$. I’m assuming this assumption has to hold for all $t$.\n- Typo: line 292, “over SGD [1]”. What is ref [1]? Seems hard coded." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "I believe SGDM is still used (although not really with decaying step size that this paper focuses on), so the provided theory for heavy-tailed noise can be relevant. The authors seem honest with their results, for instance explicitly stating the the provided results do not improve over standard SGD." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides high-probability convergence and generalization bounds for stochastic gradient descent with (heavyball) momentum (SGDM). Sharper bounds under PL assumption and Bernstein condition are also provided." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The amount of improvement compared to related work seem very marginal. For instance, compared to Li & Orabona, a related work that this manuscript often refers to, the improvement is only from $\\log(T/\\delta)$ to $\\log(1/\\delta)$. I get that Theorem 3.1 also holds for heavy tailed noise ($\\theta > 1$), but in that case the authors admit the bound does not improve over standard SGD.\n- I am not an expert in generalization bound for stochastic gradient methods. Yet, a quick search shows [1], which is present in the references of this paper, but I could not find where it was cited. At least some comparison to [1] should be provided.\n- SGDM with decaying step size is quite outdated at this point, with cosine or more sophisticated sche dules, for instance [2]. As such, I’m not sure how useful these results are.\n- Experiments are shown for convex examples like logistic regression and Huber loss. What is the point of these experiments, when the entire paper is about non-convex bounds?\n- Writing should be vastly improved. Many paragraphs and sentences are dense and long, hurting readability quite a bit: e.g., line 210 to 215 is one sentence, Remark 3.2, entire experimental setup from 466 to 492 is one paragraph… etc).\n\n[1] Ramezani-Kebrya et al. (2024) “On the Generalization of Stochastic Gradient Descent with Momentum”\n\n[2] Defazio et al. (2024) \"The Road Less Scheduled”" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 for my main question.\n\n- Line 1043: could the authors explain why $\\xi_t$ is sub-Weilbull (or alternatively why we can apply Lemma B.4 on $\\xi_t$)?\n\n- eq. 1: miss brackets around $\\exp$" }, "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 systematic analysis of high probability bounds of SGDM under various scenarios.\n- Regarding the technical aspect, this paper comes up with a uniformed analysis for different types of gradient noises, including sub-Gaussian, sub-Exponential, and heavy-tail noises, using the family of sub-Weibull distributions, which captures the degree of \"heavy-tailness\" by the parameter $\\theta$. This is also reflected in all results which naturally implies that heavier tails leads to worse bounds." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies high probability bound, including both convergence bound (corresponding to ERM loss) and generalization bound (corresponding to population loss) of SGDM in the non-convex and smooth regime. Besides the general non-convex case, this paper also analyzes the scenario under PL condition and shows accelerated bounds with this additional assumption. The results work under a relaxed assumption of stochastic gradientsa, where the noises of stochastic gradients are generalized to the family of sub-Weibull distributions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I don't agree with authors' claim that this paper is the first one studying the generalization bound of SGDM and that the two referred papers, Li & Orabona 2020 and Cutkosky & Mehta 2021, provide convergence bound (defined w.r.t. the ERM loss $\\nabla F_S$). I believe both referred papers actually provide generalization bound, e.g., see Li & Orabona Theorem 1 and Cutkosky & Mehta Theorem 2. Both theorems provide upper bounds on the generalization loss $\\nabla F$ instead of the ERM loss $\\nabla F_S$. Also I'd like to point out the difference in terms of data sampling: this paper considers sampling with replacement from a fixed dataset so that there could be repeated data, while the referred papers sample an i.i.d. data from the unknown population distribution in every iteration. This subtle difference is not mentioned in the paper. If my understanding is correct, the results of this paper is less novel than it claims.\n\nAs a related question, does Theorem 3.1 (and similarly Theorem 3.5) also hold for the generalization loss $\\nabla F$? It seems to me that if we slightly modify Assumption 2.4 and 2.8 by replacing $\\nabla F_S$ with $\\nabla F$, then the same analysis in the proof in C.1 still holds: smoothness (eq. 5) is not affected; the martingale difference concentration on page 20 and the sub-Weibull concentration on page 21 still hold with $\\nabla F_S$ replaced by $\\nabla F$. In other words, is it true that the same analysis (under slightly different assumptions) provides a generalization bound?\n\nAs a disclaimer, I'm not an expert in the field of learning theory, and I could be wrong with my understandings. I will be glad to reevaluate the results if the authors point out I'm wrong." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024sharper,\ntitle={Sharper Bounds of Non-Convex Stochastic Gradient Descent with Momentum},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x45vUUY4nT},\nnote={under review}\n}" }, "abstract": { "value": "Stochastic gradient descent with momentum (SGDM) has been widely used in machine learning. However, in non-convex domains, high probability learning bounds for SGDM are scarce. In this paper, we provide high probability convergence bounds and generalization bounds for SGDM. Firstly, we establish these bounds for the gradient norm in the general non-convex case. The derived convergence bounds are tighter than the theoretical results of related work, and to our best knowledge, the derived generalization bounds are the first ones for SGDM. Then, if the Polyak-{\\L}ojasiewicz condition is satisfied, we establish these bounds for the error of the function value, instead of the gradient norm. Moreover, the derived learning bounds have faster rates than the general non-convex case. Finally, we further provide sharper generalization bounds by considering a mild Bernstein condition on the gradient. In the case of low noise, their learning rates can reach $\\widetilde{\\mathcal{O}}(1/n^2)$, where $n$ is the sample size. Overall, we relatively systematically investigate the high probability learning bounds for non-convex SGDM." }, "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", "nonconvex optimization", "stochastic gradient descent" ] }, "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/fb6c886a350cf84c7e046e98a2508262721a5393.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": "Sharper Bounds of Non-Convex Stochastic Gradient Descent with Momentum" }, "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": "* Is it possible to have small discriminative image-text models as the baselines (e.g., CLIP [1] and ALBEF [2])?\n\n[1] Radford, Alec, et al. \"Learning transferable visual models from natural language supervision.\" International conference on machine learning. PMLR, 2021.\n\n[2] Li, Junnan, et al. \"Align before fuse: Vision and language representation learning with momentum distillation.\" Advances in neural information processing systems 34 (2021): 9694-9705." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* **Methodology**: In general, it's a sound approach. By designing the heuristics, the approach enables a rule-based system to **produce** training datasets, **evaluate** the responses, and **post-train** the models using RLHF. Similar insights were also approved in [1].\n\n* **Writing**: The paper is well-written and easy to follow. The experimental designs can support the claims in this paper.\n\n[1] Mu, Tong, et al. \"Rule-Based Rewards for Language Model Safety.\" Advances in Neural Information Processing Systems (2024)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "* **Topic**: Medical VLMs, Rule-based Rewards, RLHF\n* **Summary**: This paper proposed to improve medical VLMs through a rule-based framework. It includes (i) using a rule to generate instruction tuning data, (ii) using the rule to evaluate the response, and (iii) using the rule-based rewards to perform RLHF to reduce hallucination. The experimental results show that the proposed approach achieves better performance than OpenFlamingo, LLaVA, and LLaVA-Med, etc." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* One of my biggest concerns about this paper is that it didn't show the potential of such an approach in medical **generative** VLMs. To be specific, the modeling can be replaced with a **discriminative** classification model since the responses from the MVLM are \"pseudo\" free-text responses, and they can be replaced with text templates + classification labels. So is a **large** generative VLM (7B parameters) really needed to solve the proposed 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": 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": "What does line 136: \"followed by a rigorous evaluation to ensure no hallucinations are introduced within the template set\" mean specifically? What was the criteria used in this evaluation, or what were some examples hallucinations / examples of how hallucinations were checked in the templates." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Strong and sound method with clear performance improvements over many experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a new alignment algorithm for VLMs in clinical settings, addressing hallucination where models generate inaccurate or inconsistent responses. The authors propose a rule-based approach that grounds VLMs in clinical reasoning by creating large-scale visual instruction data and developing a reward function to ensure responses align with medical knowledge across multi-turn conversations. This avoids high costs of RLHF and they used it to create Dr-LLaVA." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Small things:\n- Figure 1 needs to be bigger. The conversation text cannot be read unless I zoom in alot.\n- Typo \"rule\": line 144 our conversational diagnostic system leverages ruke-based representations (Fig. 3) - should be \"rule-based\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "In addition to the weaknesses detailed above, my additional questions are listed below: \n\n1. In Line 136, the authors state that they conduct “a rigorous evaluation to ensure that no hallucinations are introduced within the template set.” Can the authors provide additional details with respect to this evaluation and the numbers/types of hallucinations identified?\n\n2. To the best of my understanding, the single-turn QA performance values reported in Table 1 for question-level accuracy ($A_Q$) and conversation-level accuracy ($A_C$) should be equal, since there is only one question per conversation. Why are the values nearly 20 points apart for Dr-LLaVA?" }, "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 introduces an approach for improving the quality of textual outputs generated by vision-language models in the medical setting. The proposed approach intends to mimic clinical reasoning rules typically employed by physicians.\n2. The proposed approach is capable of understanding/generating multi-turn conversations in the medical setting, unlike many existing medical VLMs. The proposed approach can also handle diverse styles of interactions with clinicians. \n3. The authors demonstrate that their approach leads to performance improvements over several existing methods in this domain and conforms closely with clinical reasoning pathways.\n4. The paper is clearly written and is well organized." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Vision-language models often generate textual outputs that are not grounded in the input image-text information. The authors propose a new alignment algorithm that uses rule-based representations of clinical reasoning (obtained from a physician) to ground VLMs in medical knowledge. These representations are used to (i) generate visual instruction tuning data and (ii) derive a rule-based reward function that evaluates clinical validity of VLM responses. The resulting algorithm eliminates the need for human involvement in data construction or reward model construction. The algorithm is used to develop Dr-LLaVA, a conversational VLM fine-tuned for analyzing bone marrow pathology slides. Dr-LLaVA is shown to outperform several VLM baselines and can handle diverse styles of interactions with clinicians." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Limited Scalability:** The application area explored in this paper is narrow in scope (bone marrow pathology classification), and it seems that in order to expand to other medical applications, a lot of manual effort is necessary. In particular, (1) a physician will have to manually generate a set clinical reasoning rules in order to extend this approach to other medical application areas and (2) the rule-based reward model will need a new set of manually-constructed keywords (as in Table B.5). This weakness limits the widespread usability of this approach. \n\n2. **Additional Methodology Details:** Some important details with respect to the construction of the dataset and the rule-based rewards are missing, as detailed below:\n\n a. **Multi-Turn Conversations:** Each conversation in the synthesized conversational dataset consists of five question-answer pairs (Line 135). Why do all conversations include exactly five QA pairs, when it’s possible that the conversation could end in fewer turns if a leaf node of the decision tree is reached early (e.g. low quality image)? \n\n b. **Dataset details:** No details with respect to the composition of the synthesized dataset are provided. How large is the dataset? What is the class distribution with respect to the diagnoses? Are there sufficient samples for each of the possible decision pathways in the clinical reasoning decision tree?\n\n c. **Keyword Matching:** The authors use a manually pre-defined list of keywords in order to compute rewards and accuracy metrics. How accurate is the process of keyword matching, and how often do false positives / negatives arise?\n\n3. **Need for finer-grained evaluations:** The performance metrics in Table 1 are aggregate metrics. How well does the proposed approach perform in comparison to baselines across each diagnosis and each possible decision pathway?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "The paper is missing an ethics statement, which is essential in such a research involving patient information which is sensitive dataset." }, "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": 1 }, "primary_area": null, "questions": { "value": "1. How many whole slide images did you use to generate 16K datasets?\n2. Do you plan to release the model and the dataset?\n3. What is the source of the clinical reasoning pathways?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Rather than RLHF, the paper uses a hierarchy set of rules based on clinical reasoning pathways to train a medical VLM. This approach is a way around the lack of human annotations. \n2. Not only conventional conversational settings, but also diagnosis related and random order question sequence settings were tested and Dr-LLaVA performs better than the baseline models on all settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper used rule-based rewards generated from clinical diagnostic pathways instead of human feedback based rewards for reinforcement learning to train a medical VLM, Dr-LLaVA. The rule-based rewards include consistency and correctness. 16,340 bone marrow pathology patches and hematopathologist's annotations were used to generate conversations for training. Dr-LLaVA performed better across all settings when compared to baseline models which are mostly further supervised fine-tuned with the same conversations but without the reward function." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. GPT-4 is used to generate the conversations, and a rigorous evaluation is done to ensure no hallucinations are introduced. However, the authors provide no details on the specific evaluation methods used to ensure this, which limits the reproducibility of the work. Additionally, this lack of transparency undermines the objective of reducing hallucinations with reinforcement learning, as the GPT-4-generated dataset could still contain incorrect statements.\n2. The essential details on clinical decision making pathways, dataset generation, and hallucination evaluations are either missing or located in the appendix rather than the main text, making it challenging to fully understand the methodology without the appendix. Since these are important in the work, the absence from the main text disrupts the flow and limits the study’s context and reproducibility. Including these details in the main body, or at least providing a summary with clear references to the appendix for further depth, would make the paper more accessible and cohesive.\n3. The presentation of the work needs improvement, especially in organization and clarity to be publishable quality. Currently, the layout requires readers to frequently go back and forth, as text descriptions are located far from the tables and figures that they describe. This disrupts the flow and makes it challenging to follow the paper. For instance, Figure 3 is introduced in Section 2 before Figure 2, which creates confusion in the narrative structure. Additionally, abbreviations are often used without prior definition; for example, Table 2 and Table 3 feature abbreviated column headings like \"CQ-R\" and \"Hcc,\" but their meanings are not clarified until much later in the text or, in some cases, are only defined in the Appendix. Including these definitions upfront or alongside the tables would enhance readability. Also, typos, such as \"ruke\" instead of \"rule\" on line 144 needs to be corrected.\n4. The paper lacks experiments specifically designed to address the issue of misalignment between image and text and to demonstrate how Dr-LLaVA mitigates this problem, despite this being a primary motivation for using RL over SFT. However, without empirical evidence or ablation study showing that RL successfully reduces misalignment, the paper's contribution is weakened.\n5. The paper lacks a comprehensive qualitative evaluation, presenting only a single qualitative example set in Figure 4. This limited approach does not provide enough insight into the model's practical application and performance across varied cases. Conducting a broader qualitative evaluation, ideally involving human assessments by clinicians, would allow for a more in-depth assessment of the model's reliability and relevance in real-world clinical settings. \n6. The paper lacks a section on limitations, which is essential for providing a balanced view of the work's contributions and areas for future improvement. \n7. The paper is missing an ethics statement, which is essential in such a research involving patient information. An ethics statement provides transparency about how ethical considerations were addressed, including data handling, privacy safeguards, and compliance with relevant regulations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024aligning,\ntitle={Aligning Multimodal Models for Clinical Reasoning using Rule-based Rewards},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x4W8P7ybTE},\nnote={under review}\n}" }, "abstract": { "value": "Vision-Language Models (VLM) can support clinicians by analyzing medical images and engaging in natural language interactions to assist in diagnostic and treatment tasks. However, VLMs often exhibit \"hallucinatory\" behavior, generating textual outputs not grounded in contextual multimodal information. This challenge is particularly pronounced in the medical domain, where we do not only require VLM outputs to be accurate in single interactions but also to be consistent with clinical reasoning and diagnostic pathways throughout multi-turn conversations. For this purpose, we propose a new alignment algorithm that uses rule-based representations of clinical reasoning to ground VLMs in medical knowledge. These representations are utilized to (i) generate visual instruction tuning data at scale, simulating clinician-VLM conversations with demonstrations of clinical reasoning, and (ii) to derive a rule-based reward function that automatically evaluates the clinical validity of VLM responses throughout clinician-VLM interactions. Our algorithm eliminates the need for human involvement in training data generation or reward model construction, reducing costs compared to standard reinforcement learning with human feedback (RLHF). We apply our alignment algorithm to develop Dr-LLaVA, a conversational VLM finetuned for analyzing bone marrow pathology slides, demonstrating strong performance in single and multi-turn medical conversations." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Medical 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/d431d8d0f0285b86d674c0c695dc1c890aa7d8b1.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/3dfaa354b304aa771a23354e3accfcc47ae6dc8e.zip" }, "title": { "value": "Aligning Multimodal Models for Clinical Reasoning using Rule-based Rewards" }, "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": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The parameter $c$ mentioned in line 96-97 (referred to as the field variables or channels) seems a bit ambiguous here as the following PDE doesn't contain anything about $c$. Would it be possible for the authors to provide more detailed explanation for the field variable/channel $c$ here? (This also highly relates to the parameter $N_c$ appearing in equations (3) and (5).)\n\nReferences:\n\n[1] Bruna, Joan, Benjamin Peherstorfer, and Eric Vanden-Eijnden. \"Neural Galerkin schemes with active learning for high-dimensional evolution equations.\" Journal of Computational Physics 496 (2024): 112588.\n\n[2] Gajjar, Aarshvi, Chinmay Hegde, and Christopher P. Musco. \"Provable active learning of neural networks for parametric PDEs.\" In The Symbiosis of Deep Learning and Differential Equations II. 2022.\n\n[3] Gao, Wenhan, and Chunmei Wang. \"Active learning based sampling for high-dimensional nonlinear partial differential equations.\" Journal of Computational Physics 475 (2023): 111848." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Extensive numerical experiments on multiple PDEs are provided to validate the effectiveness of the proposed methodology. Also, details about the numerical experiments, such as the neural network models and training procedures, are included for the sake of completeness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provided a bechmark called AL4PDE, which unifies active learning (AL) with neural PDE solvers. Specifically, it studies how several state-of-the-art neural surrogate model may be applied to solve parametric PDEs under a solver-in-the-loop (AL) setting. A complete set of numerical experiments on various tasks is included to justify the effectiveness of AL based methods compared to methods based on random sampling." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Though the authors have conducted a literature review on how AL has been used for solving other problems from scientific ML, such as PINN and direct prediction, it seems to the reviewer that the authors have missed a few important references like [1,2]. It might be meaningful for the authors to include these work and briefly discuss them in the introduction. \n\n2. Given that this work aims for a complete benchmark on various tasks, the authors might consider including some more experiments on high-dimensional PDEs, just like the setting of [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": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the strengths and weaknesses." }, "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-presented and easy to follow.\n\nThe proposed framework is novel in extending neural PDE methods with active learning methods.\n\nThe benchmark includes various batch selection strategies and neural PDE solvers, covering recent and classical works." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a benchmark framework for neural PDE solvers under active learning (AL) settings (AL4PDE). It provides a modular benchmark with various parametric PDEs and AL methods. The experimental results show that AL significantly reduces average and worst case errors compared to random sampling and yields reusable datasets across experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One key benefit of AL is data efficiency, which is also stressed in the paper. It is important to show how much data reduction can be achieved with reasonable model performance.\nCurrent experiment section only shows performance comparison of different active learning methods and lacks the \"offline\" performance, which is training the model with full dataset and evaluate its performance. \n\nAt line $88$, the author claims \"We demonstrate that using AL can result in more accurate surrogate models trained in less time.\" As mentioned above, this claim is not supported with empirical evidence as the experimental section only compares active learning performance, which cannot demonstrate improvement in accuracies regarding offline performance.\n\nThe novelty of this framework seems limited, as it is a combination of existing AL and neural PDE methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Could you please add baselines form DoE?\n\nCould you link your benchmark code to existing open source code for uncertainty quantification (e.g. UQ 360)?\n\nCould you explain the tradeoffs of reducing the dimensionality of features vs. implementing translational invariance in terms of data efficiency of the feature-based AL?\n\nCould you add examples that do not use periodic boundary conditions?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Contributing a benchmark in active learning for PDE solvers fills a needed gap in computational infrastructure for PDE solvers that is key to the central challenge of data efficiency.\nThe article is pedagogical and presents clearly the capability of the benchmark." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Thie article introduces an active learning benchmark for neural PDE solver. It compare exploration-exploitation tradeoffs based uncertainty (epsitemic uncertainty of an ensemble of models with top-K and SBAL) or features (using dimensionality reduction using Gaussian sketching with Core-Set and LCMD). The authors then show a benchmark of these method on 1D and 2D parametric PDEs adding the baseline of sampling uniformly at random to represent the lack of active learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The authors should help compare methods of Bayesian active learning and those of the field of design of experiments (DoE), which is missing in the literature review. For example, instead of using a baseline of uniform sample, Latin Hypercube sampling should be provided, as well as more sophisticated DoE methods. This benchmark effort is an opportunity to bridge these areas of research and communities that try to solve the same problem with a slightly different point of view and a different approach.\n\nThe benchmark should broaden the UQ methods by connecting to existing efforts (for example, the open-source UQ 360). Any UQ method that provide a confidence interval should suffice for active learning as the spread of the confidence interval can be a proxy of the uncertainty.\n\nIn the implementation details, the tradeoffs of the choice of taking the spatial average over the features to make make feature-based AL translation invariant are not discussed. It seems that the averaging creates a significant dimensionality reduction that may outweigh the benefits of a translational invariance in terms of data efficiency.\n\nAll the implementations use periodic boundary conditions, which significantly limits the scope of applications." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "No question" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The well-designed modular framework provides a solid foundation for further research on active learning in the context of PDEs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a modular active learning (AL) framework for training surrogate models of partial differential equation (PDE) solvers. It introduces a numerical solver for generating PDE samples, several surrogate models, batch selection strategies, and acquisition functions designed for active learning within this framework." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the framework is tailored for PDE problems, the implemented acquisition functions are orthogonal to PDE problems i.e. they are AL methods that are used in general domains. As a framework of AL for PDE, at least some PDE-specific AL methods such as adaptive sampling [1], also mentioned in Related Work, should be also implemented.\n1. The paper’s scope in terms of the surrogate models, acquisition functions, and types of PDEs studied is quite limited, impacting its practical applicability.\n\n[1] W. Gao and C. Wang, Active Learning Based Sampling For High-dimensional Nonlinear Partial Differential Equations, Journal of Computational Physics, Vol. 475, 2023." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A extensible benchmark to evaluate pool-based active learning for neural PDE solvers." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024active,\ntitle={Active Learning for Neural {PDE} Solvers},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x4ZmQaumRg},\nnote={under review}\n}" }, "abstract": { "value": "Solving partial differential equations (PDEs) is a fundamental problem in engineering and science. While neural PDE solvers can be more efficient than established numerical solvers, they often require large amounts of training data that is costly to obtain. Active Learning (AL) could help surrogate models reach the same accuracy with smaller training sets by querying classical solvers with more informative initial conditions and PDE parameters. While AL is more common in other domains, it has yet to be studied extensively for neural PDE solvers. To bridge this gap, we introduce AL4PDE, a modular and extensible active learning benchmark. It provides multiple parametric PDEs and state-of-the-art surrogate models for the solver-in-the-loop setting, enabling the evaluation of existing and the development of new AL methods for PDE solving. We use the benchmark to evaluate batch active learning algorithms such as uncertainty- and feature-based methods. We show that AL reduces the average error by up to 71\\% compared to random sampling and significantly reduces worst-case errors. Moreover, AL generates similar datasets across repeated runs, with consistent distributions over the PDE parameters and initial conditions. The acquired datasets are reusable, providing benefits for surrogate models not involved in the data generation." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Active Learning", "Neural PDE Solvers", "Scientific Machine Learning", "Benchmark", "Framework", "Neural Operators" ] }, "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/3ba9c3f08a15e3c39295bfb2bd7993fd6e20419b.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/ec12a9b33858131c5e3251c9629a610ba2f08e58.pdf" }, "title": { "value": "Active Learning for Neural PDE Solvers" }, "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": "- What are the theoretical guarantees for recovering the means through the ``largest k maxima of $\\mathcal{J}$\" in the MUSIC algorithm? \n- What is L in algorithm 1 and section 2.2?\n- From eq. 17, each $f_d$ can in general be $\\Theta(1)$. As per the definition of f in eq. 13, doesn't this imply an exponential dependence on the dimension in the sample complexity given by eq. 15?\n- In Theorem 2, what is the requirement on n?\n- Why can't the algorithm in Vempala & Wang (2002) be further used to directly estimate the means of the mixtures? How does this compare to the MUSIC algorithm?\n\n-- Typos:\n- Line 388: expectated -> expected" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper does an adequate job at describing the problem setup and the proposed algorithm.\n- The proposed approach could have utilities in applications with known low-frequency bias in the fourier domain.\n- Numerical results suggest improved computational efficiency over the EM baseline with similar performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an algorithm for estimating the means, the order (number of components), and the mixing distribution (the distribution over components) of a Gaussian mixture model through the measurements in the fourier space of the empirical measure of independent samples. Assuming the knowledge of the covariance, the work provides lower-bounds on the sample-complexity for estimating the order and mixing distribution. The paper further suggests applying PCA to improve the algorithm's computational complexity by estimating the low-dimensional subspace spanned by the means. Lastly, the paper numerically evaluates the proposed approach against EM (Expectation Maximization) baselines and criteria for estimation of the model order." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The theoretical analysis in the work is limited. The paper only provides lower bounds on the sample-complexity for estimating the model order and mixing distribution without a discussion of the sample-complexity for recovering the means themselves. The paper doesn't describe the conditions under which the mixture means can be estimated through the local minima of the MUSIC imaging function apart from lower-bounds on the necessary discretization. The PCA-based result is also directly adapted from Vempala & Wang 2002 and is not combined with the analysis for the main algorithm. Ideally, the sample-complexity guarantees should be compared with information-theoretic/minimax-optimal lower-bounds at a given scaling of separation distance and the covariance.\n- The proposed algorithm is an extension of Liu and Hai Zhang 2024 to multi-dimensional random variables and therefore possesses limited novelty. \n- It's unclear what assumptions on the fourier decomposition of the covariance are assumed for the proposed algorithm to be sample-complexity efficient. Even for the sample-complexity lower bounds for estimating the order, the dependence on the fourier decomposition of the covariance is hidden in the quantity $f$. It should be clarified how $f$ scales under a scaling of $\\Delta$ and $\\Sigma$ which allows recovery." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Please see above" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper is well written. Both the problem and the techniques used in the paper are novel and relevant." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors tackle the problem of parameter estimation from a GMM with iid samples when even the number of components is unknown. To do so, the authors exploit Fourier measurements of the samples - a novel technique in itself and proposes an algorithm with linear time complexity. The authors further show the use of PCA in dimension reduction for improved computational guarantees." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have some questions regarding this work\n\n1. I am worried that instead of the number of components being a hyperparameter, we now have the cutoff frequencies as a hyperparameter to the algorithm. Why is the latter better than the former? What is L in Algorithm 1?\n\n2. It is critical that the MUSIC algorithm is inserted into the main text. Otherwise the reader who is not acquainted with it is not able to understand it at all. \n\n3. The algorithm is very poorly written. The function \\Psi_n(t) is not referred to appropriately. In Step 2, what is \\mu? What is the quadratic program written in Step 4? The entire algorithm is a black box\n\n4. There is no intuition provided for the MUSIC algorithm. \n\n5. The lower bound in Remark 324-330 talks about the sample complexity when components is known - is that a contribution of this work? In any case, for K=5 and Delta=1/2, the dependence in lower bound is 2^{18}. On the other hand, the upper bound in Theorem 1 has dependence of 2^{16}. Why is the upper bound smaller than the lower bound? What am I missing?\n\n5) The fact that PCA reduced \"time\" complexity needs to highlighted appropriately - confusion with sample complexity in several places" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 look at 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 paper addresses the important problem of Gaussian mixture model estimation, with particular emphasis on determining the unknown number of mixture components. The authors propose a Fourier series-based method designed to simultaneously extract both the number of components $K$ and the model parameters $\\mu_i's$ and $w_i's$. Through extensive simulations, the authors demonstrate that their proposed algorithm performs comparably to or better than the traditional EM algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes two algorithms for estimating parameters in Multivariate Gaussian Mixture Models with a known common covariance matrix. The first algorithm leverages the MUSIC (Multiple Signal Classification) algorithm, along with a quadratic program, to simultaneously determine the model order $K$, estimate mean parameters, and estimate the mixing weights $w_i$. For small problem dimensions $d$, the authors propose gridding a bounded region in $R^d$ using $N^d$ data points to approximate an appropriate Fourier function, which is then used to determine the model order and parameters. For high-dimensional problems, the authors employ a Principal Component Analysis-based approach to first perform dimensionality reduction, and then apply the gridding technique (the first algorithm) to estimate the parameters." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper, while well motivated, lacks concrete theoretical guarantees. For instance: \n\n1. It is not clear how accurately the number of components $K$ of the model can be estimated by the MUSIC algorithm (even in small dimensions in presence of noise). More concretely, is the estimate (say K-hat) for the number of components a consistent estimator for the true number of components $K$? What happens to this estimate when the dimension of the problem is large.\n\n\n\n2. Are the estimates of weights and mean vectors are consistent? The authors suggest to run their algorithm with a larger number of components when the true number of components are unknown (page 14 lines 752-755). It is not clear what happens in that scenario. To best of my under standing, one selling point of the paper (compared to K-means or the EM algorithm) was that it detects the number of components accurately. It is not clear from the presentation of the paper, how justified is this claim. \n\n\n\n3. Finally, the paper recommends to grid a bounded region in the space $R^d$ using $N^d$ data points, and hence it is not efficient even when $d$ is small. On the contrary, the competing algorithms like K-means and EM-algorithm do not face such issues. \n\n\n\n4. The paper seems to be an extension of the MUSIC algorithm for one dimension [1]. It would be nice to know what are some technical challenges that are novel in this setting. \n\n[1]. Xinyu Liu and Hai Zhang. A fourier approach to the parameter estimation problem for one-\ndimensional gaussian mixture models. arXiv preprint arXiv:2404.12613, 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": 1 }, "primary_area": null, "questions": { "value": "- Please see the weakness question above for questions regarding definitions and nomenclature?\n\n- Line 210: I believe the last term should have \\sigma_max(\\Sigma), rather than \\sigma_min(\\Sigma). Note that t^{T} \\Sigma T > ||t|^2 \\sigma_{\\min}(\\Sigma) and hence 1/(t^{T} \\Sigma T) < 1 /(||t|^2 \\sigma_{\\min}(\\Sigma)) and hence -1/(t^{T} \\Sigma T) > -1 /(||t|^2 \\sigma_{\\min}(\\Sigma)). This changes the result in line 214, so the dependence on n also changes to sigma_max. Would it change any other results in the paper? For example, would it change Equation 15?\n\nLine 318: In theorem 1, it is stated that \\Delta >= R_{D, K} holds with probability at least 1 - delta. From Definition 2, \\Delta seems to be a property of the underlying distribution and does not have anything to do with samples. So I am not sure what this theorem means." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper proposes a new algorithm for GMMs and characterizes its complexity. The results are novel and would be of interest to researchers in the field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper considers the problem of estimating the means and the number of components of the GMM from samples. It proposes a new approach based on estimating the characteristic function (based on Fourier transform) and characterize the number of samples and time required by the algorithm to estimate the means and the order of the GMM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My biggest concern about the paper is that the current writing assumes the reader is familiar with many technical terms which are not commonly used in machine learning or TCS conferences, for example terms such as 'spectral estimation algorithm', 'stable recovery', 'cutoff frequency', which are stated and used without being formally defined. While researchers in statistics may be familiar with these terms, the paper is difficult to follow for ML theory researchers. The paper also has some typos or mistakes (see the questions section), which makes it hard to follow. Here are few suggestions which might improve the flow (in no specific order):\n\nSome notations are used before they are defined, e.g.,\n- Line 136: D is not defined at this point.\n- Line 143: w_{min} and \\Delta are not defined yet.\nLine 142, 144: what is \"stable recovery\" and \"Stable estimation\"? Does it mean with high probability or something more subtle? It might be good to formally define it.\nLine 233, 249: what is L?\nLine 250: what is \\tilde{K}?\nLine 377: what is M?\n\nSome parts can use a reference\n- Line 192: why is the variance of asymptotic normality ( 1- |\\phi(t)|^2)? Please add a citation.\n- Line 246: why does Nyquist-Shannon sampling theorem imply that to estimate the mean, one needs h < \\pi/R?\n\nSome terms may not be easily understood by ICLR researchers\n- Line 194: why is multiplying by e^{t^T \\Sigma t} considered as modulating? While this is not critical to the story, there are many terms used like this, which may not be familiar with readers.\n\nI believe due to space constraints, the MUSIC algorithm is provided in the Appendix, while this is fine, it might be good to have some explanation of the algorithm in the main paper.\n\nLine 236: Algorithm 1: what is the difference between w and \\pi? Is there a reason to use \\hat{w}_i to denote an estimate of \\pi_i?\n\nLine 285: what is \\epsilon(t)? in Equation 13? Since this is a fixed quantity, by display equation in line 197, isn't y(t) = modulated characteristic function when epsilon(t) = 0? Also what does the assumption ||epsilon(t) ||_\\infty \\leq sigma mean?\n\nRemarks in line 320 and 324: The abstract currently suggests the provided bounds are the min-max rates for order estimation. However, after reading these two remarks, I realized that these are rates for a specific algorithm. I believe this confusion is not intentional, but would be good to clarify." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Do we know if the proposed algorithm can achieve the lower bound?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper provides sample complexity lower bounds." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides a lower bound on sample complexity for learning mixture of Gaussians.\n\nIt proposes a PCA based approach to solve learning mixture of Gaussians, which is comparable to EM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I couldn't find theoretical guarantees for the proposed algorithm." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "In the 'Primary Area' section, the authors\nchose to classify their paper in \n'semi-supervised, and supervised representation learning'. For the N iid input samples, \nis there any vector of known classes (i.e. ground truth)\nused in their estimation algorithm?\n\nCan the authors provide some applications \nof their algorithm in these two areas?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper studies theoretically the popular \nproblem of unsupervised learning of estimating the GMM \ndistribution of N input samples.\nIt states its goal clearly: it is about \nthe usage of Fourier operations, which lower the \ncomplexity of the learning algorithm.\nIt uses concise language, it presents its model and \nvalidates the complexity efficiency versus some variants\nthe EM algorithm.\nIt defers the detailed \ndiscussion of MUSIC for the appendix. \nThe above points make it well written\nand clearly presented.\n\nFrom a first study, the model seems mathematically \nsound and the suggested Fourier operations\nform a learning algorithm which seems to be novel.\n\nThe assumptions of the model are restrictive: the\nauthors acknowledge that in Lines 530 - 538." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors study the problem of using iid\nsamples to estimate the D-dimensional Gaussian Mixture \nModels (GMM)\nparameters using an algorithm with linear complexity,\nwhich challenges the current baselines (EM algorithm)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper chooses to analyze the complexity of the \nmodel and avoids allocating a discussion \nin practical applications of GMM estimation. That \nwould make it stronger as a machine learning study." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning High-dimensional Gaussian Mixture Models via a Fourier Approach},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x4jPW4p55i},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we address the challenge of learning high-dimensional Gaussian mixture models (GMMs), with a specific focus on estimating both the model order and the mixing distribution from i.i.d. samples. We propose a novel algorithm that achieves linear complexity relative to the sample size $n$, significantly improving computational efficiency. Unlike traditional methods, such as the method of moments or maximum likelihood estimation, our algorithm leverages Fourier measurements from the samples, facilitating simultaneous estimation of both the model order and the mixing distribution. The difficulty of the learning problem can be quantified by the minimum separation distance $\\Delta$ and minimal mixing weight $w_{\\min}$. For stable estimation, a sample size of $\\Omega\\left(\\frac{1}{w_{\\min}^2 \\Delta^{4K-4}}\\right)$ is required for the model order, while $\\Omega\\left(\\frac{1}{w_{\\min}^2 \\Delta^{4K-2}}\\right)$ is necessary for the mixing distribution. This highlights the distinct sample complexities for the two tasks. For $D$-dimensional mixture models, we propose a PCA-based approach to reduce the dimension, reducing the algorithm’s complexity to $O(nD^2)$, with potential further reductions through random projections. Numerical experiments demonstrate the efficiency and accuracy compared with the EM algorithm. In particular, we observe a clear phase transition in determining the model order, as our method outperforms traditional information criteria. Additionally, our framework is flexible and can be extended to learning mixtures of other distributions, such as Cauchy or exponential distributions." }, "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": [ "Gaussian Mixture Models(GMM)", "Parameter Estimation", "Model Order Selection", "Super-resolution", "Line Spectral Estimation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/e37fdaf563524d045779b67e0b3257ddd364e4c2.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/45691c41e4bf5806b152a1864f0ed0dc9af7fb33.zip" }, "title": { "value": "Learning High-dimensional Gaussian Mixture Models via a Fourier 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": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "The paper tackles the problem of single object classification in an image. This is a VERY populated and saturated field, with techniques improving two tenths of a point with niche tricks being published. More specifically, the paper tackles efficient computing for edge devices. Again, this is a field that started years ago with dozens of publications and even products already running on edge devices around the world.\nAll this work is completely ignore by the paper. Instead, the paper suggests a naive and hand tuned reduction approach, and compares to a vanila resnet-50. Even in this dated comparison, there is no clear merit for the proposed approach. \nThere are still many questions left unanswered unfortunately: \n* How would self-learned features behave for the same computational cost? \n* What other alternatives for compression are there? \n* How does one tackle more complex scenes? \n* How does the method fair against shuffle net, mobile net and their many followups?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is concise, and explains the idea in simple words\nThere might be some interesting insight in using a very concise representation" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method for efficient image classification. The main insight is distilling images to two slim representations - either a contour or dominant points of this contour - and use it for classification, inducing minimal computational costs both during inference and training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Unmotivated problem\nNon noval solution\nLimited applicability\nPoor results\nInadequate evaluation" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "+ Consider updating the paper with the recent state-of-the-art models (refer to https://paperswithcode.com/sota/image-classification-on-fashion-mnist) results on fashionmnist dataset.\n+ Please consider using larger-sized datasets for quantitative results." }, "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 achieved \n1) multiple outcomes (listed in the summary), making it unique.\n2) results comparable to SOTA methods on multiple datasets (FashionMNIST, Flavia, and Folio).\n+ The proposed method achieves impressive inference time (both on the server and Jetson Orin configurations, as shown in Figure 4) and total time on both devices (workstation and edge computing, as shown in Table 2).\n+ The paper is well-written and easy to understand/follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an approach to leverage efficient polygonal representations of input images. Polygonal representation offers a concise and flexible depiction of images. The proposed method transforms input images into polygonal forms using either dominant points or coordinates of contours. The polygonal representation 1) substantially reduces the computational burden associated with processing large image datasets, 2) accelerates the training process, 3) conserves computational resources, 4) facilitates improved generalization of the trained models, and 5) is suitable for real-time applications and resource-constrained environments. The polygon forms are used to train deep neural networks. The proposed method achieves comparable performance to SOTA methods, mitigate overfitting, and produce lightweight models suitable for edge computing." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper is limited by qualitative results (restricted to one or two samples). It is unclear if there are complex cases and how the model performs.\n- Experiments are limited to small-sized datasets.\n- Resnet-50 is a slightly old method and is the only method used to compare with PolygoNet (Contours and DP, as shown in Table 1). It appears more like a technical report than a comprehensive comparison with recent state-of-the-art methods in the problem domain.\n- The margin of improvement in the results due to the proposed method is negligible (Table 1) on all three datasets. Though F1 Scores are comparable, the accuracy values are lower.\n- The proposed method is limited to extracting geometric/shape features from 2D images while ignoring color or other 2D feature details.\n- The paper does not provide an ablation study to understand the contributions of individual components within the proposed method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The proposed architecture uses self-attention, which can potentially increase computation times over simpler architectures. Did the authors experiment with simpler architectures? If so, the results could be interesting to include.\nDid the authors experiment with PointNet-like architectures, since the input is composed of contour/dominant points?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The paper is well written. In particular the first 5 pages are very well written, clear and easy to follow, including also an overview of the MATC algorithm\n- The idea is definitely interesting: using only contours (or, more generally, keypoints) to do classification can lower the computational requirements, which is important for edge devices, and it is also going in the direction of human-inspired techniques compared to processing every input pixel using CNNs or ViTs\n- The proposed algorithm is effective in significantly reducing the runtime and number of operations compared to the ResNet baseline" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a pipeline to classify input images with lower computational requirements compared to full-scale CNNs, potentially suitable for edge devices. The pipeline consists in running classical contour-extraction algorithms, optionally applying the MATC approach to extract dominant points from the contours, and then feeding the extracted points (contour or dominant) to an ad-hoc classification network architecture composed of self-attention with positional encoding and 1D convolutions, to leverage global and local context. Experiments are carried out on FashionMNIST, Flavia and Folio datasets, which include objects on regular backgrounds, comparing the results of the proposed method to a ResNet baseline. The proposed method uses significantly less FLOPs compared to the baseline, which translates to significantly lower inference times. The approach using dominant points uses marginally less operations compared to using the full contours, but it requires significantly longer overall time, albeit still less than the baseline, due to the MATC algorithm. Both proposed approaches achieve F1 scores comparable with the baselines, but reduced accuracy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Even though the paper is generally well written, the first half of the paper includes repetitions when describing the proposed pipeline. The network architecture description at page 6, instead, could benefit from a more descriptive figure.\n- A good part of the paper is used to describe the MATC-based approach, only for it to be shown as slightly worse than the contour-based approach and with significantly higher runtime. Can the authors elaborate on the importance of this variant of the proposed pipeline? For example if they see use cases in which it can be preferred over the contour-based variant. Otherwise, the paper could also be reformatted to more effectively target the question \"how much information is needed for classification\", to which the answer could be that contours already provide a good amount of information, but dominant points carry less information and require more execution time. If targeting this question, additional experiments would be required (such as including internal contours/edges, colours, etc).\n- The paper claims \"enabling our pipeline to demonstrate consistent performance across diverse and challenging conditions\", however the experiments are carried out on datasets with very simple conditions, such as regular backgrounds, which can significantly simplify the task of contour extraction, key to the proposed pipeline. Either more challenging datasets should be employed, or the claim should be relaxed, and also clarified in the introduction that the pipeline is only suitable to images which can be easily binarized.\n- The experiments only include one baseline: ResNet. A milestone of vision architectures, ResNet is certainly important, but a bit outdated compared to more recent architectures. ViTs should be included (and they usually require more computation, as the authors also note, which could benefit the comparison to the proposed approach), as well as other architectures targeted to low-compute devices, such as YOLO. Without these baselines, it is difficult to assess the validity of the method, even in controlled conditions such as the ones offered by the chosen datasets." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "The proposed method is limited (see 2,3,4,5 above) and underestimates the task and system (see 1, 2, 3 above). Please clarify if I have a misunderstanding in the weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1 Fast inference speed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work proposes a contour and dominant point-based image classification model. By only considering the contour and dominant point, the work can speed up the training and inference speed of the neural network." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1 In lines 68-72, the work mentions, 'This methodology aligns with cognitive processes observed in human visual perception, where recognition is often based on key structural features rather than exhaustive pixel-by-pixel analysis Biederman (1987); Koffka (2013).' This motivation/insight oversimplifies the image recognition task and human perception. Human perception and cognition systems are much more complex than this. Biederman (1987)'s work addresses the basic level of the cognition process. Human cognition operates on multiple levels and involves lateralization, where both hemispheres of the brain contribute to different aspects of recognition—such as broad categorization and fine-grained identification. This multiple-level, lateralized system integrates structural, contextual, and surface information, which goes far beyond the simplified structural approach suggested in the work. I encourage the authors to take a look at this literature review [2].\n\nBy the way, there is an interesting game called 'Who's That Pokémon?' In this game, the player needs to guess the name of the Pokémon based on its contour, which is more challenging than using pixel-by-pixel information.\n\n2 In lines 23-25 and lines 65-67, the work mentions that the proposed method can filter out the background noise. This is an interesting direction, as spurious correlations [1] often arise from the background. However, the proposed method relies on a thresholding mechanism to filter out irrelevant content. I do not believe this approach is robust, especially in scenes with diverse backgrounds. The qualitative results are all based on datasets with simple backgrounds. Using the segmentation model is one way as mentioned in the conclusion, but the work does not consider integrating the model, limiting the work.\n\n3 Only considering contour and dominant points is risky and underestimates the complexity of image recognition/classification tasks. Many image classification tasks depend on more than just contour recognition. For example, how can the proposed method distinguish different kinds of balls, such as basketballs, soccer balls, ping-pong, volleyballs, and baseballs, without knowing the texture, color patterns, or surface details? How can the proposed method distinguish the brand of vehicles? The Flavia dataset used to demonstrate recognition is favorable to the approach, as the shape and contour of the leaves are different across categories. If the proposed method wants to prove the classification ability, considering the examples I mentioned is more convincing.\n\n4 The work emphasizes inference speed and tests it on edge devices, which is a highly relevant real-world consideration. However, the work does not provide a real-world demo; instead, it is tested on an existing dataset offline. Deploying this in a real-world scenario is a challenge that is not addressed. For instance, real-world scenes often have complex backgrounds, include multiple objects, and feature objects in various poses. None of these real-world considerations are taken into account in the work.\n\n5 The work may not work well for non-rigid objects and objects with various poses. Although the datasets contain non-rigid objects such as clothes, these objects are well-posed in the dataset.\n\n6 The baseline is just ResNet50. Only adopting one baseline is not convincing.\n\n[1] Kim, Younghyun, et al. \"Bias-to-text: Debiasing unknown visual biases through language interpretation.\" arXiv preprint arXiv:2301.11104 (2023).\n\n[2] Palmeri, Thomas J., and Isabel Gauthier. \"Visual object understanding.\" Nature Reviews Neuroscience 5.4 (2004): 291-303." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024polygonet,\ntitle={PolygoNet: Leveraging Simplified Polygonal Representation for Effective Shape Classification},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x4lmFlfFKX},\nnote={under review}\n}" }, "abstract": { "value": "Deep learning models have achieved significant success in various image-related tasks. However, they often encounter challenges related to computational complexity and overfitting. In this paper, we propose an approach that leverages efficient polygonal representations of input images by utilizing either dominant points or coordinates of contours. Our method transforms input images into polygonal forms using one of these techniques, which are then employed to train deep neural networks. This representation offers a concise and flexible depiction of images. By converting images into either dominant points or contour coordinates, we substantially reduce the computational burden associated with processing large image datasets. This reduction not only accelerates the training process but also conserves computational resources, rendering our approach suitable for real-time applications and resource-constrained environments. Additionally, these representations facilitate improved generalization of the trained models. Both dominant points and contour coordinates inherently capture essential features of the input images while filtering out noise and irrelevant details, providing an inherent regularization effect that mitigates overfitting. Our approach results in lightweight models that can be efficiently deployed on edge devices, making it highly applicable for scenarios with limited computational resources. Despite the reduced complexity, our method achieve performance comparable to state-of-the-art methods that use full images as input. We validate our approach through extensive experiments on benchmark datasets, demonstrating its effectiveness in reducing computation, preventing overfitting, and enabling deployment on edge computing platforms. Overall, this work presents a methodology in image processing that leverages polygonal representations through either dominant points or contour coordinates to streamline computations, mitigate overfitting, and produce lightweight models suitable for edge computing. These findings indicate that this approach holds significant potential for advancing the field of deep learning by enabling efficient, accurate, and scalable solutions in real-world applications. The code for the experiments of the paper are provided at \\url{https://anonymous.4open.science/r/PolygoNet-7374}" }, "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": [ "Shape Classification; Polygonal representation; Computational Efficiency; Self-Attention Mechanism;" ] }, "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/818aba7578003bf60baa64a11f1b477254548ead.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": "PolygoNet: Leveraging Simplified Polygonal Representation for Effective Shape Classification" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the weakness section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper is very well-written and easy to follow.\n- The proposed method of conducting LLM-based learning on graphs without a GNN component is novel and makes sense.\n- The proposed hash-based structural similarity calculation is novel to me." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates an interesting problem of leveraging LLM for learning on text-attributed graphs. The authors propose a method called SKETCH which adapts LLM for graphs by retrieving both structural and semantic information. To be specific, the semantic-based retrieval is built upon some off-the-shelf pretrained retrievers, and the similarity score is calculated by the embedding similarity search. On the other hand, the structure-based retrieval is designed to fetch related neighbors from the graph with a novel hash-based Jaccard similarity estimation. The semantic similarity score and structural similarity score are merged to select the final neighbors, which are put into the LLM together with the center node for problem-solving. The authors then conduct experiments on three real-world datasets to demonstrate the effectiveness of their proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Some model designs are not well-illustrated. For example, how the sampled neighbors and center text is finally fed into the LLM? What kind of instruction or prompt are you using? Do you train the model or just use do direct prompting?\n\n- Some experiments on larger datasets or other tasks other than node classification can be helpful. The experiments are mainly focused on 10k-size graphs. Can the method be scaled to a large graph with millions of nodes? Node classification might not be enough to demonstrate the strength of the proposed method. It would be interesting to try on some more advanced LLM-based graph reasoning benchmarks [1].\n\n- How many neighbors are finally selected? Is the model performance sensitive to the number of selected neighbors? Is there any scalability issue?\n\n- Typos: (1) “where |S| is the size of the text-attributed nodes” should it be |V|? (2) The equation in line 224 needs one further “=” to be complete.\n\n\n\n[1] Jin, B. Graph chain-of-thought: Augmenting large language models by reasoning on graphs. ACL 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. Is the method reproducible? Please provide your repository if possible. \n2. Is the GCN-generated embedding also leveraged in the method? Has the author carefully considered the differences between the proposed method and GCN, or should these differences be discussed further?\n3. The title is **LARGE LANGUAGE MODELS BASED GRAPH CONVOLUTION FOR TEXT-ATTRIBUTED NETWORKS**. Does this imply that feature aggregation based on a weighting mechanism quantified by semantic and structural proximity is a better way to present it?\n4. \"While effective, this method may miss the rich semantic nuances in textual data.\" This is probably a crucial starting point. Are there any references or experiments to support this claim?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Due to the missing semantic information during the message-passing process, the author has proposed a fused framework based on LM and graph heuristics, which is easily scalable.\n2. The author has conducted extensive experiments to demonstrate the performance improvement and computational efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The author has introduced a new method that combines a pre-trained language model (PLM) with a graph heuristic (Common Neighbor) for semi-supervised node classification. In this approach, the PLM generates semantic proximity by incorporating a weighted sum of token-level embeddings. This semantic information is then fused with local graph structure, such as the common neighbor heuristic, by weighting the local connections according to their semantic proximity." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. When introducing the background of GCN and RAG, some crucial papers are not cited. For instance, the structure of text-attributed graphs encompasses textual information from various nodes. Inspired by concepts from Retrieval-Augmented Generation (RAG) \\cite{one in NLP}{one in graph}{one in tag}, we propose integrating an additional corpus during the training process.\n \n2. The paper requires additional revision for better and more fluent logical flow. For example, GNNs primarily depend on node-level aggregation via graph convolutions to compute weighted sums of neighboring features. While effective, this method may overlook the rich semantic nuances in textual data. Incorporating such nuances enables a more granular understanding of the relationships between tokens, leading to improved flexibility and adaptability.\n\n3. The method is not well presented. In the section on aggregated learning of retrieved content, a weighted sum of semantic and structural proximity is introduced, but the difference from graph convolution is not carefully studied or justified.\n\n4. Empirical Demonstration: The results are reported without running on 5-10 random seeds or multiple data splits.\n\n5. It is not clear how to calculate semantic proximity and structural proximity for a node classification 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": 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. The foundational work of Retrieval-Augmented Generation (RAG) [1] is not cited. Given that the primary contribution of this paper lies in graph retrieval-augmented generation, it is crucial for the authors to provide a comprehensive discussion of significant prior works [2-4] in related fields.\n\n2. Only texts in documents (nodes) are used, and connections (graph structure) between texts are not considered in the generation phase. The authors presents several drawbacks in TAG modeling, such as \"the text representations and graph structure are trained independently from their respective aspects, potentially resulting in sub-optimal integration between the two modalities\" and \"the separate processing stages do not take into account the simultaneous optimization of the two data types, resulting in information loss and reduced robustness\". Could the authors clarify how SKETCH addresses these challenges?\n\n3. The paper lacks implementation details and accessible code. How do authors fine-tune LLMs? What is train / val / test split? What is the searching space for each hyperparameter, e.g., $k$-hop? The reproducibility claims in the article are not convincing. The claim that \"the results and related analysis reported in the paper are only a summary of those available in the code\" is ambiguous.\n\n4. What is $G$ in $G = R_{sum} + R_{struct}$?\n\n5. It is fair to use frozen / fine-tuned LLMs as baseline. However, comparing the proposed model with tailored TAG models that do not utilize external knowledge bases may be unfair. Why not include RAG-based approaches for TAGs?\n\n6. What is external knowledge data used for each dataset?\n\n7. What are promps used for proposed model (SKTECH)? Are the prompts employed for the LLM baseline the same as those used for SKETCH?\n\n8. Why SKTECH with Nomic (127M parameters) perform better than SKTECH with Llama3 (8B parameters) on Wikipedia when Nomic has much fewer parameters?\n\n---\n[1] Lewis, Patrick, et al. \"Retrieval-augmented generation for knowledge-intensive nlp tasks.\" Advances in Neural Information Processing Systems 33 (2020): 9459-9474.\n\n[2] He, Xiaoxin, et al. \"G-retriever: Retrieval-augmented generation for textual graph understanding and question answering.\" arXiv preprint arXiv:2402.07630 (2024).\n\n[3] Hu, Yuntong, et al. \"GRAG: Graph Retrieval-Augmented Generation.\" arXiv preprint arXiv:2405.16506 (2024).\n\n[4] Edge, Darren, et al. \"From local to global: A graph rag approach to query-focused summarization.\" arXiv preprint arXiv:2404.16130 (2024)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The retrieval-based approach offers a fresh perspective on handling TAGs.\n\n2. The model leverages hash-based similarity estimation to reduce computational costs in multi-hop similarity estimation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces SKETCH, a novel framework for handling text-attributed graphs (TAGs) based on retrieval-augmented generation, enhancing large language models (LLMs) for TAG-related tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Work on graph retrieval-augmented generation, which is closely related to the topic of this paper, is not discussed.\n\n2. Lack of implementation details such as hyparameter searching space, prompts used, 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": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Could you release the train/test/val splits of the three datasets? I haven't found it in Appendix A.1 and main text.\n- Could you provide more explanation on the claim that SKETCH requires fewer computational resources than other baselines?" }, "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 is well motivated and easy to follow\n- SKETCH givens a new perspective to integrate LLMs and the graph task\n- The writing and presentation of this paper is clear" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed Semantic Knowledge and Structural Enrichment framework (SKETCH) to extract the semantic and structural related information from the graph to help graph understanding and reasoning. The conducted experiments show that SKETCH could enhance the model's performance on three graph datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The evaluation is only limited to 3 datasets with less than 10 classes. InstcurtGLM [1] was evaluated on Ogn-arxiv, while GraphFormers [2] was evaluated on Product, DBLP and Wiki.\n- The improvement from Llama3-8b+GraphSAGE scenario is marginal.\n- SKETCH requires extensive hyper-parameters tuning compared to existing graph based methods (such as Llama3-8b+GraphSAGE).\n\n[1] Ye, Ruosong, et al. \"Natural language is all a graph needs.\" arXiv preprint arXiv:2308.07134 4.5 (2023): 7.\n\n[2] Yang, Junhan, et al. \"Graphformers: Gnn-nested transformers for representation learning on textual graph.\" Advances in Neural Information Processing Systems 34 (2021): 28798-28810." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "we propose a novel framework to adapt the long-context model for graph learning by retrieving both structural and text-related content." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024large,\ntitle={Large Language Models based Graph Convolution for Text-Attributed Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x5FfUvsLIE},\nnote={under review}\n}" }, "abstract": { "value": "Text-attributed graph (TAG) tasks involve analyzing both structural information and textual attributes. Existing methods employ text embeddings as node features, and leverage structural information by employing Graph Neural Networks (GNNs) to aggregate features from neighbors. These approaches demand substantial computational resources and rely on two cascaded stages, limiting scalability in large-scale scenarios and making them vulnerable to the influence of irrelevant neighboring nodes. The advancement of language models (LMs) presents new avenues for tackling this task without GNNs, leveraging their ability to process text attributes of both the target node and its important neighbors. Instead of using graph convolution modules, LMs can assign weights to these tokens based on relevance, enabling token-level weighted summarization. However, it is nontrivial to directly employ LMs for TAG tasks because assessing the importance of neighbor nodes involves both semantic and structural considerations. Additionally, the large search space presents efficiency issues for computing importance scores in a scalable manner.\nTo this end, we propose a novel semantic knowledge and Structural Enrichment framework, namely SKETCH, to adapt LMs for TAG tasks by retrieving both structural and text-related content. Specifically, we propose a retrieval model that identifies neighboring nodes exhibiting similarity to the target node across two dimensions: structural similarity and text similarity. To enable efficient retrieval, we introduce a hash-based common neighbor estimation algorithm for structural similarity and a nearest-neighbor recalling algorithm for embedding similarity. These two similarity measures are then aggregated using a weighted rank aggregation mechanism. The text attributes of both the retrieved nodes and the target node provide effective descriptions of the target node and are used as input for the LM predictor. Extensive experiments demonstrate that SKETCH can outperform other baselines on three datasets with fewer resources." }, "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": [ "Text attributed graphs", "Long-context 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/2ce49d281ee362303d95e4628c899125defa5dd1.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Large Language Models based Graph Convolution for Text-Attributed 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": 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 the weaknesses. I'll ask more questions after looking at the author's rebuttal." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The topic this paper tries to push is timely, as there have been several works in Quantization Aware Training of low-bitwidth models recently, which show how these models start working better at scale.\n\n2. This paper takes an interesting NTK perspective to justify training dynamics and generalization of binary neural networks. They also derive a scaling law of 1-bit neural networks under certain assumptions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides theoretical justification for scaling 1-bit neural networks, showing that their training dynamics converge to kernel-like behavior as model width increases. The authors demonstrate that the training loss can become arbitrarily small with sufficient network width and that the generalization difference between 1-bit and full-precision networks shrinks with scale. Using the Neural Tangent Kernel (NTK) framework, the authors guarantee training convergence to minimal loss. Preliminary empirical results confirm that 1-bit models perform comparably to full-precision networks on complex tasks, with significant efficiency gains. This work suggests that scaling quantization aware training of 1-bit neural networks are a promising direction for efficient deep learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main weakness of the paper is the lack of realistic experiments and the positioning of the paper. The authors use 1-bit LLMs to motivate the paper in the abstract and introduction, but their theory and experiments are for quantized MLPs at a very small scale. The authors should rewrite the abstract and introduction to clarify what their focus is on rather than overemphasizing low-bitwidth LLMs. For example, authors should explicitly state their focus is on quantized MLPs for learning functions, and clarify how their results may or may not extend to large-scale LLMs.\n\n2. In section 3.2, the authors mention $f$ is a two layer attention model; however, the equations suggest it’s a two-layer MLP. Symbols like $m$ are not defined clearly. This section needs better clarity.\n\n3. The choice of experiments (learning rigorous functions) seems arbitrary to me, given that the motivation of the paper is 1-bit LLMs. Can the authors explain this disparity?\n\n4. Authors should look at related papers that try to empirically derive the scaling law of low-bit (binary and ternary) quantization-aware training of LLMs like [1], which might make the arguments in their paper stronger. They also provide the model checkpoints and losses at various scales for 1-bit and 1.58-bit LLMs. I think if the authors want to position their paper as validating the scaling of low-bit LLMs using their theory, they can use the existing checkpoints from this work.\n\n[1] A Kausal et al. Spectra: Surprising Effectiveness of Pretraining Ternary Language Models at Scale" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "See above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper studies an important topic of training quantized networks, from a theoretical point of view.\n- Showing convergence of 1-bit network with quantization aware training is, to my knowledge, novel.\n- Studying the relation between 1-bit network and full-precision networks is interesting and important." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work studies theoretically the power of 1-bit networks, using NTK theory. Specifically, the authors bound the convergence time for training 1-bit quantized network using Quantization Aware Training, where the network is quantized in the forward pass, and computes the gradient with a straight-through estimator in the backward pass. They also relate the convergence guarantees to scaling laws for 1-bit networks, and complement these scaling laws with experiments. Additionally, the authors bound the difference between the quantized and non-quantized network at initialization and during training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lemma 4.1: the results in the lemma depend on $\\lambda$, which I assume is $\\lambda_{\\min}(H^*)$ (this should be properly stated). However, if this is the case $\\lambda$ should depend explicitly on $\\kappa^2$, and the dependence on $\\kappa^2$ should be tracked throughout the results. It is better to define $\\lambda$ as the minimal eigenvalue of the non-scaled Gram matrix, and have the explicit dependence on $\\kappa^2$ in the results.\n- It is worth noting that the scaling laws derived in proposition 4.3 are very different form than the Kaplan scaling laws (e.g., $N$ and $D$ in the Kaplan scaling laws are in the denominator with fractional power, and are not exponentiated).\n- The authors claim that the function implemented by the quantized network approaches the function computed by the full-precision network as the network size grows, but I do not see how this is reflected in the results. Specifically, it is worthwhile to write down an explicit bound on the difference between the two networks that goes to zero with $m$, when keeping other parameters (e.g. dataset size and network scale) constant. I currently did not see such a bound, but maybe I am missing something.\n- Related to the above, the main results that relate the quantized and non-quantized networks (Theorems 5.1 and 5.2) rely on the scaling parameter $\\kappa$ being very small. However, clearly taking $\\kappa$ to zero makes both functions go to zero, and thus trivially the difference between the two functions becomes small. But this is also true for any (bounded) function multiplied by $\\kappa$, so it seems that these results trivially hold. Do the results simply imply that all functions go to zero with $\\kappa$ and thus become close, or can these result hold when both functions are bounded away from zero (e.g. by a constant)?\n- The functions studied in 6.1 are relatively low-dimensional, which may mean that they can be approximated easily in the kernel regime by any kernel. Did you try experimenting with functions of higher dimensions?\n\nAdditionally, there are some unclear sentences/phrases in the paper, e.g.:\n- Lines 392-395, these two sentences are not clear.\n- Line 347\n- Line 260-261\n- Line 422: \"we aimed to learn rigorous functions\" - what does \"rigorous function\" mean here?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Could the authors provide evaluation on practical models, e.g., GPT-2 and validate the proposed scaling law?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- This paper studies an important problem as to scaling 1-bit neural networks. Given the increasing deployment of large foundation models, how to reduce their energy consumptions become an increasingly important problem. 1-bit quantization is a promising approach to study.\n- The authors conduct thorough experiments to validate the theoretical analysis." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper conducts a theoretical analysis on scaling 1-bit neural networks. The authors use the NTK framework to derive a scaling law formula for 1-bit quantized networks. Experiments on simulating target functions demonstrate the effectiveness of the proposed scaling law." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- This paper studies the scaling law of 1-bit neural networks. However, the analysis is performed on simple toy models. The original scaling law as proposed by Kaplan [1] is conducted on Transformers with millions/billions of parameters with huge amounts of pretraining compute. The goal of scaling law should be to study how to reliably predict the performance of pretraining with larger compute. However, the setting of this paper does not fit in the category of “scaling law”, given the small scale experiments conducted on toy models. \n- The experiments do not support the claim that there is a scaling law for 1-bit quantized neural networks. The authors do not demonstrate the exact scaling law formula, e.g. what are the values of hyperparameters in proposition 4.3 when fitting this curve to the experiments in Figure 1. Obtaining the coefficient in the scaling law is of practical importance, which can be used to predict performance of training larger 1-bit quantized models. \n- What is the justification for using NTK for studying this problem? The introduction of section 3.3 seems rather abrupt, which lacks the discussion for motivation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See Weaknesses" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) The paper introduces a variation of the soft-commite machine that can be analytically traced to show the convergence of 1-bit models.\n\n2) The convergence is analysed for the number of model parameters and the size of the training data set." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper shows that a modified 1-bit soft-commite machine converges to an arbitrarily small loss as the width of the hidden layer goes to infinity. It also gives a theoretical upper bound on the rate of convergence.\nThey also show that a 1-bit model can be trained to estimate complex functions with high accuracy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Given that the paper provides numerical results, I found it problematic that the main result of the paper (the functional form of the convergence of the training loss of 1-bit models) is not verified with numerical results. \n\n2) The finding of a \"scaling law\" in Fig.1 with four data points, no comparison with a theory curve, and significant fluctuations is, in my opinion, not a valid verification of Proposition 4.3 at all.\n\n3) The figures are not properly described/referenced in the text. \n\n4) The introduction of $\\kappa$ in the model seems a bit arbitrary, especially considering that it is necessary to \"plug an appropriate value of $\\kappa$\" (Line 391) into the model to ensure learning.\n\n5) The abstract does not adequately explain what is being done in the paper. Instead of citing previous results in the abstract, I would like to see the model and a more precise definition of the results there." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024unlocking,\ntitle={Unlocking the Theory Behind Scaling 1-Bit Neural Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x5YEibapUM},\nnote={under review}\n}" }, "abstract": { "value": "Recently, 1-bit Large Language Models (LLMs) have emerged, showcasing an impressive combination of efficiency and performance that rivals traditional LLMs. Research by Wang et al. (2023); Ma et al. (2024) indicates that the performance of these 1-bit LLMs progressively improves as the number of parameters increases, hinting at the potential existence of a *Scaling Law for 1-bit Neural Networks*. In this paper, we present the \\emph{first theoretical} result that rigorously establishes this scaling law for 1-bit models. We prove that, despite the constraint of weights restricted to $\\\\{-1, +1\\\\}$, the dynamics of model training inevitably align with kernel behavior as the network width grows. This theoretical breakthrough guarantees convergence of the 1-bit model to an arbitrarily small loss as width increases. Furthermore, we introduce the concept of the generalization difference, defined as the gap between the outputs of 1-bit networks and their full-precision counterparts, and demonstrate that this difference maintains a negligible level as network width scales. Building on the work of Kaplan et al. (2020), we conclude by examining how the training loss scales as a power-law function of the model size, dataset size, and computational resources utilized for training. Our findings underscore the promising potential of scaling 1-bit neural networks, suggesting that int1 could become the standard in future neural network precision." }, "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": [ "1-bit neural network", "quantization", "neural tangent kernel" ] }, "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/a2abe47d145314b7cd07f12f1275caf9ca7f15bb.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/f647d5e89edaaa41d6b8ae20fcdc7bba5b42fd54.zip" }, "title": { "value": "Unlocking the Theory Behind Scaling 1-Bit Neural 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": 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": "- The caption in Figure 2 should contain ideally an explanation of what is seen in the figure. At least a reminder of the loss terms would be helpful to avoid having to jump back and fourth through the paper.\n\n- How does Saner apply to non-binary gender? Does it properly use general pronouns like they?\n\n- Conceptually, it seems hard to draw a line between terms related only to the gender and to the gender and a concept. Should de-biasing replace terms like pregnant? There is no male counterpart of this so replacing it will probably remove important information.\n\n- Is the \"multilayer perception\" in line 81 supposed to be a multilayer perceptron?\n\n- The definition of a dataset in Line 104 seems odd in the sense that, if a is a protected attribute then this dataset should only contain data with protected attributes. Is that intended? I would imaging the goal is to train on a large dataset which can have samples with protected attributes but not all samples have to have protected attributes?\n\n- One added sentence for 2.2 on how a orthogonal projection is helping debiasing would be helpful for understanding.\n\n- How hard is it to re-train clip? When the semantics are lost the regularization loss may be high. How computationally heavy is it to get to the same or a similar performance level?\n\n- Table 2's caption should probably read \"racial bias\" and not recial bias." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is very clearly written. Whenever an uncertainty comes up about a term footnotes help understanding.\n\nThe analysis of debiasing techniques is very clearly presented that it could be used for a tutorial. The experimental setup is well done, the evaluation metrics such as measuring the difference between a uniform distribution and a potentially gender biased image generation seems adequate for the task. Showing generated images, i.e. with Stable Diffusion shows also a very immediate practical application of this research.\n\nThe figures help understanding of the method even though the caption could be changed to make them more self-complete.\n\nIn general, this seems a very well written, easy to understand paper addressing a very current need." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to overcome societal bias present in datasets used to train large scale multi-modal models like CLIP. The authors provide a study of debiasing methods and explain how these struggle with loosing important attribute information, e.g. gender, or depend on attribute annotations which are hard to get and even harder to get in the form of a large diverse dataset. Their contributed method Saner has four components which together aim to remove biases: attribute neutralization, feature modification and attribute annotation free debiasing and regularization losses. The paper compares the results on quite a few tasks such as text-to-image generation, text-to-image retrieval and zero-shot image classification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The attribute groups will be limited to a specific set of defined attributes, which need to be agreed upon and could be debatable, e.g. \"pregnant\" (If I understand list C in the appendix correctly). There may also be missed attributes. This does not seem like a big issue but could be one in an adversarial setting. \n\nIn general, the benefits of not needing a annotated dataset with attributes seems to be bought by needing a general list of attributes. This may not be a weakness per se but it would be interesting to either read about it in the limitations or have the author comment on why this is not a limitation.\n\nThere are general implementation details about the used architecture but none about the training process. Re-training clip with a loss to keep it's performance while having a re-projection that changes the semantics seems complex. Information about the used GPU would help understand how hard training for 5 epochs is.\n\nComparison against other state of the art in this small field is limited but understandably so and the author's did a good job re-implementing relevant approaches." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "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": 2 }, "primary_area": null, "questions": { "value": "Please see the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The problems of existing methods are straightforward, and the authors conduct several experiments to verify these phenomena.\n- The proposed method has a good performance.\n- The experiments are conducted on both text-to-image retrieval and generation tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to address societal biases in large-scale vision-language models, such as CLIP. The authors identify two limitations of existing methods: the loss of attribute information and the reliance on attribute annotations. To overcome these challenges, the authors propose SANER, a simple yet effective debiasing method for CLIP that incorporates attribute neutralization and an annotation-free debiasing loss. Extensive experiments are conducted to verify the method's effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Could the authors provide ARL results to verify the loss of attribute information? Besides, do other existing methods, such as Mapper and Prompt tuning-based debiasing, also have the lossing attribute information when debaising?\n- Although effective in debiasing CLIP, the proposed method appears ad-hoc for this specific task and lacks major technical contributions, as most components seem to be existing technologies or tricks. Therefore, this paper may not have significant influence or provide broader insights beyond the CLIP-debiasing task.\n- Why does adding only contrastive losses have a significantly negative impact on FairFace performance (as shown in Table 7)?\n- Is the proposed method only applicable to CLIP? Could the authors test other VLMs to demonstrate its broader 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": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The two identified challenges, especially the loss of attribute information, sound critical and interesting. It seems like the pipeline only involves lightweight training, as only the debiasing layer is trained. The pipeline is evaluated on two different downstream tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a CLIP debiasing pipeline that aims to address two existing challenges: 1. loss of attribute information and 2. Dependency on attribute annotations. Specifically, the paper focuses on adjusting the text feature by replacing the protected attributes with attribute-neutral terms and further regularizing the adjusted feature by two reconstruction losses. The pipeline is evaluated on Text-to-Image retrieval and generation tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although replacing the protected attribute words with an attribute-neutral word is a sound neutralization method, it requires a comprehensive list of the protected attributes (as listed in Appx. C). However, in real practice, and especially in non-binary cases, it will be very hard or nearly impossible to have a complete list. Further, since the \"attribute annotation-free debiasing loss\" relies on a set of attribute-specific descriptions, creating the attribute-specific descriptions set may lead to some concerns. For example, if the set is created by iterating all possible words and all possible groups (e.g., man, woman, male, female, he, she, etc), then the size of the set will be overlarge and increase the computation complexity. On the other hand, if only one word is selected for each group (e.g., man + female, woman + male), then how to select matching words from groups may also affect the debias performance. Also, is it possible that the generated description may have unmatched text (e.g., A woman/boy is eating salad) or text that has grammar errors (e.g., A he/she is eating salad, according to Appx. C). I would like to see more discussion on the text generation. \n\n2. The proposed pipeline tried to address the challenge by combining the debiasing loss with two extra regularization losses. The debiasing loss tries to place the adjusted feature of the attribute-neutralized text in the middle point of several features of the attribute-specific text, which seems to be fine. However, the regularization losses try to project the attribute-neuralized feature to both text and visual features of an attribute-specific feature, which seems strange. As demonstrated in Figure 2, the input to the \"Text encoder\" has been neutralized, and the attribute-specific information has been removed. Further, the debiasing layer is also trained to adjust the features to be further neutral. Thus, even there is a residual connection in the \"Feature modification\" process, the output feature should not contain any information regarding the specific attribute. Then, how is it possible that such a feature can be mapped to either text or visual features that contain specific attribute information to avoid \"the loss of attribute information\"? Is it possible that the \"retained attribute information\" is some biased information from the dataset that is implicitly polluted by the regularization loss?\n\n3. For the evaluation of the text-to-image generation, the paper only provides numerical results on debiasing performance and some visualizations but does not provide any image generation imperial results, like FID. Also, only a few SOTA are used as the baseline, which makes the comparison weak. However, this may be understandable as several baselines do not have public implementation, and it seems like the paper prepares all the baseline results." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "Have you considered running experiments with LLaVA-like models? I don't think it would be terribly difficult to replace the clip model with the proposed one, run standard llava experiments (there are a few githubs that are easy to setup for such eval), and report the results. As for debiasing experiments, I've seen people use PATA and FairFACE and just look at the next token predicted to get similar scores as the CLIP style. Showing this CLIP replacement method works on text generation as well as image generation would very much increase the impact of the work, in my opinion. However, I realize this experiment may best be left for future work." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "I will start off by saying that the lede of the paper was buried: section 5 as an experiment is astounding. I absolutely love the drag and drop replacement of this debiased model into an existing model. The idea that, 1. you don't have to retrain the text-to-image model and 2. that you don't have to retrain CLIP from scratch is incredibly compelling. I recommend the abstract/intro be rewritten to highlight this experiment.\n\nOn another note, the paper is generally well-written and easy to follow. The problem of bias in VLMs is highly prominent and this work makes an excellent contribution to the field. I am fairly familiar with this field of debiasing clip and I feel section 2.1 does a good job summarizing the various papers." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method for finetuning a CLIP model to remove social bias by an intuitive training scheme." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I don't have too many issues with the paper. One thing I would like to have seen is a more automated method for text neutralization. While word level ablation is intuitive and makes sense for the social bias angle, a less hardcoded approach seems like it would be needed for a general-purpose debiasing solution, especially for concepts that are hard to ablate at the word level.\n\nThere's a dataset called FACET that addresses some issues with PATA and FairFace. If there's time, it would be great to run experiments on it as well. This would cement the proposed method as useful and cutting edge in my opinion." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose SANER, a debiasing method for CLIP that removes societal bias without requiring attribute annotations, while preserving attribute-specific information." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024saner,\ntitle={{SANER}: Annotation-free Societal Attribute Neutralizer for Debiasing {CLIP}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x5hXkSMOd1},\nnote={under review}\n}" }, "abstract": { "value": "Large-scale vision-language models, such as CLIP, are known to contain harmful societal bias regarding protected attributes (e.g., gender and age). In this paper, we aim to address the problems of societal bias in CLIP. Although previous studies have proposed to debias societal bias through adversarial learning or test-time projecting, our comprehensive study of these works identifies two critical limitations: 1) loss of attribute information when it is explicitly disclosed in the input and 2) use of the attribute annotations during debiasing process. To mitigate societal bias in CLIP and overcome these limitations simultaneously, we introduce a simple-yet-effective debiasing method called SANER (societal attribute neutralizer) that eliminates attribute information from CLIP text features only of attribute-neutral descriptions. Experimental results show that SANER, which does not require attribute annotations and preserves original information for attribute-specific descriptions, demonstrates superior debiasing ability than the existing 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": [ "Societal bias", "CLIP", "Debiasing", "Fairness" ] }, "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/3ac1d2afcfe95d6aad2bbb3e3a5c69095a0431cc.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": "SANER: Annotation-free Societal Attribute Neutralizer for Debiasing CLIP" }, "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": "No further questions." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The theoretical foundation of h-SVGD is established, which is new and important to this research topic.\n\n2. Besides the new results, some existing results are proved with relaxed and more practical assumptions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed studied the theoretical foundation of the hybrid kernel Stein variational gradient descent (h-SVGD) method, which is a variant of the vanilla SVGD method. Specifically, the authors demonstrated the ability of h-SVGD to alleviate variance collapse, and showed the existence of a solution to the hybrid Stein partial differential equation for h-SVGD. They also showed that h-SVGD does not converge to the target distribution in the mean field limit. Experiments have been provided to show the promising properties of h-SVGD." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "To be honest, since I am not familiar with the mathematical tools used in this work, I can only judge the paper based on what the authors have done, as they claimed, and I cannot say much about its weaknesses." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "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": "- The result in Corollary 3.4 indicates that h-SVGD does not guarantee the distribution of particles converges to the target distribution. While this is reasonable, the metrics commonly used in Bayesian Neural Network (BNN) tasks, such as test RMSE and test LL, do not reflect this limitation. This raises the question of how to interpret the advantages of h-SVGD despite this theoretical constraint." }, "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 relatively comprehensive theoretical foundation for the empirical method hybrid kernel variant of SVGD.\n- This paper offers a clear explanation of the relationships and distinctions between the theoretical results of h-SVGD and those of SVGD." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the theoretical understanding of the hybrid kernel variant of Stein variational gradient descent (h-SVGD). Specifically, it provides a kernelized Wasserstein gradient flow representation for h-SVGD and, building upon this, offers the dissipation rate and large particle limit of h-SVGD. In numerical experiments, the authors demonstrate that h-SVGD significantly improves variance collapse compared to vanilla SVGD." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experimental results presented in the paper offer relatively limited support. In results on the Protein and Kin8nm datasets, SVGD does not appear to suffer from variance underestimation issues, yet its performance is still inferior to that of h-SVGD. This observation suggests that the advantage of h-SVGD in BNN tasks may not primarily stem from mitigating variance underestimation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- In Assumption (A2), the first part can be deduce from Assumption (A1). In fact, Assumption (A1) induce that $V$ is bounded. Why do you precise this assumption ?\n- Can you give an interpretation of the second part of Assumption (A2) ?\n- Can you give an interpretation of Assumption (A3) ?\n- Do you have examples of potentials $V$ that verifies Assumptions (A1)-(A2)-(A3)-(A4) ?\n- According to your analysis of h-SSVD you have a proposition of a sampling strategy that does not suffer from dimensional collaps and sample the right distribution ?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper provides interesting theoretical results on h-SVGD. It extend the known results on SVGD to h-SVGD. The experimental results effectively support the interest of h-SVGD." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Ths paper gives more theoretical insights about a sampling method named hybrid kernel stein variational gradient descent (h-SVGD). It is demonstrated that this method effictively sample a law which is not the target law (but linked to it). Moreover, it provides a descent result and a discretization quantification. Finally, it presents experiments that show the empirical interest of h-SVGD." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major weaknesses:\n- Assumption (A1) does not seem to be well written. To suppose that $\\lim_{|x|\\to +\\infty} V(x) = +\\infty$ must be a more realistic assumption to ensure that $\\pi$ could be a probability measure. Is it a writting error ?\n- There is a few intuitions and comments in the paper, making the paper hard to read. In particular, I suggest to discuss the assumptions in details. A reminder on Wasserstein gradient flow will be appreciable. Finally, Proposition 3.6 is not comment.\n\nMinor weaknesses:\n- Equation 6 : there is a typo, it might be $(x, \\cdot)$ instead of $(\\cdot, x)$.\n- line 198 : in the paper notations, it might be $\\mathcal{H}$ instead of $\\mathcal{H}_1$.\n- The remark in line 246 seems inconsistent because Assumption (A1) induce that $V$ is bounded.\n- In line 250, the constant $L$ is said to be dependent of $k_1, k_2$ and $p$. However it is not clear what is the variable $p$.\n- In line 329, the normal density do not verify Assumption (A1). In fact in this case $V$ is quadratic.\n- line 226, I suggest to be more precise about the sense of \"symetric function\" in this context." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- The authors mention convergence issues in the mean field limit. Are there any ways to mitigate this bias?\n- The authors focus on $k_2 = ck_1$ case. Have they tried using two completely different kernels?\n- The authors reference h-SVGD's promising results on image classification. Have they conducted similar experiments in this paper?\n- From an applications perspective, why is it important to avoid variance collapse?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Theoretical contributions are rigorous and sound, including the existence of a solution to the hybrid Stein PDE and the establishment of the descent lemma. A kernelised Wasserstein gradient flow interpretation is thoroughly discussed with $k_2 = ck_1$.\n- The experiments involve diverse datasets and metrics.\n- The representation of experiments is clear and the paper is well-structured.\n- Good to quantify that h-SVGD does not add additional computational costs to SVGD." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper discusses h-SVGD as a method to address variance collapse by using separate kernels for the driving and repulsive terms. The authors extend the theoretical aspects of the original paper and demonstrate that h-SVGD does not converge to the target distribution in the mean field limit. The empirical results support the theoretical findings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The title may be a bit misleading as it suggests that h-SVGD is a novel algorithm introduced in this paper, but it was previously proposed. For instance, it can reflect the paper's focus on theoretical analysis and empirical evaluation of h-SVGD, rather than introducing it as a new method.\n- The main focus on $k_2 = ck_1$ may be somewhat limiting, as it potentially restricts the exploration of using truly distinct kernels for the driving and repulsive terms.\n- While RMSE and LL have been assessed in Appendix B, h-SVGD does not show clear benefits over SVGD in these metrics. It would be better to provide a more detailed discussion in the main text about why DAMV is an appropriate metric for evaluating h-SVGD's performance, especially in relation to the variance collapse problem.\n- Using \"S-SVGD\" and \"SSVGD\" interchangeably is slightly confusing and could be more consistent." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024hybrid,\ntitle={Hybrid Kernel Stein Variational Gradient Descent},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x5l5PRtvul},\nnote={under review}\n}" }, "abstract": { "value": "Stein variational gradient descent (SVGD) is a particle based approximate inference algorithm. Many variants of SVGD have been proposed in recent years, including the hybrid kernel variant (h-SVGD), which has demonstrated promising results on image classification with deep neural network ensembles. In this paper, we demonstrate the ability of h-SVGD to alleviate variance collapse, a problem that SVGD is known to suffer from. Unlike other SVGD variants that alleviate variance collapse, h-SVGD does not incur additional computational cost, nor does it require the target density to factorise. We also develop the theory of h-SVGD by demonstrating the existence of a solution to the hybrid Stein partial differential equation. We highlight a special case in which h-SVGD is a kernelised Wasserstein gradient flow on a functional other than the Kullback-Leibler divergence, which is the functional describing the SVGD gradient flow. By characterising the fixed point in this special case, we show that h-SVGD does not converge to the target distribution in the the mean field limit. Other theoretical results include a descent lemma and a large particle limit result. Despite the bias in the mean field limiting distribution, experiments demonstrate that h-SVGD remains competitive on high dimensional inference tasks whilst alleviating variance collapse." }, "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": [ "Stein Variational Gradient Descent", "Approximate Inference", "Particle-based Variational Inference", "Gradient Flow" ] }, "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/58133e099f583da04b9c13de13226d65cafa8dae.pdf" }, "presentation": null, "primary_area": { "value": "probabilistic methods (Bayesian methods, variational inference, sampling, UQ, 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": "Hybrid Kernel Stein Variational Gradient Descent" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Please refer to the \"Weaknesses\" section for further information." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper presents an interesting question, namely whether backdoor can transfer between multiple languages. The paper conducts a large number of experiments and finds that multi-language models, when fine-tuned with data from a few languages, also affect the performance of other languages.\n\n- The organization of the paper is very complete, including chapters on attack settings and objectives, defenses against the proposed attack method, etc.\n\n- The paper conducts experiments on some closed-source models (such as gpt-4o) to verify the practical impact, which is of great significance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper leverages instruction tuning and explores the backdoor transfer abilities of large language models across multiple languages. The paper empirically analyzes and finds that some multi-language large models achieve over 90% attack success rates in more than 7 languages. These findings underscore a widespread and language-agnostic vulnerability that threatens the integrity of MLLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- This paper seems to only conduct empirical research? I think this kind of contribution may not be sufficient for top-level deep learning conferences such as ICLR. While this research is very interesting, I think that the findings of this study may not have very far-reaching significance. Can we further explore the underlying issues? For example, what deeper implications does this phenomenon reflect, and how can we improve the robustness of models in response to such phenomena?\n- The empirical research in this paper only includes the scenario of instruction tuning, which seems insufficient for an empirical study. We know that the backdoor community has proposed a large number of methods, and there are also various ways of applying large language models. Is the phenomenon revealed in this paper widely prevalent?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper was the first to investigate the cross-lingual transferability of backdoor attacks on LLM instruction tuning. The experiments for the cross-lingual attack effectiveness are well-designed, and the results are presented clearly in 6 European and 6 Asian languages (5,600 instances for each language)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper investigates the cross-lingual transferability of backdoor attacks in the instruction tuning of large language models (LLMs). The work demonstrates that backdoor attacks can effectively transfer across different languages and attack settings/objectives, even when the poisoned instruction-tuning data for specific languages is fixed (one or two). The authors provide experimental results with high attack success rates on models like mT5 and GPT-4o, across 26 different languages." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation of cross-lingual backdoor attack is week, because it seems that taking the same/similar threat model with the existing backdoor attacks on instruction-tuning (Inject a few ratio of poisoning data into the fine-tuning dataset).\n\n2. Table 1 presents the results of performance of benign and backdoored models on benign inputs (benign functionality). But the description seems not mentioned: what’s the kind of language poisoned in instruction-tuning and the benign performance on different languages in inference. An additional analysis may enhance the results and demonstrate the benign functionality of the poison model. \n\n3. The evaluations on different LLM tasks (e.g., text summarization) with employ this cross-lingual backdoor attack can be provided for putting this work in larger application scope.\n\n4. This paper provides attack results on cross-lingual transferability but lack of sufficient explanation to this property." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": [ "Yes, Privacy, security and safety" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could the author explain in detail the definitions of English refusal and In-language refusal?\n\n2. Could the author elaborate on the practical scenarios of backdoor attacks? Typically, adversaries inject backdoors into traditional PLMs, particularly in text classification tasks. For example, attackers may exploit triggers to evade spam detection systems. However, in generation tasks involving LLMs, the impact of an attack initiated by the adversary appears weaker compared to that initiated by the user. Research indicates that using instruction or scenario triggers can be more harmful than employing covertly selected triggers by attackers (see references [1-3]). In other words, what does it mean when a cross-linguistic backdoor is activated by an attacker? I believe it is more practical when users activate it, as attackers are the direct beneficiaries.\n\n**Reference**\n\n[1] TrojanRAG: Retrieval-Augmented Generation Can Be Backdoor Driver in Large Language Models\n\n[2] Backdooring Instruction-Tuned Large Language\nModels with Virtual Prompt Injection\n\n[3] Watch Out for Your Agents! Investigating Backdoor Threats to\nLLM-Based Agents" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Instruction fine-tuning-based backdoor exposes vulnerabilities in multilingual LLMs.\n2. The extensive evaluation effectively demonstrates the transferability of poisoned language to other languages.\n3. Well-written and readable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on cross-lingual backdoor attacks against multilingual LLMs via instruction tuning. Extensive evaluations show that poisoned one or two languages will affect the outputs for languages whose instruction-tuning data were not poisoned." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**1. Lack of a theoretical proof or interpretable analysis**\n\nAlthough the authors demonstrate the backdoor portability of the MLLM model through extensive evaluation, they do not demonstrate why portability exists in terms of methodology and interpretability. This is important for exposing the vulnerability of instruction tuning on MLLM.\n\n**2. Lack of novelty**\n\nSimilarly, due to the lack of instruction fine-tuning against MLLM vulnerability analysis (refer to Weakness 1), the work appears to be too engineered. In other words, there is no need for such a lengthy evaluation of this finding. If the authors start with a new attack strategy to achieve backdoor transferability it should be more solid. e.g. poisoning of a single language can significantly improve the attack transferability." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a novel attack demonstrating that backdoor attacks can transfer across multiple languages in multilingual LLMs, even when only a small fraction of instruction-tuning data is poisoned in one or two languages." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024tuba,\ntitle={Tu{BA}: Cross-Lingual Transferability of Backdoor Attacks in {LLM}s with Instruction Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x6YSsKYJuH},\nnote={under review}\n}" }, "abstract": { "value": "The implications of backdoor attacks on English-centric large language models (LLMs) have been widely examined — such attacks can be achieved by embedding malicious behaviors during training and activated under specific conditions that trigger malicious outputs. Despite the increasing support for multilingual capabilities in open-source and proprietary LLMs, the impact of backdoor attacks on these systems remains largely under-explored. Our research focuses on crosslingual backdoor attacks against multilingual LLMs, particularly investigating how poisoning the instruction-tuning data for one or two languages can affect the outputs for languages whose instruction-tuning data was not poisoned. Despite its simplicity, our empirical analysis reveals that our method exhibits remarkable efficacy in models like mT5 and GPT-4o, with high attack success rates, surpassing 90% in more than 7 out of 12 languages across various scenarios. Our findings also indicate that more powerful models show increased susceptibility to transferable cross-lingual backdoor attacks, which also applies to LLMs predominantly pre-trained on English data, such as Llama2, Llama3, and Gemma. Moreover, our experiments demonstrate the high transferability of the proposed attack: 1) the backdoor mechanism successfully operates in cross-lingual response scenarios across 26 languages, achieving an average attack success rate of 99%, and 2) the proposed attack remains effective even after defenses are applied. These findings expose critical security vulnerabilities in multilingual LLMs and highlight the urgent need for more robust, targeted defense strategies to address the unique challenges posed by cross-lingual backdoor transfer." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "backdoor attacks", "cross-lingual transfer", "LLMs" ] }, "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/94ea59618a4d9e7d79cd9665d249c9c71dc04227.pdf" }, "presentation": null, "primary_area": { "value": "foundation or frontier models, including LLMs" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/6a79ee83cc6be9191de6862c69641e589b25f3c0.zip" }, "title": { "value": "TuBA: Cross-Lingual Transferability of Backdoor Attacks in LLMs with Instruction Tuning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. First of all, it is unfair to use GPT-4o for a non-English benchmark because GPT-4o is predominantly trained for the English language. Second, why is the TFLOP, TTFT not reported for the RGB English dataset?\n\n2. Why is gpt-4o's baseline not reported for the LongBench multi-doc QA datasets? I think there are plenty of space left to include these results.\n\n3. Can the authors address the potential data contamination from its fine-tuning dataset to the 5 datasets used in the evaluation section (4 from long bench and 1 from RGB). Musique, wikimqa and hotpotqa are primarily based on wikipedia articles, and many datasets used in Table 6 are also based on wikipedia articles. Also, HotpotQA is included in table 6, why is it also used for testing?\n\n4. I think the paper would benefit if the authors show this turboRAG technique is applicable to other LLMs, such as the more recent state-of-the-art Llama3.1 8B and 70B.\n\n5. Currently, there is no link to code, model checkpoints and fine-tuning data. This raises issue of reproducibility." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper introduces a new computational technique that would speed up the inference time of an open-sourced RAG system. The paper makes the observation and subsequently the assumption that cross-attention between documents during inference is unnecessary. To avoid doing so would reduce significantly computation during inference. Therefore, this computational technique can be viewed as an attention-approximation. This assumption seems to hold for current RAG benchmarks such as RGB and LongBench multi-doc QA. This paper is written with sufficient motivation , and is able to explain its proposed computational technique with clarity. This paper identifies two technical problems with naively concatenating pre-computed kv caches, i.e. misrepresentation of positional ids, and a pre-trained LLM would suffer from distributional shift with reordered positional ids. The paper then proceeds to solve these two problems by using either composite or reordered positional ids, and then fine-tune a pre-trained LLM. In the experiment section, the paper shows that on RGB benchmark, turboRAG is only slightly worse than gpt-4o based naive RAG, and is significantly better than Qwen-7B based naive RAG. On 4 datasets from LongBench, this paper shows that turboRAG achieves 9x speedup and significant inference-time resource saving over Qwen-7B based naive RAG." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a kv caching computational technique that would significantly speed up (during inference time) the conventional RAG system in terms of time to first token (which would reduce user's wait time), and significantly reduce inference-time TFLOPs. In order to implement this technique, one would need to precompute the kv caches for all documents in a database. The paper makes the assumption that cross-attention between different retrieved documents are unnecessary, setting them to zero, which makes this computation technique an attention-approximation method. The paper reorder positional ids for the precomputed kv caches of the retrieved documents from a database, and using a fine-tuned LLM, it is able to achieve overall comparable performance on the RGB Benchmark as a gpt-4o powered conventional RAG system with some performance degradation in high-noise setting.\nThis computational technique is applicable to the following resource-constrained setting: an organization that has large storage space to store all pre-computed kv caches, has the hardware to fine-tune large language models (>=7B), has a preference to use open-sourced LLM (that use relative positions as positional embeddings) as generator rather than commercially available ones such as GPT-4o, whose user query instances and documents in database satisfy the assumption that cross-attention between documents are unnecessary for answer user queries, and is concerned about optimizing inference time." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. insufficient related work discussion. TurboRAG essentially uses an attention-approximation technique that assumes no cross-attention between different documents, this falls in line with many other work that manipulate and attempt to utilize the sparsity in attention maps. But few work along this line is mentioned and compared with the computational trick in turboRAG to convince me that it is a novel approach. This paper also reorders positional ids in RoPE, there are also many similar works that manipulate positional ids to various gains, such as efficiency in long-context inferencing, no such work is mentioned and compared with the positional id reordering trick used in the paper to convince me that is is a novel approach. Therefore, I would say the novelty and originality of this work is moderately low. \n\n\n2. I am concerned about the limitations TurboRAG imposes on the research direction of RAG. As I wrote in the summary section, TurboRAG requires model fine-tuning which rules out most commercial LLMs. TurboRAG also requires a significant amount of offline processing and storing all precomputed kv caches, this is generally not feasible with real-life database. This paper does not discuss storage cost and time for offline kv computation. Lastly, the assumption made in paper (cross-attention between documents are unnecessary) would limit future work that adapts this technique and prevent them to explore harder, more challenging and more meaningful RAG tasks where the user query requires a careful synthesis of information in different documents, such as scientific document summarization, meta analysis generation etc. In fact, while some current benchmarks (5 explored in this paper, 4 from long bench and RBG) do not require information synthesis between different retrieved documents, this limitation in benchmarking resources should not be the cause to make provincial and limiting assumptions that would negatively impact future research efforts in RAG. FYI no limitation section is included in this work even though there is almost one spare page. Because of these limitations, I would incline to reject this paper in its current version. \n\n\n3. I have some questions on the experiment evaluation sections that I will reserve to the Questions Section." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Could the authors provide an analysis of the storage requirements for TurboRAG compared to conventional RAG systems? This information would help readers understand the trade-offs involved.\n\n2. How does TurboRAG impact memory usage during inference, especially for scenarios with many retrieved documents? An analysis of memory consumption would provide a more complete picture of the system's efficiency.\n\n3. How does TurboRAG compare to other recent optimization techniques for RAG systems or long-sequence processing, such as efficient attention mechanisms (e.g., Performer, Reformer) or other RAG caching and optimization strategies?\n\n4. how does TurboRAG scale with increasing document collections or query complexity?\n\n5. The paper focuses on a specific LLM architecture. Have the authors tested or considered how TurboRAG might apply to other popular LLM architectures? This information would be valuable for understanding the broader applicability of the approach." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "*Originality:*\n1. Introduces TurboRAG, a novel approach to accelerate Retrieval-Augmented Generation (RAG) systems by precomputing key-value (KV) caches for document chunks offline.\n2. Proposes innovative techniques to handle attention masks and position IDs to maintain model accuracy while using precomputed KV caches.\n3. Redesigns the RAG inference paradigm by transforming the online computation of KV caches for retrieved documents into a hybrid offline-online process.\n\n**Quality:**\n1. Provides a comprehensive experimental evaluation across multiple benchmarks, including RGB and LongBench.\n2. Demonstrates significant performance improvements, achieving up to 9.4x speedup in time-to-first-token (TTFT) compared to conventional RAG systems.\n3. Conducts thorough regression tests to ensure the proposed modifications do not negatively impact the model's general capabilities.\n4. Presents detailed ablation studies on different configurations (TurboRAG-composite and TurboRAG-reordered) and analyzes their performance under various noise ratios.\n\n**Clarity:**\n1. Well-structured paper with clear sections outlining the problem, methodology, and experimental results.\nIncludes informative figures (e.g., Figure 1 and Figure 2) that effectively illustrate the differences between standard RAG and TurboRAG approaches.\n2. Provides clear explanations of technical concepts, such as the attention mask matrix and position ID rearrangement.\nSignificance:\n3. Addresses a critical performance bottleneck in RAG systems, potentially enabling their application in latency-sensitive scenarios.\nAchieves substantial improvements in TTFT without compromising accuracy, which could have broad implications for real-world RAG applications.\n4. Proposes a method that is applicable to most existing large language models without requiring modifications to the models or inference systems.\n5. Reduces computational resource utilization during online inference by 98.46% compared to standard RAG, significantly increasing the maximum supported batch size and enhancing throughput.\n\nOverall, the paper presents a novel and significant contribution to the field of RAG systems, offering a well-executed and clearly explained approach to improve their performance while maintaining accuracy. The potential impact on real-world applications and the broader applicability of the proposed techniques add to the paper's significance in the field of natural language processing and information retrieval." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "1. This paper introduces TurboRAG, a novel approach to improve the performance of Retrieval-Augmented Generation (RAG) systems without sacrificing accuracy. \n2. A new pipeline that decomposes the prefill stage of conventional RAG systems into offline and online phases, significantly reducing the overhead of key-value (KV) cache computation.\n3. Techniques to handle attention mask and position IDs to maintain model accuracy, including:\n\n\n a. Independent attention between document chunks. \n b. Rearranged position IDs for concatenated KV caches\n\n\n5. A fine-tuning approach to adapt language models to the new attention and position ID patterns.\n6. Substantial improvement in time-to-first-token (TTFT) performance, achieving up to 9.4x speedup (8.6x on average) over state-of-the-art multi-document QA benchmarks without compromising accuracy.\n7. The authors demonstrate that TurboRAG maintains comparable accuracy to conventional RAG systems on document QA tasks, even under high-noise conditions. They also show that the approach does not significantly impact the model's general capabilities across various tasks.\n8. TurboRAG's key innovation lies in precomputing and storing KV caches for document chunks offline, then directly utilizing these caches during online inference. This approach significantly reduces computational overhead and improves inference efficiency, particularly for applications with strict latency requirements.\n9. The paper provides experimental results on multiple benchmarks, including RGB and LongBench, to validate the effectiveness of TurboRAG in terms of accuracy and performance. The authors also discuss the impact on batch size scaling and overall system efficiency." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Limited exploration of trade-offs:**\nThe paper focuses primarily on the benefits of TurboRAG but does not thoroughly explore potential drawbacks or limitations. For instance: The authors do not discuss the storage requirements for precomputed KV caches, which could be substantial for large document collections. There's no analysis of how TurboRAG might impact memory usage during inference, especially for scenarios with many retrieved documents. A more balanced discussion of these trade-offs would provide a clearer picture of TurboRAG's applicability in different settings.\n\n **Limited comparison to other optimization techniques:** The paper primarily compares TurboRAG to a naive RAG implementation. However, it doesn't extensively compare against other recent optimization techniques for RAG systems or long-sequence processing, such as Efficient attention mechanisms (e.g., Performer, Reformer) and Other caching strategies or optimization approaches for RAG systems . A broader comparison to other RAG optimization approaches in addition to native RAG and also to other LLM architectures would help contextualize TurboRAG's contributions within the current state of the art.\n\n **Limited discussion of scalability:** The paper demonstrates impressive speedups, but doesn't extensively discuss how TurboRAG scales with increasing document collections at scale or query complexity. Additional experiments or analysis on scalability would strengthen the paper's claims about TurboRAG's broader applicability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- As referred to in the weaknesses, how long must the retrieved prompt be in order to offset the latency created by moving the KV cache from the CPU to the GPU? It seems that it went from 10x to 4x with only a difference of around 4k tokens.\n- Did you have any experiments on using the KV cache representations to enhance retrieval itself? It seems awkward to have to encode all documents with two models offline.\n- Is there any reason why you would expect the Naive RAG model would underperform Turbo reordered RAG model in english LongBench QA tasks? This seems very strange to me.\n- The fact that HotpotQA and DuReader are part of the fine-tuning data to enable TurboRAG and also the main experimental setting is makes it harder to tell if the method can truly generalize. Are we at least fully certain that there is no data leakage?" }, "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 clear.\n- Allowing LLMs to pre-compute the KV cache of a document for later use could significantly speed-up RAG applications and reduce their computational requirements.\n- Their fine-tuning methodology is simple, intuitive and apparently quite effective in allowing models to leverage independently retrieved KV-cache information.\n- Their approach leads to no performance degradation in tasks outside of RAG." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors of this paper propose the TurboRAG framework, which allows documents to be encoded offline by an LLM and used later for any retrieval augmented generation task. Their main contribution is the idea that a model can be fine-tuned in order to enhance its robustness to missing sections of the KV-cache, since the retrieved document KV-caches will be independent of each other (retrieved documents will not be able to attend to each other directly). This allows for a 5 to 10 times speedup in terms of the \"time-to-first-token\" without any significant performance degradation in both RAG and standard generation tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- It is unclear whether the efficiency improvements are significant when the retrieved documents are short (efficiency improvements are only measured with 8-16k tokens. I believe that most RAG settings will be mostly working with shorter prompts than that. For example, the supporting passages in multi-hop QA datasets within LongBench are actually only around 200 tokens each.\n- The metrics presented in the results sections are not well specified.\n- The table column and row titles are, especially for Table 2 and 3, need to be properly capitalized and formatted." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "N/A" }, "rating": { "value": 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 follow, making the proposed techniques accessible to readers.\n\n2. The proposed pipeline for RAG demonstrates a significant decrease in time to first token (TTFT), showcasing practical improvements in efficiency." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes caching key-value (KV) pairs for all documents within a Retrieval-Augmented Generation (RAG) system to reduce the overhead of KV cache computation during inference. Additionally, it introduces simple techniques for reordering position embeddings to optimize performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the idea of precomputing the KV cache to reduce TTFT is effective, it is not particularly novel. Prior work, such as [1], has already explored precomputing KV caches to achieve similar objectives. Although this paper mentions RAGCache, it lacks a thorough discussion that differentiates the proposed methods from existing approaches. Including a detailed comparison with RAGCache in both methodology and experimental results would strengthen the paper’s contribution.\n\n2. While precomputing document KV caches can effectively reduce TTFT, it increases storage costs, as each document needs a separate set of KV caches for different models. It is important for the paper to mention this storage issue so that readers can understand the potential trade-offs involved.\n\n[1] RAGCache: Efficient Knowledge Caching for Retrieval-Augmented Generation" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024turborag,\ntitle={Turbo{RAG}: Accelerating Retrieval-Augmented Generation with Precomputed {KV} Caches for Chunked Text},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x7NbaU8RSU},\nnote={under review}\n}" }, "abstract": { "value": "Current Retrieval-Augmented Generation (RAG) systems concatenate and process numerous retrieved document chunks for prefill which requires a large volume of computation, therefore leading to significant latency in time-to-first-token (TTFT). To reduce the computation overhead as well as TTFT, we introduce TurboRAG, a novel RAG system that redesigns the inference paradigm of the current RAG system by first pre-computing and storing the key-value (KV) caches of documents offline, and then directly retrieving the saved KV cache for prefill. Hence, online computation of KV caches is eliminated during inference. In addition, we provide a number of insights into the mask matrix and positional embedding mechanisms, plus fine-tune a pretrained language model to maintain model accuracy of TurboRAG. Our approach is applicable to most existing large language models and their applications without any requirement in modification of models and inference systems. Experimental results across a suite of RAG benchmarks demonstrate that TurboRAG reduces TTFT by up to 9.4x compared to the conventional RAG systems (on an average of 8.6x), but reserving comparable performance to the standard RAG systems." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Retrieval-Augmented Generation; Large Language Models; Precomputed KV Cache" ] }, "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/b32eddd973d058edb6548701e90c43f983ca460c.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": "TurboRAG: Accelerating Retrieval-Augmented Generation with Precomputed KV Caches for Chunked Text" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "What is the specific meaning of the mean $\\mu_c$ in the Eq.(5)? It appears that the mean $\\mu_c$ is merely a constant without practical significance. And how can it be ensured that the weak bisimulation metric effectively extracts task-relevant features?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The introduction presents a compelling motivation and provides a clear background explanation.\n\nThe proposed weak bisimulation metric is a straightforward yet effective approach, and it is easily adaptable to other bisimulation-based algorithms." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a weak bisimulation metric in sparse reward settings. Compared to the previously strict bisimulation metric methods, the weak bisimulation metric introduces two primary enhancements: (1) it relaxes the reward difference term through a trainable Gaussian distribution, alleviating potential representation collapse caused by sparse rewards; (2) it strengthens the extraction of equivalent task features by accumulating state transition distribution differences accordingly. Experimental results on DMC, Meta-World, and Adroit demonstrate that, the weak bisimulation metric indeed improves performance in sparse reward tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**1. Insufficient Coverage of Related Work.** \n\nThe current manuscript lacks adequate discussion of prior work related to bisimulation representation collapse in sparse reward tasks. Notable approaches, such as constructing bisimulation metrics through intrinsic rewards and inverse dynamics [1] or adopting action-based bisimulation to eliminate reward dependency [2], are absent. To strengthen the comparison, I recommend that the authors clarify how their proposed method differs from or improves upon these existing techniques. Specifically, they could discuss the relative advantages of their approach compared to the intrinsic reward method in [1] and the action-based method in [2].\n\n**2. Errors and Ambiguities.** \n\nThis paper contains several errors in expression and ambiguities that could affect clarity and accuracy. For example, “−” is used instead of “+” in Eq. (5) and Eq. (17). Additionally, in line 402, the statement \"We choose complex *walker_run*, *walker_walk*, *reacher_hard*, and *quadruped_run* tasks with sparse reward properties\" incorrectly categorizes *Walker* and *Quadruped* as sparse reward tasks, which are typically considered dense reward tasks as per [3]. To improve clarity, I suggest that the authors carefully review these equations and statements. It may also be helpful for them to provide clarification if their categorization of tasks differs from the standard definitions in [3], to ensure readers understand any intentional deviations from established terminology.\n\n**3. Experimental Design and Comparison Issues.** \n\nThe experimental setup has several issues:\n\n(1)The selection of comparative algorithms lacks sufficient representativeness. DrQ-v2 and DrM, while valuable, are not RL algorithms specifically designed to use bisimulation for extracting task-relevant representations. It would strengthen the study to include additional bisimulation-focused algorithms, such as MICo [4] and RAP [5], as well as other methods addressing bisimulation in sparse reward tasks (e.g., [1] and [2]). I suggest that the authors clarify their rationale for the current baseline selection and consider a broader comparison with bisimulation-based approaches to provide a more comprehensive evaluation.\n\n(2)In Figure 4, It is unfair to compare SRL (build upon DrQ-v2) proposed in this paper with DBC, as DrQ-v2 demonstrates significantly superior performance than DBC.\n\n(3)In the selection of experimental tasks, it is necessary to include experiments with a noisy background. Additionally, a more in-depth qualitative analysis of the relationship between introduced redundant information and interference from noisy backgrounds is required.\n\n(4)This paper would benefit from ablation experiments to analyze the contributions of individual modules. Specifically, an ablation study comparing the model's performance with and without the Gaussian distribution term in the weak bisimulation metric could clarify its impact and importance within the framework.\n\n**4. Reliance on Assumption A.1.**\n\nThe paper’s theoretical results rely heavily on Assumption A.1, which presumes that the sparse-reward expectation remains less than or equal to a sufficiently small constant $C_1$. However, the assumption lacks precise quantification, particularly regarding what constitutes a “sufficiently small” constant, potentially limiting the applicability of the results. I recommend that the authors either provide empirical evidence from their experiments to support the validity of this assumption or discuss the possible implications if this assumption does not hold in certain environments.\n\n[1]Kemertas M, Aumentado-Armstrong T. Towards robust bisimulation metric learning. NeurIPS, 2021.\n\n[2]Rudolph M, Chuck C, Black K, et al. Learning Action-based Representations Using Invariance[C]. RLC, 2024.\n\n[3]Tassa Y, Doron Y, Muldal A, et al. Deepmind control suite[J]. arXiv preprint arXiv:1801.00690, 2018.\n\n[4]Castro P S, Kastner T, Panangaden P, et al. MICo: Improved representations via sampling-based state similarity for Markov decision processes[C]. NeurIPS, 2021.\n\n[5]Chen J, Pan S. Learning representations via a robust behavioral metric for deep reinforcement learning[J]. NeurIPS, 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. What is the formal definition of $\\mathbb{M}$ in Theorem 3.1? What is the definition of $\\pi$-simulation and what is a least fixed point?\n2. Can you explain the logic in line 234 and 235? Theorem 4.1 only reasons about the sup over all states rather than some specific state pair.\n3. By introducing a Guassian random variable in the definition 4.2, the distance can now be negative. Besides, when $s_i=s_j$, this metric does not yield zero, not even on average. Why does it make sense\n4. I am assuming scalability is the ability to remain effective when performing on more complex tasks. In what sense is SRL scalable?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Most part of this paper is easy to read and follow.\n2. The experiment results shows that the resulting algorithm is consistently better than baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the problem of representation learning in deep reinforcement learning. It points out the fundamental challenges of previous (approximations of) bi-simulation metrics. It proposes SRL, which relaxes bi-simulation metric, aiming at solving the intractable reward difference and collapse in the sparse reward setting. Furthermore, it considers continuous differences over the transition distribution to tighten the metric. Finally, experiments are conducted to show the advantage of this algorithm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing about some concepts can be improved. While I understand bi-simulation is not new. Detailed definitions of relevant concepts should be provided, at least in the appendix. For example one would not be able to understand Theorem 3.1 if one does not read other papers. Besides, definition 4.2 is also not strict as it contains random variable in the definition of a deterministic map by default. This makes subsequent statements unclear as well.\n2. If I am understanding correctly, this paper contains overstatements on theorems. See question 2.\n3. Definition 4.2 seems unnatural to me, see question 3. More explanations should be provided\n\nTypos:\n1. In definition 4.2, the minus sign should be plus." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Why is the assumption A.1 not included in the main text, before the theorems that rely on them are presented?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The empirical results presented in Sec. 5 seem encouraging." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a \"reward-free\" bisimulation-like metric to tackle the known representation collapse issue of bisimulation metrics in sparse reward environments. The expected-reward difference term in the $\\pi$-bisimulation metric is replaced by a noise distribution and the transition distribution distance term is replaced via unrolling by a T-step discounted sum of future distances. Experiment results are shown over 3 pixel-based control suites against a few baseline algorithms." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The presentation seems too sloppy to me. The paper is littered with vague, unclear or incorrect statements that make it hard to read. \n2. The introduction section is too dense and particularly lacking in good use of scientific language.\n\n3. Some examples to the above two points:\n- L13: \"possess the superiority\"\n- L19 \"intractable reward difference\": Monte Carlo estimation makes this quite tractable.\n- L23: \"pure distribution internally\": unclear what this means.\n- L52: why \"fundamentally\"?\n- L53: \"equivalent\" task-relevant features: equivalent how?\n- L75: \"certain favorable properties\"?\n- L78: \"effective relaxation and strengthening in specific aspects\"?\n- L160: stacking frames does not lift partial observability\n- L163: why is the reward a function of $\\phi_\\omega(s_t)$ and not $s_t$?\n- L216: \"In optional $d_\\pi$\": unclear what is being said here.\n- L240: \"But seriously,...\"\n- etc.\n\n4. The motivation behind the particular alterations to the well-studied bisimulation formulation seem arbitrary and poorly motivated, sounding more like a marketing pitch than scientifically rigorous ideas. It is unclear to me how task relevance, which inherently depends on the reward function, is maintained when the reward difference term of bisimulation is replaced by some Gaussian noise arbitrarily.\n5. The method is called \"Scalable\" Representation Learning, but it isn't clear to me how the scalability of the proposed approach is any different from prior work, e.g., MiCO by Castro et al. (2022).\n6. The T-step unrolling of the transition distance seems like a hack and seems overly susceptible to high modelling errors due to compounding.\n7. The identification of the representation collapse issue of $\\pi$-bisimulation in sparse reward environments (the central focus of this paper) and Theorem 4.1 is incorrectly attributed to Liao et al. (2023), when in fact this was first studied by Kemertas et al. (2021); see their more general Lemma 2 with $c_R=1, c_T=\\gamma$. The latter work is not mentioned until the end of Sec. 4.3 in Page 7.\n8. Experiments do not compare to Kemertas et al. (2021), whose modifications of DBC (addition of embedding normalization, intrinsic rewards and inverse dynamics regularization) substantially improved performance, especially in sparse reward environments and would therefore comprise a stronger baseline. \n9. L395: unclear why \"challenging size of 2e5\" is selected as the replay buffer size. This may be unfairly disadvantaging baseline methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Are there cases where the Gaussian assumption might not hold, and what alternatives would work better in those situations?\n\nCould the performance of SRL be further improved by considering more complex forms of the transition distribution (e.g., non-Gaussian distributions), or is the flexibility provided by the Gaussian noise sufficient for most sparse-reward settings?\n\nHow does SRL compare with methods designed for dense-reward tasks? While the paper focuses on sparse-reward environments, could the approach offer any advantages in settings where rewards are more frequent?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The introduction of a weak bisimulation metric that relaxes strict assumptions while maintaining bisimulation's theoretical advantages is a novel and important step forward. The experiments are extensive, covering a variety of challenging tasks across multiple domains (DMControl, MetaWorld, Adroit), and the proposed SRL consistently outperforms baseline methods, including state-of-the-art approaches. The paper is generally well-structured, with a clear explanation of the problems with traditional bisimulation metrics and a well-motivated introduction of the weak bisimulation metric." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a scalable representation learning approach (SRL) designed to improve the stability of representations under sparse-reward settings in reinforcement learning (RL). The core contribution is the introduction of a weak bisimulation metric, which relaxes traditional bisimulation metrics by eliminating the reliance on reward differences and incorporating a trainable Gaussian distribution. This relaxation is intended to address the challenges of representation collapse and degeneration that traditional bisimulation metrics face in sparse-reward environments. Additionally, the approach introduces continuous differences in the transition distribution to enhance task-relevant feature extraction. The proposed SRL is empirically validated on several sparse-reward RL benchmarks, such as DeepMind Control Suite, MetaWorld, and Adroit tasks, where it outperforms state-of-the-art methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper assumes that Gaussian noise can effectively relax reward differences, but the implications of this assumption could vary across different environments. More detailed ablation studies that explore different noise distributions or parameterizations might provide further insights into the generalizability of the approach." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "To overcome bisimulation metrics' unstable representations in sparse reward settings, we present a weak bisimulation metric-based scalable representation learning approach for deep reinforcement learning, which outperforms SOTA baselines." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024weak,\ntitle={Weak Bisimulation Metric-based Representations for Sparse-Reward Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x7Q0uFTH2a},\nnote={under review}\n}" }, "abstract": { "value": "Recent studies have shown that bisimulation metrics possess the superiority of essentially extracting the features related to reinforcement learning tasks. However, limited by strict assumptions and the inherent conflict between metrics and sparse rewards, they suffer from serious representation degeneration and even collapse in sparse reward settings. To tackle the problems, we propose a reward-free weak bisimulation metric-based scalable representation learning approach (SRL). Specifically, we first introduce the weak bisimulation metric, which bypasses the intractable reward difference, instead leveraging a trainable Gaussian distribution to relax the traditional bisimulation metrics. Particularly, the Gaussian noise creates a flexible information margin for the metric optimization, which mitigates potential representation collapse caused by sparse rewards. Additionally, due to its pure distribution internally, the metric potentially mitigates representation degeneration resulting from inconsistent computations under strict assumptions. To tighten the metric, we accordingly consider continuous differences over the transition distribution to enhance the accuracy of the initial transition distribution difference, strengthening the extraction of equivalent task features. We evaluate SRL on challenging DeepMind Control Suite, MetaWorld, and Adroit tasks with sparse rewards. Empirical results demonstrate that SRL significantly outperforms state-of-the-art baselines on various tasks. The source code will be available later." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Deep reinforcement learning", "Weak bisimulation metric", "Representation learning", "Sparse reward" ] }, "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/782de28076e77056f2a57acc702ad3d38e314c7f.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": "Weak Bisimulation Metric-based Representations for Sparse-Reward 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": 3 }, "primary_area": null, "questions": { "value": "- What are the main differences between this work and the work by [1]?\n\n- The authors say that \"We can clearly observe that the self-generated synthetic data has higher Min-50%++ scores than the other calibration data. It indicates that the self-generated synthetic calibration data is indeed similar to the training data, confirming the validity of\nusing self-generated data as a proxy for the training data.\". The conclusion is not entirely clear to me, can you explain how to conclude that synthetic calibration data is similar to the training data in this figure?\n\n- While the paper aims to enhance general capabilities, the impact of using domain-specific calibration data for pruning models intended for specialized tasks remains unclear. do the authors have any intuition for that?\n\n[1] Miles Williams and Nikolaos Aletras. On the impact of calibration data in post-training quantization\nand pruning. In Lun-Wei Ku, Andre Martins, and Vivek Srikumar (eds.), Proceedings of the\n62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers),\npp. 10100–10118, Bangkok, Thailand, August 2024. Association for Computational Linguistics.\nURL https://aclanthology.org/2024.acl-long.544." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper effectively challenges the common assumption that post-training pruning methods are robust to the choice of calibration data. Recognizing the challenge of inaccessible training data, the paper introduces a \"self-generating then sampling\" strategy for constructing suitable calibration data. The paper provides a detailed examination of various aspects related to the self-generating calibration data strategy" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the role of calibration data in post-training pruning for large language models (LLMs). The authors find that calibration data similar to the training data yields better performance when pruning LLMs for model compression. As many training datasets for LLMs are inaccessible, the authors propose a strategy to create synthetic calibration data, which outperforms commonly used datasets in experiments. This strategy involves generating synthetic text using the LLM and then filtering out low-quality data. This synthetic data is more similar to the training data and ultimately leads to better performance for pruned LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the paper shows a correlation between training data similarity and pruning performance, it doesn't explain why this connection exists. The paper's evaluation primarily centers on overall model performance. Investigating how calibration data affects the pruning of individual model components like attention heads or specific layers could be beneficial. This granular analysis would offer a more complete picture of how calibration data impacts different parts of the LLM." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Suggestion (I): decompose or improve the table to highlight matching or exceeding the performance of using calibration data from the actual training set and exceeding the performance compared to calibration datasets belonging to other distributions. \n\nSuggestion (II): avoid redundancy in repeating the literature review and possibly summarize the questions in the introduction. In the literature review, the name of the technique corresponding to each citation could be mentioned as well.\nSuggestion (III): improve the abstract to better reflect the outcomes of the paper and be easier to read.\nSuggestion (IV): Mention somewhere that the paper will first proceed by answering the calibration data related questions and then propose a new a novel technique for its generations. Typically, one expects the main novel contribution to come first." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper productively expands on prior work to answer unanswered follow up questions related to the influence of calibration data on pruning and delivers insightful findings through a set of reliable experiments.\nIt proposes a novel and intuitive approach for the synthesis of calibration data and evaluates it empirically and theoretically while experimentally justifying major hyperparameter choices. They show that the approach can improve by up to 2.6% over using an out-of-distribution calibration dataset.\nThe paper also clearly describes background, relevant pruning approaches, the problem statement and proposed approach for calibration data synthesis as well as experimental results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Large language models with numerous parameters substantially increase deployment and inference complexity and costs. To mitigate this, post-training parameter pruning can be used which exploits the fact that neural networks are often over-parametrized. It operates selectively removing redundant parameters while aiming to preserve performance as measured using a sample of calibration data.\nThe key contributions of this paper are: (i) a (plausibly) novel data synthesis strategy for calibration data, and (ii) an investigation into the effects of size, quality, and distribution of calibration data, across different pruning hyperparameters.\nAdditionally, the paper examines major hyperparameter choices within their strategy and perform additional analysis to show that their synthesis method generates data that is distributed similar to the training data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main results are not so well represented. In Table 2, the proposed calibration data synthesis approach frequently falls behind other sources of calibration data. It’s not highlighted in the table (e.g., using colors or otherwise) whether each source was present in the training set of the evaluated LLM. That is, it makes sense to have separate comparisons for the proposed approach with each of (i), data the model was not trained on and (ii), data the model was trained on, but these seem to be mixed up in one table making it hard to interpret the quality of the results by looking at the table. The statement “Overall, our self-generated synthetic calibration data outperforms other baseline calibration data in language modeling and commonsense reasoning tasks and…” is not well justified because the remaining of the paragraph focuses on Wikipedia and C4 and its not obvious from the table that it outperforms all sources consistently over all tasks. \n\nThe paper involves some redundancies. For instance, the introduction as well as background seem to closely repeat the literature review. The questions are mentioned in the introduction then later again in section 3. Moreover, the choice of words in some of the sentences used is inadequate. For instance, the use of “value more” in “We fill this blank and surprisingly observe that the effects of calibration data even value more than designing advanced pruning strategies.” Take note as well that the paper does not convey that this “values more” than designing more advanced pruning strategies and that’s nontrivial to prove. Constructs such as “while different calibration data’s impact on pruning performance still lacks systematical exploration.” also make the abstract harder to read compared to if it was something like “while the impact of calibration data used has been…”." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Where does Figure 6 reflect the results of magnitude-based pruning?\n2. Are the conclusions and method presented in this paper applicable to LLM quantization?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. This paper introduces a criterion and construction strategy for choosing calibration data in post-training pruning, supported by extensive experimental validation.\n2. The authors conduct experiments on various LLMs and pruning methods, with multiple repetitions, to eliminate the effects of randomness.\n3. The paper is well-organized, clearly presenting the empirical studies, methodology, experiments, and results, making it easy for readers to follow the authors' arguments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the impact of calibration data in the post-training pruning of LLMs, which shows that calibration data significantly affects pruned model performance as pruning difficulty increases, surpassing the improvements from advanced pruning algorithms. The authors also find that using training data or data similar to it as calibration data significantly boosts pruned model performance. Since pre-training data is often unavailable for advanced LLMs, the paper proposes a strategy for self-generating calibration data. Extensive experiments on multiple LLMs and pruning methods confirm the effectiveness of the proposed synthetic calibration data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper only conducts experiments on unstructured and semi-structured pruning settings and does not validate the effectiveness of synthetic calibration data in more practical structured pruning.\n2. The synthetic calibration data is not a method first proposed by the authors. A recent work by Shin et al.[1] also proposed synthetic calibration data. However, the authors do not discuss the differences between that work and the others.\n3. This paper only uses data from Wikipedia to generate synthetic data. Why do you not validate the effectiveness of synthetic data generated from other sources?\n\n[1] Shin, Sungbin, et al. \"Rethinking Pruning Large Language Models: Benefits and Pitfalls of Reconstruction Error Minimization.\" arXiv preprint arXiv:2406.15524 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 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": "Could the authors provide further clarification on how efficient their proposed calibration data synthesis method is, e.g., what are the minimum data points it needs to generate for calibration?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. **Originality:** In addition to the models included by Williams & Altetras (EMNLP2024), the authors also tested with DCLM-7B. This model is designed to showcase the effectiveness of systematic data curation. They propose a self-generating calibration data synthesis strategy. \n2. **Quality:** The paper provides a systematic exploration, supported by experimental results demonstrating how different calibration datasets affect the performance of pruned models.\n3. **Clarity:** The writing is reasonably clear and easy to follow. The objective is straightforward. \n4. **Significance:** The findings have significant implications for practitioners in the field, although it has been highlighted by previous work already." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates the impact of calibration data in the pruning of large language models (LLMs). This work mainly repeats some work that has been done by Williams & Altetras (EMNLP2024), which investigates the impact of calibration data in the pruning and quantization of large language models. The authors present evidence that the quality and type of calibration data can impact pruning performance, at times more so than advanced pruning methods themselves, reflecting the results done by Williams & Altetras (EMNLP2024). They propose a self-generating calibration data synthesis strategy to create effective calibration datasets when access to training data is limited." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Lack of Novel Contribution:** The study builds on important findings by Williams & Aletras (EMNLP2024), and the findings are already been proven and previous work has done further comparing quantization and pruning. \n2. **Lack of downstream tasks experiments:** the authors only consider pruning performance and it does not necessarily reflect the downstream tasks. Previous work done by Williams & Altetras (EMNLP2024) has done a much more comprehensive evaluation of a wide range of downstream tasks. \n3. **No explanation on pruning performance:** The paper primarily evaluates \"pruning performance,\" but fails to provide a clear explanation of this metric. It's unclear whether this refers to pruning error, signal-to-noise ratio (SNR), or another measure. The authors neither explain their calculation method nor cite a source for this metric. \n4. **Experimentation with Diverse Datasets:** The experiments predominantly focus on a narrow range of calibration datasets and models. Including a broader set of datasets could provide more generalizable results and strengthen the conclusions drawn about the effectiveness of their proposed methods.\n5. **Validation or discussion of choices in methods:** There are some variables actually can be potentially impact the results, such as why 5000 samples from the Wikipedia data for generation, and why eliminate the top 20%." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024beware,\ntitle={Beware of Calibration Data for Pruning Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x83w6yGIWb},\nnote={under review}\n}" }, "abstract": { "value": "As large language models (LLMs) are widely applied across various fields, model compression has become increasingly crucial for reducing costs and improving inference efficiency. Post-training pruning is a promising method that does not require resource-intensive iterative training and only needs a small amount of calibration data to assess the importance of parameters. Previous research has primarily focused on designing advanced pruning methods, while different calibration data's impact on pruning performance still lacks systematical exploration. This paper investigates the effect of calibration data on post-training pruning and demonstrates that using calibration data similar to the training data yields better performance. Based on this finding, we propose a self-generating synthetic calibration data strategy to sample suitable calibration data for LLMs in practical scenarios with inaccessible training data. We conduct experiments on the DCLM, LLaMA-2, and LLaMA-3 models, and the results show that the proposed method outperforms commonly used calibration data." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "calibration data", "post-training pruning", "large language models" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/612762032ae529c4c51b222e9167ec9df2782d49.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Beware of Calibration Data for Pruning Large Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The idea of utilizing intermediate data to smooth the domain adaptation process sounds reasonable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a two-stage domain invariant representation learning method to address large co-variate shifts in unsupervised domain adaptation. The approach uses intermediate, unlabeled data to create smoother transitions between source and target domains, aiming to enhance classification performance under challenging conditions. The authors claim their method outperforms existing UDA models, especially when co-variate shifts are significant." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Poor clarity and organization. The paper is challenging to read, with many grammatical errors, convoluted language and unclear explanations of the methodology.\n2. Lack of rigorous validation: The theoretical claims, especially the effectiveness of two-stage learning and parameter optimization, lack sufficient mathematical justification and empirical support. In line 62, the authors claim that \"intermediate data (unsupervised) between source and target to ensure simultaneous domain invariance between source and intermediate data and invariance between intermediate and final target data\". Doesn't this imply the source and target data are domain invariant as well?\n3. The text is verbose and repetitive, making it difficult to extract key insights and understand the novelty compared to prior methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How would the method perform if an intermediate, unsupervised domain were unavailable or of low quality? Are there alternative approaches, such as synthetic data generation or transfer learning, that could help mitigate this dependency? Could the authors provide guidance on selecting or creating intermediate datasets for cases where a semantically related domain is not readily available?\n\n2. Can the authors provide specific examples or case studies to demonstrate the practical impact of the proposed parameter tuning framework on model performance? Would an experiment isolating the effects of the parameter tuning framework help clarify its practical benefits? If so, could the authors consider adding one to the paper?\n\n3. How does the proposed method address specific limitations of existing UDA techniques, such as domain-adversarial training or correlation alignment? Could the authors provide a comparison experiment or detailed analysis that highlights the advantages and trade-offs of this two-stage approach relative to traditional UDA methods?\n\n4. How does the proposed method handle highly heterogeneous domains where intermediate data is noisy or contains varying domain characteristics?\n\n5. Could the authors include ablation studies to isolate the effects of each stage in the two-stage process? This might help clarify the contributions of each component in achieving domain invariance." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper addresses a critical limitation in UDA by focusing on large co-variate shifts in two-dimensional data domains, which are common in real-world applications such as autonomous driving, human activity recognition." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a two-stage domain-invariant representation learning approach for UDA under large co-variate shifts. It uses intermediate unsupervised data to bridge the gap between source and target domains. Additionally, a theoretical framework is proposed for parameter tuning without requiring target labels." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main limitation of the proposed method is its dependency on an intermediate, unsupervised domain that is semantically related to both the source and target domains. This dataset may not always be available or feasible to collect, especially in real-world applications with limited resources.\n2. The practical application of the proposed method is not fully demonstrated. The benefits remain largely theoretical, making it hard for readers to grasp its relevance without concrete examples of its impact on model performance.\n3. Lacks a deep discussion that contextualizes how this approach builds upon or diverges from existing UDA methods. An analysis/experiment can be done to show how the proposed method addresses specific limitations in prior approaches, such as domain-adversarial training or correlation alignment techniques." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. If the reverse validation like idea is always effective?\n2. This is more like a training strategy by commenting on step-by-step manner comparing to end-to-end manner." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The two-stage domain adaptation method, i.e. two-stage DANN sound good and interesting.\n2. The reverse validation based idea is interesting and not frequent in domain adaptation community.\n3. The two-stage strategy can be scalable to other method such as CORAL." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a simple two-stage domain adaptation method by feeding source domain, intermediate domain and target domain into the model. The intermediate domain overcomes the large covariate shift problem that is widely a challenge of domain adaptation. This paper further proposes a free parameter indicator with reverse validation strategy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Domain adaptation has undergo a wide study in the past decade. However, with the multimodal large language model, domain gap has been allievated from another way, such as CLIP based [1,2]. Therefore, in this submission, the impact of this proposal may be weak.\n2. There lacks sufficient comparisons to previous SOTAs in recent works [3], particularly large vision-language model based and prompt based [4].\n3. The writting should be further improved for easier reading.\n4. Using intermediate domain as a bridge is not new because there are a wide research in DA with intermediate state [6, 7].\n5. As the algorithms of 2 stage DANN shows, the intermediate domain is required. I thinks how to obtain intermediate domain is still an open question. This is not discussed.\n6. Transformer based DA models should also be discussed since it has been widely used in domain adaptation [8, 9]. \n\n[1] Learning transferable visual models from natural language supervision. In ICML, pages 8748–8763. PMLR, 2021.\n[2] Ad-clip: Adapting domains in prompt space using clip. In ICCV, pages 4355–4364, 2023.\n[3] Gradient Harmonization in Unsupervised Domain Adaptation. In IEEE TPAMI, 2024.\n[4] Domain adaptation via prompt learning. In IEEE TNNLS, 2023.\n[5] Domain-Agnostic Mutual Prompting for Unsupervised Domain Adaptation. In CVPR, 2023.\n[6] Semi-Supervised Domain Generalization with Evolving Intermediate Domain. In PR, 2023.\n[7] Manifold Criterion Guided Transfer Learning via Intermediate Domain Generation. In IEEE TNNSL, 2019.\n[8] Tvt: Transferable vision transformer for unsupervised domain adaptation. In WACV, 2023.\n[9] Safe self-refinement for transformer-based domain adaptation. In CVPR, 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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": "Use cases of the proposed methods are too limited. It was designed for two dimensional data. What if the source and target data are in high dimensional space?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. A new approach was proposed\n2. It provides an automated free-parameter tuning method without needing access to target ground truth labels." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors of this manuscript propose a two-stage domain-invariant representation learning method, which uses semantic intermediate data to bridge the gap between source and target domains. This method improves classification performance even under large covariate shifts by learning domain-invariant features and optimizing task discriminability through source labels. The paper also introduces a theorem for optimizing free parameters by measuring the gap between trained models and target labeling rules. The proposed method outperforms previous UDA techniques across 38 tasks in 4 representative ML datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. English was not used properly:\n a) line 52: \"did not be\"\n b) line 498: \"It can be read that\" sounds awkward.\n ...\nLots of sentences in this manuscript are not authentic, making it hard to follow the manuscript's content. I strongly recommend authors taking times to improve the presentation of this work.\n\n2. No related works section.\n\n3. The x, y axis labels for figure 4-6 are not easily visible." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 1 }, "primary_area": null, "questions": { "value": "Please refer to weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper introduces a novel and effective solution to a practical challenge in machine learning by using intermediate data to bridge large domain gaps.\n2. The proposed method is versatile as it can be integrated with various domain invariant representation learning techniques.\n3. The authors derive a theorem for measuring the gap between trained models and unsupervised target labelling rules for hyper-parameters searching." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses large co-variate shift problems in UDA, particularly when dealing with two-dimensional co-variate shifts. The proposed method uses an intermediate unsupervised dataset to bridge the gap between source and target domains, learning domain-invariant features simultaneously between source-intermediate and intermediate-target pairs, which helps achieve better domain adaptation compared to direct source-target adaptation. The authors also derive a theorem for measuring the gap between trained models and unsupervised target labelling rules, which helps optimize free parameters without access to target labels. The proposed method is validated on classification 4 datasets including 38 UDA tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThis paper requires substantial improvement in terms of writing quality and clarity of presentation. (1) Many notations are without proper definition, (e.g., D_S, \\hat{y}^S_{i,j}, N appear before being formally introduced); (2) The language is frequently imprecise and contains awkward expressions that hinder understanding. I strongly recommend the authors to thoroughly revise the mathematical notations with clear definitions and improve sentence structures and word choices for greater presentation.\n2.\tThe authors assume the existence and availability of appropriate intermediate domains that perfectly fit their \"two-dimensional domain shift\" framework, but do not adequately address how to identify or construct such intermediate domains in real-world applications. For example, while it is intuitive to have MNIST-M as an intermediate domain between MNIST and SVHN, in most real-world scenarios, it is unclear: (1) How to systematically identify the two dimensions of domain shift; (2) How to obtain or construct suitable intermediate domain data; (3) What to do when clean intermediate domains are unavailable or when domain shifts are more complex than two-dimensional. Without addressing these practical concerns, the proposed method may have limited applicability in real-world domain adaptation problems.\n3.\tThe technical novelty of the proposed method is limited. The approach is essentially a straightforward extension of existing domain-invariant learning methods. It merely splits the domain adaptation loss into two components (L_domain(S,T) + L_domain(T,T')) and applying standard adversarial training techniques, which lacks novel methodological designs in terms of loss function formulation, optimization strategy, or network architecture.\n4.\tThe proposed parameter selection method is incremental. The proposed approach is essentially a straightforward adaptation of the existing Reverse Validation (RV) method to the two-stage domain adaptation setting, without substantial methodological innovations.\n5.\tThe experimental evaluation of the paper is not up to current standards in domain adaptation research. The comparisons are limited to classical UDA methods like DANNs and Deep CORAL, while ignoring numerous recent advanced techniques that have shown significant improvements in handling large domain gaps." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Novel method of domain invariant representation learning for large co-variate shift in unsupervised domain adaptation problem" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024two,\ntitle={{TWO} {STAGES} {DOMAIN} {INVARIANT} {REPRESENTATION} {LEARNERS} {SOLVE} {THE} {LARGE} {CO}-{VARIATE} {SHIFT} {IN} {UNSUPERVISED} {DOMAIN} {ADAPTATION} {WITH} {TWO} {DIMENSIONAL} {DATA} {DOMAINS}},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x8jxf3byli},\nnote={under review}\n}" }, "abstract": { "value": "Recent developments in the unsupervised domain adaptation (UDA) enable the unsupervised machine learning (ML) prediction for target data, thus this will accelerate real world applications with ML models such as image recognition tasks in self-driving. Researchers have reported the UDA techniques are not working well under large co-variate shift problems where e.g. supervised source data consists of handwritten digits data in monotone color and unsupervised target data colored digits data from the street view. Thus there is a need for a method to resolve co-variate shift and transfer source labelling rules under this dynamics. We perform two stages domain invariant representation learning to bridge the gap between source and target with semantic intermediate data (unsupervised). The proposed method can learn domain invariant features simultaneously between source and intermediate also intermediate and target. Finally this achieves good domain invariant representation between source and target plus task discriminability owing to source labels. This induction for the gradient descent search greatly eases learning convergence in terms of classification performance for target data even when large co-variate shift. We also derive a theorem for measuring the gap between trained models and unsupervised target labelling rules, which is necessary for the free parameters optimization. Finally we demonstrate that proposing method is superiority to previous UDA methods using 4 representative ML classification datasets including 38 UDA tasks. Our experiment will be a basis for challenging UDA problems with large co-variate shift." }, "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": [ "domain invariant representation learning", "unsupervised domain adaptation", "image recognition", "signal processing", "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/08f5b0073a2cc72926c2ec2bb2c33e49f41ec017.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": "TWO STAGES DOMAIN INVARIANT REPRESENTATION LEARNERS SOLVE THE LARGE CO-VARIATE SHIFT IN UNSUPERVISED DOMAIN ADAPTATION WITH TWO DIMENSIONAL DATA DOMAINS" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "I dont have any follow-up questions." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1) the paper is easy to read\n2) The paper confirms a fact that is well know by most data science / ML prectioners, the complexity of the data set matters for classification performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an analysis of the influence of hyperparameter tuning and training \"data complexity\" (the author call it \"complexity meta-features\") on the performance of two classic classification algorithms:SVMs and Random forests. the paper includes run extensive experiments on 290 OpenML tabular datasets. The author's end with a summary of their findings: dataset complexity matters the most." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The content, experiments and conclusions of the papers are very outdated. It reads like a paper that was written 15-20 years ago. Most citations are from many years ago. Hence the contribution are practically irrelevant to the current state of ML / data science in 2024. Hence there is no significant contribution or relevance to the ICLR community.\n\n2) Furthermore, the main conclusion of the paper is that dataset complexity (class overlap, dimensionality, etc) matters when training a classifier (RF or SVM). These are well known fact that are thought in introductory ML class and hence there is no new information provided here." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "please check the weakness." }, "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 gap by directly comparing the impacts of dataset complexity and hyperparameters on model performance. Previous studies have often examined these factors separately, but this research provides a unified framework to assess their relative influence on bias and variance. \n\nThe paper comprehensively considered nearly 300 datasets and over 300 configurations, enabling a more convincing conclusion.\n\nApart from the numerical results, the paper also includes very detailed arguments of why this happens and what this indicates.\n\nSome of the estimated coefficients shown in the OLS summary table do align with our common understanding of how random forests deal with bias-variance tradeoffs, e.g. coefficients associated with min samples leaf, bootstrap, and max features." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the relative influence of dataset complexity and hyperparameters on classification model behavior, specifically for RF and Kernel SVM. \\\nAuthors utilize the fANOVA framework and OLS to quantify the influence on bias and variance brought by dataset complexity and hyperparameters. \\\nBased on the analysis across 290 datasets and 304 hyperparameter configurations, the study finds that dataset complexity meta-features—such as class overlap, data sparsity, and class imbalance—have a more substantial impact on bias and variance than hyperparameters." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. lack of details on the experiment design. \n\n (1) as the paper claims, nearly 300 datasets of varying sample sizes, response categories, and feature dimensions are used. Why are they comparable? I believe 0-1 loss is not a typical loss function people use for multiclassification problems. And high-dimensional datasets wouldn't react necessarily the same as n>p datasets in terms of hyperparameters. \n \n(2) hyperparameters like C and gamma have huge variations in scale. How is it being included in OLS? Is it logarithmized?\n\n (3) For readers not familiar with meta-features of datasets, it would be very helpful to at least sketch some general ideas of how these meta-features are defined. Are those features immune to data transformation? The same for how fANOVA works. \n\n\n2. lack of details on why the experiments are conducted in such a way. \n\n (1) In my perspective, the number of trees is RF's one of the most important hyperparameters. Why is this not considered?\n \n(2) I believe the neural net is the framework that people are most curious about. The authors also mention it in the introduction. Why is that not considered? \n \n(3) Based on the pymfe package, there are plenty of meta-features that characterize data complexity from different perspectives. Why specifically these 3, N1, T2, C1, are chosen?\n\n3. Some of the results that are confusing to me. \n \n(1) if those meta-features are immune to data transformation, how can we benefit from your research even though we know that data complexity itself is much more important than tuning hyperparameters? if not, shouldn't you include some examples of how bias and variance are reduced after some preprocessing of the data that reduces data complexity? For example, class imbalance issues can be alleviated by reweighing samples or bootstrapping.\n\nIn general, I do agree that the data's quality is much more important than tuning parameters. If the data is always linearly separable, I believe logistic regression would suffice. It's just the data quality is not something we can work on but the model and the model's parameter choice. Please correct me if I am wrong. \n\n (2) Based on your OLS example, the features included are all significant neq to 0. If the trend is determined, does it mean that choosing a smaller c or some certain kernel can always help with the prediction?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "No questions, but recommendations for future submission:\n\n- drop the analysis of dataset complexity vs. hyperparameters; just\n focus on the impact of hyperparameters on the bias/variance\n trade-off. It is sufficiently important topic on its' own. Impact of dataset complexity is not useful because 1) it is well understood that dataset complexity has major impact on classifier performance 2) it is not clear what to do about it.\n\n- include multi-class problems\n\n- include XGBoost and neural networks in the analyses. I don't think\n one can make practically useful conclusions without including those\n state-of-the-art classifiers. I would personally also add logistic\n regression\n\n- consider adding image classification datasets in your analyses. If\n the behavior - in terms of bias/variance, and hyperparameter tuning\n contribution - is quite different from tabular data, that would be a\n valuable result\n\nI think this *could* become a good paper, but not without extensive\nrevision, which is not feasible in the ICLR timeframe.\n\n\nMinor points\n\n3.3\n---\n\n- 2 out of 3 is not really \"generally\". Please just be specific: for\n N1 and T1, higher values indicate greater classification\n difficulty. For C1, lower values indicate greater classification\n difficulty. That would be simpler and easier to read. \n\n \"Higher values of these meta-features generally indicate greater\n classification difficulty (except for C1).\"\n\n4.2.2\n-----\n\n- \"When considering variance, a similar trend emerges. According to\n the fANOVA results (Figure 3b), C1 continues to dominate, accounting\n for 37.78% of the variability in variance.\"\n\n I wouldn't call this a similar trend. For bias, C1 accounts for 71%,\n for variance, 38%. Please point out that the C1 impact on variance is significantly lower than on bias. This suggests other factors play greater role. Discuss what those factors might be." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- extensive experiments\n\n- clearly written, easy to read\n\n- the topic of bias/variance tradeoff is very important" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors compare impact of dataset complexity and hyperparameter tuning on the performance of binary Random Forest and SVM classifiers. They perform extensive experiments which support their finding that the dataset complexity has dominant impact on the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I think the paper sends, in essence, a wrong message to readers.\nAuthors are basically suggesting that hyperparameter optimization\nisn't useful, which I disagree with. As I write this, I am tuning\nhyperparameters of a neural network classifier, and the AUROC has gone\nfrom 0.55 to 0.75, exclusively due to the (gradually improving) choice\nof hyperparameters. The conclusion is at best narrowly limited to SVM\nand RF, but that makes the manuscript 1) not that useful, given\nlimited scope 2) misleading, since many readers may walk away with a\nwrong impression that the conclusions apply generally \n\nTo be clear, not disagreeing with the idea that hyperparameter tuning\nhas a natural limit, and going beyond that may require additional or\ndifferent data. But the paper leaves an impression to the reader that\nhyperparameter tuning doesn't help in general, which I disagree with. At a minimum, the title should say that the results are limited to SVM and RF binary classifiers. \n\nAlso keep in mind that \"optimizing dataset complexity\" is a vague and\nhardly actionable advice. I personally don't quite know how to\noptimize dataset complexity, whereas hyperparameter tuning is well\nunderstood. This should be clearly stated/discussed." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. fANOVA is one of many ways of decomposing model predictions. Why this approach? And are there any potential issues due to taking into account pairwise interactions only?\n2. l. 60: How can normalization or scaling affect the intrinsic complexity of the datasets? Intuitively, wouldn't a reasonable measure of complexity be invariant to these? Of the three in the paper, C1 is invariant. T2 is not, because it depends on PCA, but that just makes me wonder if T2 even makes sense. I wouldn't want my dataset to become more complex, just because I convert a feature from meters to centimeters. N1 is based on a two-sample test, so, while I didn't go into the details of the test, I'd assume that we'd like all of our two sample tests also to be invariant to scaling. ... To clarify, I'm not criticizing the choice of not scaling, we can argue either way. But I am concerned about the use of complexity metrics that are not invariant.\n4.But not normalizing does have an effect on SVM and regularization? This is not how we would apply SVM in practice.\n5. N1 is a bit outdated. Two sample tests have progressed a lot in recent years. In particular, tests based on machine learning models directly or using classification performance as a proxy.\n6. It seems that N1 would fail to be attributed when the dataset is so complicated that RF and SVM can perform well, but the test in N1 doesn't?\n7. The attribution to C1 for SVM is to me the most surprising result in the paper. Any explanations of this difference between SVM and RF? How correlated are N1 and C1?\n8. l. 658: How reasonable is this assumption that there are no hidden confounding factors?\n9. Parameter configurations were generated by sampling uniformly and independently from each hyperparameter range? I'm asking because of the following scenario. Let's say that the optimal range for a parameter is relatively narrow (0 - 0.05), while after that (0.05 - 1.0) the model performs majority class baseline poorly. Because the range of good values is small, this, as a variable in a regression, would not be that important. So, it would not get much of an attribution in the experiment, but it is definitely important in practice. In other words, isn't the importance of a hyperparameter determined by the difference it can make, not by the variability of the performance over some arbitrary set of its values?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Well structured and clearly written paper.\n- A many ways a very comprehensive experiment.\n- Tackles an important issue for ML practitioners." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on the attribution of predictive model behavior. In particular, it describes a comprehensive empirical comparison of the influence of hyperparameters and dataset meta-features on the bias and variance of classifiers. The analysis utilizes functional ANOVA. The main result is that most of bias and variance can be attributed to dataset characteristics, as opposed to hyperparameters." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I have two main concerns. The first is that the experiments, while in some ways very comprehensive, are in other ways very limited:\n\n- Only two classification algorithms.\n- No missing values in the dataset is a very strong criterion.\n- A limited number of complexity measures.\n\nThe second is that when carefully interpreted, the results are not that general or actionable:\n- N1 is in essence a classifier (probably along the lines of LDA). So the results can basically be summarized that most of the bias and variability of RF and SVM can be explained by running another reasonable classifier and seeing how it performs. That of course makes perfect sense, but it can also be derived from what we already know, that classifiers tend to perform similarly (the differences between classifiers are less than the differences between datasets).\n- We should be more careful when interpreting the result that model performance can be attributed more to dataset characteristics than to hyperparameters. First, it is the nature of commonly used classifiers that they are relatively robust in terms of hyperparameter selection - being easy to tune is what makes them popular. Second, . And third, the range of several parameters is limited. For example, would results change if max_features was allowed to go below 0.1 or above 0.9? Or if 20 different kernels were considered? Similarly, the experiments are limited to 1500 features, which diminishes the importance of regularization.\n- In practice, I can in most cases freely tune the parameters and select models. I can't really change my problem (or dataset) though.\n- The paper does not consider model selection, which I would in this context consider as part of hyperparameter tuning. I would not be surprised that a lot more can be attributed to model selection than to tuning the parameters in this paper. Choosing a different model is also actionable.\n\nThere are also other methodological concerns (see Questions)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024attributing,\ntitle={Attributing Model Behavior: The Predominant Influence of Dataset Complexity Over Hyperparameters in Classification},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x8mr9zGkpr},\nnote={under review}\n}" }, "abstract": { "value": "Understanding how different factors contribute to model performance is critical for advancing the attribution of model behavior. Previous research has independently explored the effects of hyperparameters and dataset complexity meta-features on classification performance. However, a key knowledge gap remains in understanding how these factors comparatively influence model behavior, particularly bias and variance. This study bridges this gap by investigating which factors, hyperparameters or complexity meta-features, exert a greater influence on the bias-variance in classification algorithms, specifically Random Forests (RF) and Support Vector Machines (SVM). Using 290 diverse OpenML tabular datasets and 304 hyperparameter configurations, we employ functional analysis of variance (fANOVA) to quantify the impact of these factors. Our findings reveal that dataset complexity meta-features exert a more significant influence on both bias and variance than hyperparameters for both RF and SVM models. To further substantiate our findings, we conducted an analysis based on the Manipulation Theory of Causation. This analysis demonstrates that optimizing dataset complexity can simultaneously reduce bias and variance, while hyperparameter tuning often leads to a bias-variance trade-off with limited impact on overall performance. To the best of our knowledge, this research is the first to directly compare the effects of hyperparameters and complexity meta-features on bias and variance in classification, contributing new insights into model behavior attribution." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Model Behavior Attribution", "Complexity Meta-features", "Hyperparameters", "Bias-Variance Decomposition" ] }, "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/9117c4a5a601eabbe50787ac800ccba1852ce91d.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/26c499da5b57639f990615d2e806e9fef921c474.zip" }, "title": { "value": "Attributing Model Behavior: The Predominant Influence of Dataset Complexity Over Hyperparameters in Classification" }, "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": "line 239: DPO is not really an SFT method. it is a simplfied form of RLHF.\n\nline 365: should it be such as 1-0 metric, or something that takes into account the magnitude of the difference between the rewards.\n\nline 377: for DPO, are you computing the impliicit reward that the DPO derivation uses (Ratio of log probs) ?\n\nline 455 - 463: This finding if it can be replicated in other settings would be quite interesting (that reward accuracy during training is not reliable)." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "It is true that practitioners systematically under-account for the effect of annotation noise in preference-based alignment / RLHF. annotated data is frequently taken as ground truth, and there is less data cleaning performed than there should be.\n\nThe revised dataset could be an additional resource for the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper claims to be a an analysis of the role of human preference proxies (direct preferences and reward models) in LLM alignment. The authors do qualitative analyses of the errors in a commonly used alignment dataset and re-label it. They then show that this leads to a significant gain in alignment scores for both DPO style algorithms and PPO-style reward model + policy learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "After having gone through the paper, it seems the main contribution really boils down to a re-labeling of the (widely used) HH-RLHF dataset. While this will make it a useful resource, it is unclear to me what the scientific value of the work is. It is entirely expected that re-labeling the dataset to reduce noise, should improve accuracy on the downstream tasks.\n\nWhat further insights are gleaned by this exercise? There is an inventory of the data labeling losses which are mildly interesting. However, the authors give absolutely no information about the labeling process. So what can a reader takeaway that is useful? let's grant that the downstream results indicate that the labeling process of this paper is better than the labeling for the original dataset, how do we know why?\nWas the entire dataset re-labeled by the authors? this is not a process that can scale to other problems, were lower quality raters still need to be used. The inventory of \"pitfalls\" may be helpful but it is unclear whether the categories and proportions will generalize to other datasets/ domains.\n\nThe discussion section 5 does not seem to add any new insights not familar in the literature. 5.1 is more of a related work section (very minimal) and 5.2 goodharts law is well known and not quite relevant to the work done here.\n\nThis paper feel like a better fit for a conference like Findings'EMNLP." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Where is the re-labeled preference dataset?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The paper's classification of human proxies into two levels (direct preference data and reward models) offers a clear framework. This seems simple but is actually very **novel**. This is the first time I have ever seen such a classification.\n- By re-labelling the HH-RLHF dataset and providing a cleaner version (CHH-RLHF), the authors make a very practical contribution that could improve the reliability of the preference learning dataset.\n- By discussing Goodhart's Law, the paper raises awareness of the risks of over-optimizing proxies. This is indeed important for us to keep in mind: alignment isn’t just about maximising scores but ensuring genuine alignment with human intent." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper explores the challenges of using human-made proxies, like labelled data and reward models, to align large language models (LLMs) with human preferences. The authors find that these proxies often don’t accurately reflect genuine human values, which can lead to unintended model behaviours. Therefore, they re-labelled a popular preference learning dataset (HH-RLHF) to create a cleaner version, called CHH-RLHF, which they hope will support more reliable alignment. Their experiments show that better proxies lead to improved alignment, but they caution against relying too much on any single measure, as it can be over-optimized and lose accuracy. The authors urge researchers to prioritise verifying proxy reliability to prevent these hidden risks from undermining model alignment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- “This work aims to alert researchers to the importance of verifying human proxies before using them, whether for alignment optimization or evaluation.” Several research papers discuss this issue, with reward over-optimization being one of the most well-known. This topic has been widely discussed.\n\n- Relabeling HH-RLHF is valuable, but the quality of the HH-RLHF dataset is somewhat criticized in the community. The more important question is how we can collect user preferences in a scalable and reliable manner rather than simply re-labelling them. This paper is valuable for providing a re-labelled dataset, but in my opinion, it does not meet the standard of ICLR.\n\n- “However, they only establish a leaderboard for reward models, failing to reveal the relationship between alignment performance and reward model quality.” I agree, and it’s already been discussed in the community that achieving a higher score in RewardBench does not necessarily translate into better downstream task performance. How do the authors of this paper address this issue?\n\n- Table 3 is not convincing, as Figure 4 already shows that “Starling-34B” provides a better score. Why not use 34B? Instead, a 6.9B model was chosen. I wouldn’t consider this a pitfall; experiments clearly indicate that a larger base model performs better.\n\n- In Figure 5, all models are using 6.9B or 7B reward models: why not use the 34B version?\n\n- As the authors have already empirically shown that a better reward model leads to a better policy model, where is the pitfall?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- grammar L084: If not, what impacts they may have on alignment performance?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The high opposite re-label rates for HH-RLHF are insightful and highlight various issues with working with depending heavily on such datasets. In this setting, three of four annotators choose the rejected response as better, highlighting critical issues with the dataset.\n\nThe experimental design, discussion and framing are insightful. Experiments such as the correlation between human and model evaluations are particularly valuable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work explores the various complexities of alignign to human preferences, considering both the more direct capture of human feedback data as a preference proxy, and the one step removed reward model representation of human preference. It provides various experimental insights into how faithfully these reflect actual human preference. The authors analyze and provide a cleaned version of HH-RLHF, and use this as a basis for further analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The DPO experimental settings are clear, in terms of training on both the original HH-RLHF and improved CHH-RLHF datasets. However, this is unclear for the PPO setting and even on re-reading, I am unsure which results (if they exist) show the comparison between the effects of RMs trained on HH-RLHF vs CHH-RLHF on downstream LLM performance evals. If these experiments have not been run, I would strongly recommend running them as they would speak to the impact of the massive effort to re-label/clean up this data. If this is not possible due to resource constraints, I'd recommend discussing this in detail.\n\nOverall, this is thorough work which is will motivated and has involved important and substantial effort, both in terms of the CHH-RLHF annotation and experimental design, much of which is extremely insightful. I feel that currently, the poor framing of the work within the context of existing critical investigations of the space and lack of clarity around the overall narrative limit the work's potential impact. Both these points can be addressed to make this a potentially high impact contribution." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "* Some details on HH-RLHF would be helpful. Datasize? How was it originally annotated? \nIe, in Figure 3, when 0.07% is said to be “Empty” – how many samples is that? \n\n* Figure 4 – why did you include GPT-J? Also, how big of a model is GPT-J?\n* Figure 4 - why is there no correlation score?\n* Does the newly labeled data include annotations from all 4 annotators? If you have already gone through the effort to re-label the data, having this more granular information could allow for better alignment methods as opposed to binary (preferred vs. rejected) labels.\n* What is the purpose of Figure 5?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The re-labeled dataset CHH-RLHF is definitely a contribution for the community. \n* The authors do an exhaustive list of experiments, although it is not clear how the set of models chosen for each experiment was decided on." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper raises an important question on whether current day alignment datasets are truly representative of human preferences. To address this question, the authors study to abstractions of human preferences - \"Level-1 Proxy\" (aka training data) and \"Level-2 Proxy\" (reward models).\n\nStarting with Level-1 Proxy: The authors use HH-RLHF, a popular dataset, and offer a taxonomy to categorize the training data (Section 3.1), such as toxic responses, \"inverted\" responses (ie, the rejected response is actually better than the chosen respose), etc.\nThe authors then carefully re-label HH-RLHF using 4 human annotators, and find that a large chunk of the data falls under the above pitfalls.\nRe-training on the cleaned data (CHH-RLHF) using DPO, they demonstrate immediate improvements in the model's \"reward score\" (Section 3.2.2, Table 2).\n\nFor Level-2 Proxy: The authors then study the impact that a reward model can have on the alignment process. Similar to Level-1, the authors start with a taxonomy of pitfalls (Section 4.1.2). However, if I understand correctly, the taxonomy does not seem to be used anywhere? (Ex:, “Score for Empty Responses” indicates when a reward model gives high scores to an empty response – how often does something like this happen?)\n\nThe authors then claim that current reward models are not adequate enough to be used for training aligned models (Section 4.2).\nThe authors define an accuracy metric to define the quality of a reward model, and assess a wide range of reward models (Table 4), demonstrating a wide range of accuracy scores.\nThe authors study the impact that a reward model can have on alignment algorithms - PPO and PRO – and find that suboptimal reward models lead to suboptimally aligned models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* My biggest question is what new information have we learned from this? I believe that the takeaways are that better data leads to better models, and better reward models lead to better aligned models. While a cleaned dataset is a contribution worth applauding, I don’t think there is anything else that the paper offers that we didn’t already know. \n\n* The authors have a wide range of empirical results - but it is not clear how the choice of models used for each experiment was made.\n\n* Figure 4: The authors claim that Starling-34B’s evaluations show good correlation with human evaluations (But is missing a correlation score?), and also shows that Starling-34B has the highest reward model accuracy (Table 4) and using Startling-34B as a reward model leads to a better model (Figure 6). Doesn’t this go entirely against the authors claim that current day reward models (level-2 proxy) are inadequate? \n\n* Table 4: I don’t think assessing DPO for reward modeling is reasonable. There seems to be an assumption being made that the policy model from DPO would also serve as a good reward model – and the authors justify this assumption with the paper title of DPO (“Your language model is secretly a reward model”) – I don’t think this is reasonable, especially because prior work shows that policy models learn very different things from reward models (https://arxiv.org/abs/2405.19534).\nThe reason I mention this is that if you take out the DPO results in Table 4, the claims that the authors make (ie, reward models often do worse than 50/50 chance) is significantly weakened. My takeaways from Table 4 is that larger models lead to better reward modeling accuracy, and larger models (Starling-7B, 34B) do not seem to have the issues that the authors discuss." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024elephant,\ntitle={Elephant in the Room: Unveiling the Pitfalls of Human Proxies in Alignment},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x8z8hCjtcY},\nnote={under review}\n}" }, "abstract": { "value": "The demand for regulating the behavior of large language models (LLMs) has ignited research on alignment algorithms, the essence of which is to align LLMs' generations with human preferences. Due to infeasibility of humans directly participating in the training or generation of LLMs, existing alignment algorithms choose to align with human preferences carried by proxies, i.e., preference data or reward models. However, whether these human proxies faithfully represent human preferences remain under-explored. We categorize human proxies into two levels based on the degree to which they directly embody human preferences: Level-1 Proxy (preference data) and Level-2 Proxy (reward models). We empirically examine the faithfulness of both levels of proxies and its impacts on alignment performance.\nWe notice that current algorithms tend to overlook the faithfulness of these proxies in reflecting human preferences; many works even directly use reward models as their automatic evaluators without any correlation verification. Current literature of alignment overly focuses on optimizing algorithms, rendering the faithfulness of human proxies an \"elephant in the room\"—something extremely important yet largely overlooked. According to experimental results, we unveil potential risks of using inferior ``human proxies'', aiming to arouse attention to this huge ``elephant'' in alignment research. We summarize existing pitfalls from different angles and provide a re-labeled preference dataset and insights about reward model usage to facilitate the healthy development of alignment\\footnote{This work contains examples that potentially implicate stereotypes, associations, and other harms that could be offensive to individuals in certain social groups.}." }, "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": [ "Pitfalls; Human Proxies; Alignment" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/34fbc7e0927cb185993e65c39992d0efcbae5805.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": "Elephant in the Room: Unveiling the Pitfalls of Human Proxies in Alignment" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "None." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper tackles a very relevant problem.\n\nI believe the proposed approach has great merit. Training a dedicated representation network sounds like a good idea (as it is also well motivated in the paper) and might have a big impact on various quantum applications.\n\nThe paper explains the approach well and features several amazing visualizations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper tackles the well-known issue of quantum state control, i.e., steering a quantum system towards a target quantum state. To this end, the authors run reinforcement learning with the addition of a learned state representation encoded in a representation network. They show that their approach can steer quantum systems in a meaningful way." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main problem is that the presented approach is not sufficiently analyzed. Several design decisions in the algorithm and in the study are not really discussed and not analyzed. There is no ablation study or a comparison to other candidate approaches. There is no application of the approach to other similar issues.\n\nThe training behavior and training properties of the presented approach are neither shown nor discussed. Thus, this paper's contribution to an AI community is unclear.\n\nThe setting that reinforcement learning operates in is not formalized in a standard way and thus hard to follow. Giving a standard MDP-style definition would help here.\n\nSeveral typos persist. Most importantly, all references are formatted incorrectly." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 well does the proposed approach generalize to recalibrations? Is overfitting the training device a potential issue to be considered? \n\nCould the proposed method be used as an extension to existing QRL frameworks (e.g., [1-4])?\n\n[1] van der Linde et al., 'qgym: A Gym for Training and Benchmarking RL-Based Quantum Compilation', 2023. \n[2] Altmann et al., 'Challenges for Reinforcement Learning in Quantum Circuit Design', 2023.\n[3] Kölle et al., 'A Reinforcement Learning Environment for Directed Quantum Circuit Synthesis', 2024.\n[4] Rietsch et al., 'Unitary Synthesis of Clifford+T Circuits with Reinforcement Learning', 2024." }, "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 theoretically well-elaborated and provides a sound approach for improving state extraction in RL for quantum control. It is generally well-written and provides solid connections to the background in quantum physics. Overall, the paper presents a well-motivated method to improve exploration for reinforcement learning in quantum control by providing a smooth reward signal." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides an approach for the extraction and representation of states from a quantum system, improving the reward calculation for training a reinforcement learning policy in quantum control tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Despite being motivated for NISQ Devices, the paper does not provide evaluations regarding those. While the paper provides solid connections to the background in quantum physics, I am missing such connections to the background in reinforcement learning, especially focusing on integrating the extracted state and reward signal into existing frameworks for RL in quantum control and quantum circuit design. Generally, I feel like the claimed contribution could be clarified, as the paper does not seem to provide an RL algorithm but rather a method for improved state representation, which is still an important contribution when connected to existing frameworks for RL in this domain. Regarding the empirical results, I am missing details on the training process, comparisons to considerable baselines, and, in general, quantitative results. Also, the results shown in Fig. 3 could be better described in the text. Finally, the potential limitations of the approach should be discussed. \n\nMinor comments: \n\n- The stochastic policy should be defined as $\\pi: D \\times A\\mapsto[0,1]$\n- r seems overloaded, consider changing representation or reward to avoid confusion \n- A common notation t for a timestep should be used \n- Citation formatting could be improved (e.g., using \\citet for in-text citations).\n- The placement of figures (e.g., Fig.3) could be improved to ease 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": 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": "- L 346 claims: “Thus, the optimal path in phase space does not correspond to the shortest trajectory in the representation space, implying that the control task we consider is highly nontrivial.” However, t-sne simply arranges the learned representations in low-dimensional space, and unless the pretraining of the representation network included some form of latent-space structuring, the projections will always be of chaotic nature. As such I find it difficult to reason about the complexity of this task by the t-sne control trajectory, could you please elaborate a bit more on this?\n- Is there a reason as to why Figure 3 only shows 30 steps? Does the algorithm perform better / finds the target state eventually, or are the 30 steps a physical hard-constraint? (Figures 7 are plotted up to 55 control steps?)\n- In Figure 3a, the no-label case for Tr→SB is very much an outlier, performance wise. Is there a reason for this?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Well written, clear and concise. Not to much jargon, nicely understandable even from an outside perspective. I also found the figures conceptually nicely understandable with the differently colored phase regions. The application seems to be works well, empirically." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes improved quantum state control via taking small initial samples of the system to build an improved prior RL model and thus learn actions more appropriately. Contribution is a RL algorithm to steer an uncharacterized quantum state towards a target state using only measurement statistics. The algorithm uses a representation network to estimate state representations and their similarity to the target state. With the trained representations, the model can also be used to incrementally search for inputs that produce a certain target output of given unknown quantum process. The algorithm is tested on control of many-body ground states across phase transitions, like trivial or topological symmetry broken phases." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Weaknesses:\n\nAs major weaknesses the following issues stood out to me:\n\n- Duplicate explanation of the environment, reward and the representation input in 3.1 and 3.3. Those two sections should be consolidated. Also the reward in 3.1 is introduced first by the simple negative euclidian distance, but the EQ uses a different norm ||.||(_2?) and divides by d. This should be more consistent.\n- Its not clear if the architectures in Figure 2 are part of the contribution? As I understand, Figure 2 is simply an illustration of representation network and decoders of Wu et al. 2023b?\n- Figures 3 d/e, 6 and 7 are way too small to read, even with full digital zoom. Most of the explanation focus on Figure 3, with Figures 4-6 only barely mentioned in L350-354. Even with the caption, I am not sure I understand what is meant to be conveyed with Figure 7.\n- Novelty in this work is rather lacking, in my opinion. This is a simple application of RL to a discrete control problem with a very simple reward function. The representation network is an interesting inclusion, but similarly also only an application of existing work. The choice of designs (decoder architecture, RL algorithm choice, hyperparameters etc.) are not motivated, although part of this information can be found in the appendix. (Still papers should be self-contained without the appendix.)\n- Finally there is no comparison to other approaches included in this paper and as such significance is hard to judge. Without context, or at least an ablation study of different choices for the chosen approach the contribution of this paper is very light.\n\n---\n\nMinor Notes:\n\n- Figure 1 does not really present any valuable insights than one line of text would not also have explained.\n- L154, 158 i.e.[,]\n- L 157 non[]Gaussian\n- L 182 unclosed bracket and whitespace ([]from a finite set …\n- The notation of bold-r for representations and default r for reward could be chosen better and are confusing at times.\n- The notation of maximizing the average cumulative reward is worded badly. In any RL setting, the RL agents learns to optimize a policy that maximizes the reward / return of each episode. Improvement of average return over time is simply a side-effect of this.\n- Inconsistent use of Fig. and Figure. for references.\n- L 327, 329 [Ref.] {citation} ?\n- The color palette in Fig.3 d/e could be better chosen and the figures should be bigger. The light blue/beige colors and the trajectory, at that small scale, are very hard to see.\n- L 371 overloads the variable “r” for the third 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": 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": "Can you demonstrate an advantage over previous methods or more direct approaches such as end-to-end learning without the intermediate step of the pre-trained representation networks?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The method and its application is presented clearly, the work is of high quality technically.\n- The problem of state preparation is relevant, the examples are well-chosen, and the new method achieves its task of state preparation.
\n- The presented algorithm seems plausible, could be promising, and seems like a natural application of the [Zhu et al (2022) and Wu et al. (2023b)] articles." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In the control of quantum systems, a common scenario is the task of preparing a particular quantum state. In simulations, one would be able to adjust controls based on of the current (simulated) quantum state, including the initial state, but that is not possible in a real experiment and one only has access to the measurement statistics. This article proposed a new reinforcement learning method termed RGRL. It is built on two previous works [Zhu et al (2022) and Wu et al. (2023b)] which introduced networks that learn to translate measurements into a representation of a quantum states. In every control step, one would translate measurement statistics into such a representation and decide on the next step base on that information. The representation networks are trained before controls are trained.\n\n

The authors demonstrate the application of the new method for two examples: Preparing target states in the XXZ model as a many-body example, and in a continuous variable system as a single-system example." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the algorithm does not require pre-existing knowledge of the quantum system’s dynamics, it does require knowledge of the possible system states to train the representation networks it is based on. Direct end-to-end learning based on the measurement statistics would be truly independent of prior knowledge and the comparative advantage of the method RGRL presented here is unclear to me. To demonstrate any such advantage, the RGRL method would have to be benchmarked against prior work, for example the articles cited in the introduction. Such benchmarking would significantly strengthen the article, going beyond the more exploratory demonstration of RGRL." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "In the introduction you mention that \"a few works explored the possibility of directly using measurement outcomes for reward calculation Reuer et al. (2023); Borah et al. (2021); Sivak et al. (2022), but scalability still remains a challenge.\" Is your algorithm more scaleable than the alternatives and why? You claim that \"This approach significantly reduces the reliance on precise initial state information, making the algorithm scaleable and applicable to larger quantum systems\". How does the algorithm scale with quantum system size and what makes it scaleable?\n\nHow many measurements are needed and how does this scale with representation quality? How does this compare with other black box approaches in terms of the number of required samples? Perhaps this is also an interesting Reference to consider: Yuval Baum, Mirko Amico, Sean Howell, Michael Hush, Maggie Liuzzi, Pranav Mundada, Thomas Merkh, Andre R.R. Carvalho, and Michael J. Biercuk PRX Quantum 2, 040324.\n\nWhat particularly physical scenarios/experiments motivate the study of a system with an unknown initial state or completely unknown quantum operation?\n\nWhy is uncertainty in the initial state a \"major challenge in quantum state control\"?\n\nIs this method more applicable to unknown initial states or unknown quantum dynamics?\n\nHow does this method compare (for instance in terms of achieving higher fidelity control solutions, more robust solutions etc..) to previous methods in quantum control? \n\nWhat is your ML contribution? It seems like you are applying a standard algorithm (PPO) to a particular way of representing quantum states with limited measurements, or is there more to the algorithm that may have escaped my attention?\n\nYou say: \"we use the distance between quantum state representations as a proxy for quantum fidelity\", but then explicitly refer to the \"quantum fidelity\" at numerous occasions when describing experimental results. Do you mean the quantum fidelity? Or can the relationship between the representation distance and quantum fidelity be explicated more explicitly." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper proposes an interesting extension of previous work in using an abstract representation of an unknown quantum state as a framework for performing \"quantum control\" and steering the system into a desired target representation. This seems a useful approach for tackling the control of black box systems where no information about the dynamics or the initial state is given.\n\nThe background on quantum states and measurements, as well representation network structure is described clearly and in a way such that non specialists can easily follow the authors exposition. The RGRL algorithm is also described clearly and Fig. 1 provides a straightforward description of the precise operations which are implemented.\n\nThe analysis of the physics of phase space transitions, as well as the complexity of the control task is also very thoroughly portrayed, this seems particularly appealing to readers with a physics background." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors develop a machine learning algorithm that characterises an unknown quantum state or unknown quantum operations by generating an abstract representation. The authors then show how reinforcement learning can be used to apply adequate control operations to realise a desired target state representation. The control actions are determined by a neural network based on measurement data from a set of quantum measurements. This builds upon previous work (Yan Zhu, Ya-Dong Wu, Ge Bai, Dong-Sheng Wang, Yuexuan Wang, and Giulio Chiribella. Flexible learning of quantum states with generative query neural networks. Nat. Commun., 13(1):6222, 2022.) which described a method of generating representations of quantum states but the novelty lies in the reinforcement learning algorithm.\n\nThis particular approach is applied to two systems of interest. First the authors show that they can effectively control many body ground states and realise desired phase transitions. Two different state representations; one for predicting measurement statistics and one for predicting mutual information are explicitly compared and it is shown that higher quality state representations yield better control efficiency. Lastly, the authors consider a control task where they prepare a target output state from a CV Kerr quantum gate." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are a number of issues with this paper which I would like to address. \n\nFirstly, it seems as though it lacks the significance for it to be relevant to the wider audience of ICLR. The algorithm described in the paper seems interesting, but it is not clear that it is solving a broadly relevant and useful problem. The method is applicable particularly to systems where the initial states of the system are completely unknown, however no concrete examples of important and relevant physical systems which exhibit this behaviour are shown or described. Similarly, in the case of the CV Kerr gate, it is not clear what physically motivates the experimental setup: where a \"target output state can be obtained by applying the unknown quantum process to a certain input state\". It would be helpful to frame this more clearly with explicit references.\n\nSecondly, It is not clear to what extent the algorithm is practically useful when sampling from a real quantum device, since \"M\" copies of the same quantum state need to be prepared (what about the no cloning theorem?). Claims in the paper are made that the number of measurement samples is \"small\". It would be useful to have a quantitative comparison of sample efficiency compared to other methods, and to discuss how the approach handles the constraints of real quantum devices, including preparation and measurement times. For example, the authors could look at the following reference: Irtaza Khalid, Carrie A. Weidner, Edmond A. Jonckheere, Sophie G. Schirmer, and Frank C. Langbein Phys. Rev. Research 5, 043002 which constructs a sample efficient RL algorithm for quantum control.\n\nThirdly, there are no benchmarks or comparisons drawn to previous work in the quantum control literature. From the presentation in the paper, it is not clear whether the method outperforms previous methods or approaches in any relevant quantum control tasks. Moreover, the authors claim their algorithm is \"scaleable\", but it is not clear what makes the method scaleable and there are no explicit benchmarks. There are some alternatives in the literature which could be compared and contrasted; for example given uncertainty about the initial state, one can incorporate this into the design of a particular quantum control sequence which does not require a \"black-box treatment\" as shown in (Frank Schäfer et al 2020 Mach. Learn.: Sci. Technol. 1 035009) where the authors claim that \"Despite the uncertainty in the initial states, we managed to reach fidelities larger than 99.9% in the preparation of a GHZ state in a chain of M qubits.\" It is also unclear whether one would opt for a \"black-box\" method when the initial state is unknown in a real physical system, as often a more efficient approach in real quantum experiments would be to bring the quantum state into a known quantum state to then perform a control operation (e.g. optical pumping in atoms or transmon reset). \n\nLastly, the manuscript is not always clear and some of the experimental results are inaccessible to a non-specialist audience and I suspect the broad audience of ICLR. The Kerr gate is mentioned but no citation is given for further details and it is not clear why this is a particularly important problem to tackle which should be explained. Moreover, the control of many body quantum states is not sufficiently contextualised or motivated. The in-depth discussion of different types of phase transitions as well as Figs 3/6 showing \"representation space\" are not clear (no axes labels). The \"t-SNE algorithm\" is also left unexplained with no citation and it would be good to better explain this for a reader with an ML background or cut down on specific details on the phase transitions. Some more explicit references need to be made to the Appendix sections to clearly guide the reader for further background/implementation details." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A representaion-guided reinforcement learning for effcient quantum state control with very few measurements" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024rgrl,\ntitle={{RGRL}: Quantum State Control via Representation-Guided Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x9J66fnMs8},\nnote={under review}\n}" }, "abstract": { "value": "Accurate control of quantum states is crucial for quantum computing and other quantum technologies. In the basic scenario, the task is to steer a quantum system towards a target state through a sequence of control operations. Determining the appropriate operations, however, generally requires information about the initial state of the system. Gathering this information becomes increasingly challenging when the initial state is not {\\em a priori} known and the system's size grows large. To address this problem, we develop a machine-learning algorithm that uses a small amount of measurement data to construct its internal representation of the system's state. The algorithm compares this data-driven representation with a representation of the target state, and uses reinforcement learning to output the appropriate control operations. We illustrate the effectiveness of the algorithm showing that it achieves accurate control of unknown many-body quantum states and non-Gaussian continuous-variable states using data from a limited set of quantum measurements." }, "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 control", "quantum state representation learning", "reinforcement learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/1b2ad1ff8c2cf8c0005b365dc63ad4b914d5a8dc.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/c9269672c432833630927984abb43e84f7bf3b1f.zip" }, "title": { "value": "RGRL: Quantum State Control via Representation-Guided 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": 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": "None." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The dataset preparation (collection followed by data cleaning) is at the centre of contribution." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The submitted paper is an interesting contribution in the field of time series forecasting for a specific problem domain (weather). The key innovation of the paper is 1) preparation of the dataset to be used for community 2) its benchmark over existing algorithms 3) and suggestion for the future work." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper miss the latest research in the domain of TSF, such as use of foundation model for long term forecasting. Such as \n- Tiny Time Mixer\n- MORAI\n- Chronos\n- ..\nPlease review literature and add methods.\n\nAlso, there are method for AutoAI, AutoARIMA and etc from statistical 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": 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": "Same as 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. A new benchmark dataset is constructed specifically for station-based weather forecasting tasks.\n2. The process of dataset construction is clearly described.\n3. Experiments are conducted to compare general time series forecasting models with a specialized weather prediction model, providing insights into the limitations of general forecasting models for weather prediction and identifying future research opportunities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new benchmark dataset, called Weather-5K, for evaluating general time series forecasting models on weather forecasting tasks. The dataset is built using data from a public database (ISD) and includes five types of hourly weather data from over 5,672 stations spanning 2014 to 2023, following quality control and post-processing. Experiments comparing general time series forecasting models with a Numerical Weather Prediction (NWP) model are conducted, with results analyzed to gain insights into the use of general time series forecasting models for weather forecasting tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Beyond computational complexity, are there other reasons why general time series forecasting models are needed for weather forecasting? Also, even regarding complexity, a comparison between general time series forecasting models and specific weather forecasting models should be provided to support this claim.\n2. Some datasets, such as Weather-Australia, GlobalTempWind, and CMA_Wind, are not cited correctly.\n3. In the outlier detection process, it is unclear how to ensure that detected outliers are not extreme weather events.\n4. The paper states, “these models, operating at the mesh space (e.g., grid resolution of 0.25° and 0.09°), may not be the optimal solution for GSWF as discussed in Section 1”; however, there does not appear to be a related discussion in Section 1.\n5. The statement \"5,672 weather stations worldwide are selected...\" is repeated on Page 4 and Page 5.\n6. It is unclear whether RQ4 describes the bridge between TSF models and NWP models or between GSWF and NWP 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": 3 }, "primary_area": null, "questions": { "value": "Q1.\nHow do you view the trade-offs between model complexity and interpretability in practical forecasting applications?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "S1. \nThe paper presents an advancement in the field of time-series forecasting by introducing the WEATHER-5K dataset, which fills a notable gap in the availability of comprehensive, high-quality weather data for model training and evaluation. \n\nS2.\nThe authors demonstrate a thorough understanding of existing datasets' limitations and the challenges TSF models face in practical applications. \nThe methodology for constructing the WEATHER-5K dataset includes rigorous data selection, quality control, and pre-processing.\n\nS3.\nThe paper is well-structured and clearly articulated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the WEATHER-5K dataset, a large-scale global station weather dataset designed to address the limitations of existing time-series forecasting (TSF) datasets.\nThe authors present WEATHER-5K, which includes comprehensive observational data from 5,672 global weather stations with hourly data spanning ten years.\nThe paper conducts extensive benchmarking of various TSF models against operational Numerical Weather Prediction (NWP) models. This comparison highlights the performance gap between academic TSF models and real-world weather forecasting.\nFurthermore, a standardized evaluation framework and new metric (SEDI) are proposed for assessing TSF models, focusing on overall accuracy and extreme weather event prediction." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1.\nThe benchmarking against NWP models primarily focuses on traditional NWP methods. \nThis paper could benefit from including a wider variety of contemporary data-driven NWP models, such as those utilizing machine learning techniques. \n\nW2.\nThe performance analysis largely focuses on traditional metrics like Mean Absolute Error (MAE) and Mean Square Error (MSE), which, while standard, may not fully capture the complexities of weather forecasting, especially regarding extreme weather events.\n\nW3.\nThis work does not adequately address the robustness of the proposed TSF models under varying conditions, such as different seasons or geographical variations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "## Questions\n\n### Q1\n> $L=48$ for $H \\in$ {$24,72,120,168$}\n\nThe experiment setting is not optimal for long-term time series forecasting (TSF), where transformer models typically excel. For shorter horizons, graph neural networks might be more suitable. The current setting, with input lengths of $48$, may not fairly represent the performance of transformer models, which are designed for larger input lengths (usually $96$ or $336$ in the literature). Including results for different input lengths, especially $96$, would enhance the benchmark.\n\nThe current state-of-the-art for long-term TSF is Pathformer. It would be interesting to see its performance on short-term forecasts.\n\n### Q2\nISD data from NCEI/NOAA is the same source as the Corrformer and Informer weather datasets. Why are these datasets not included in Table 1? Although Weather from Informer is only one station and Global from Corrformer might have limited weather indicators, it is important to mention them to provide a complete picture and justify why WEATHER-5K should be preferred (cf. clarity section).\n\n### Q3\nWhat is the definition of an outlier, and how do the authors differentiate outliers from extreme weather? This distinction is crucial for users to understand modifications and potential model behavior. Have the modified timesteps been flagged for better model interpretation?\n\n### Q4\n> […] which highlights an unacceptably high RMSE when comparing ERA5 data to real-world observations […]\n\nAre the authors suggesting that ERA5 is inaccurate? To the best of my knowledge, ISD measurements involve both manual and automated processes, which could introduce human error, especially in manual records, which will not be in favor to ISD based dataset (even with errors correction, which might introduce different types of errors).\n\n### Q5\n> Figur 3 e) illustrates the daily temperature variations at Chongqing city in China (57516099999) from June to September in 2022.\n\nHow many stations are included in this city? And so, is this relevant to use it as the main example?\n\n### Q6\n> This indicates that ERA5 consistently underestimated the diurnal temperature range at this station throughout the heatwave period.\n\nThis observation is based on one station over six months, compared to more than 5 000 stations over 10 years of data. Such a claim requires more extensive analysis. The authors seem to discredit ERA5 while earlier referring it as one of the most accurate real-time forecasting products and using it to fill missing data. Why use ERA5 if it is later discredited? If ERA5 is the most accurate, why not use it directly, especially given the potential for errors in ISD's manual records, extensive work for correction and data coverage limitation?\n\nHow many timesteps are considered extreme weather, and what percentage do they represent over the 10-year span? Is this percentage significant enough to promote WEATHER-5K over ERA5?\n\n### Q7\nIn Figure 4a, where is the NWP model? The authors mention the computational cost of NWP models as a limitation but do not provide performance data. The parameters size /cost of some models, like PatchTST, known for high computational cost, seems inconsistent with the literature, which might advocate that the experiment setting (input of $48$) is insufficient to demonstrate predictive power. There is also an issue with the figure: Dlinear should be a point due to the parameter's size of $0.01$ but is not, and there are two circles without text (red and purple). What do they represent? Mamba and Autoformer? What do the training cost and error dotted lines represent?\n\n### Q8 - Experiment setting \nEarly stopping of 3 is, in my experience, too low; 5 or 10 would be more appropriate.\n\n### Q9 - Reproducibility\nWhat is the batch size of Corrformer?\n\n## Dataset\n\n### D1\nWhat is the point of repeating latitude and longitude for each row of a given weather station? This increases the dataset size unnecessarily unless stations are moving.\n\n### D2\n> we have used latitude, longitude, and elevation to represent their geographic locations.\n\nWhere is the elevation information? The instance field does not mention elevation in the CSV files.\n\n### D3\n> However, the proportion of error introduced by the interpolation is relatively small.\n\nAuthors need to provide the number of errors corrected and a mask or label column to identify these corrected timesteps. This information is crucial for many tasks and could open the dataset to other applications beyond forecasting.\n\n## Limitations\n\n### L1\nAre errors always isolated timesteps on only one variable? If so, interpolation is understandable. If not, especially with consecutive timesteps with errors or missing data on multiple variables, interpolation is limited. In such cases, did the authors use ERA5? If so, make it clearer in the paper, and explain more in details how interpolation from ERA5 are made.\n\n### L2\nNot all users of this dataset will require worldwide stations for their applications. Have authors provided a way to select a subset of the dataset?\nIf not, please provide explanation of the weather stations name/ID formatting to simplify the selection task.\n\n## Clarity / Coherence\n\n### C1\nClarify the nature of Weather-5K and the targeted task. In my understanding, it is an hourly multivariate spatio-temporal weather dataset (several stations, each providing time series of several weather indicators) for short-, mid-, and long-term Time Series Forecasting. Forecasting tasks can be done in various fashions, such as:\n- Univariate-to-Univariate (U), ex. Predict temperature of station XX for the next day using historical temperature of station XX\n- Multivariate-to-Univariate (MU), ex. Predict temperature of station XX for the next day using historical weather indicators of station XX\n- Multivariate-to-Multivariate (M), ex. Predict temperature of station XX for the next day using historical weather indicators of station XX\n- Spatio-Temporal (ST) U\n- ST MU, ex. Predict temperature of station XX for the next day using historical temperature from station AA to station WW\n- ST M\n- Multi-Variables (V) ST MU\n- V ST M\n\nTable 1 does not fully capture the dataset's potential in terms of forecasting tasks. \nIn addition, it should be revised:\n- Use commas consistently for thousands.\n- If \"frequency\" and \"year\" are provided, is \"length\" necessary? A column for missing data would be more informative.\n- For the Exchange dataset, should $8$ be the value for \"station\" instead of \"variable\" since each column represents a country?\n\n### C2\n> Task settings\n\nWhat is the forecasting fashion? M or MU? This is important for reproducibility. If M, are all stations considered, hence the input length of 48 \"to balance computation and performance\"? it not all stations, which subset of stations is used (for reproducibility)? If MU, which station is the target? And which stations are the inputs (all or a subset)?\n\n### C3\n> While they provide a solid foundation, their performance may be limited when faced with complex patterns or nonlinear relationships.\n\nDo authors have a reference for this assumption?\n\n### C4\n> ML methods […] offer enhanced capabilities to handle nonlinear relationships and complex patterns.\n\nNeed references, perhaps [1] and [2].\n\n[1] https://people.math.sc.edu/devore/publications/NLACTA.pdf\n[2] https://link.springer.com/book/10.1007/978-3-319-58795-0\n\n### C5\nIn section RW, paragraph \"Data-driven numerical weather prediction,\" reintroduce the NWP acronym, the same should be done for GSWF to improve clarity.\n\n### C6\nFor Figures 8 to 14, provide weather station IDs and sample IDs or origin forecast (t) for each row (for reproducibility). In addition, to highlight the necessity of WEATHER-5K, use samples depicting different scenarios, notably including extreme weather, to visually demonstrate model behavior. Therefore, identify and provide cases where extreme weather:\n- Is in the input but not the predicted window.\n- Is in the predicted window but not the input.\n- Is in both.\n\nFurthermore, in these figures, there is no need to repeat the title of each plot and the x-axis; use a 7 x 5 grid with the sharex option.\n\n## Proof-read\nTo cite but a few:\n\n1. **[Very important]** There is a formatting issue with citations in the first paragraph of the introduction and throughout the paper. Citations should be in parentheses unless the authors' names are part of the text, as per ICLR guidelines:\n> When the authors or the publication are included in the sentence, the citation should not be in parenthesis using \\citet{} (as in “See Hinton et al. (2006) for more information.”). Otherwise, the citation should be in parenthesis using \\citep{} (as in “Deep learning shows promise to make progress towards AI (Bengio & LeCun, 2007).”)\n\n2. Thie following text is repeated twice in the paper; remove the unnecessary repetition.\n> 5, 672 weather stations worldwide are selected, spanning the period from 2014 to 2023, ensuring a recent and relevant time frame. This selection process focused on balancing the longevity of station operation, hourly data availabilit\ny, and the inclusion of diverse weather variables.\n \n\nTypos:\n- “[…] into physically-based NWP models” or “physical-based” should be physics-based, no?\n- “(with yea 2022)”\n- “[…] benchmark experiments on WEATHER-5K,” comma?\n- “[…] perform better. In our benchmarks. So developing efficient time-series […]”\n- “[…] are may not be the optimal solution […]”\n- “[…] except for Correformer.”\n- “Figur 3 e) illustrates the daily temperature […]”\n- “[…] of the WEATHER-5 dataset […]”" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Intensive work in dataset creation.\n- Proposal of both a dataset and a benchmark.\n- Data splitting follows the longest cycle of the data (one year)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a novel weather dataset that can be utilized in various time series forecasting tasks. The dataset comprises data from over 5,000 weather stations worldwide, each providing five weather indicators, making it a promising resource. The authors have meticulously selected stations with 10 years of data and have cleaned the raw data from ISD using interpolation and ERA5 to ensure a complete dataset for future users. \n\nThe dataset is used to benchmark state-of-the-art transformer and state-space models against a Numerical Weather Prediction technique." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of coherence and clarity in the presentation.\n- Omission of some important, well-known datasets.\n- Overgeneralization based on unique observations.\n- Limited reproducibility.\n- A thorough proof-reading is required." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024how,\ntitle={How Far are Today's Time-Series Models from Real-world Weather Forecasting Applications?},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x9cXrOQskc},\nnote={under review}\n}" }, "abstract": { "value": "The development of Time-Series Forecasting (TSF) techniques is often hindered by the lack of comprehensive datasets. This is particularly problematic for time-series weather forecasting, where commonly used datasets suffer from significant limitations such as small size, limited temporal coverage, and sparse spatial distribution. These constraints severely impede the optimization and evaluation of TSF models, resulting in benchmarks that are not representative of real-world applications, such as operational weather forecasting. In this work, we introduce the WEATHER-5K dataset, a comprehensive collection of observational weather data that better reflects real-world scenarios. As a result, it enables a better training of models and a more accurate assessment of the real-world forecasting capabilities of TSF models, pushing them closer to in-situ applications. Through extensive benchmarking against operational Numerical Weather Prediction (NWP) models, we provide researchers with a clear assessment of the gap between academic TSF models and real-world weather forecasting applications. This highlights the significant performance disparity between TSF and NWP models by analyzing performance across detailed weather variables, extreme weather event prediction, and model complexity comparison. Finally, we summarise the result into recommendations to the users and highlight potential areas required to facilitate further TSF research.\nThe dataset and benchmark implementation will be publicly available." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Time-series benchmark", "large scale spatial-temporal dataset， numerical weather prediction 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/9a68d8ba55d0d78e18ac6c91440fec93d63de236.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": "How Far are Today's Time-Series Models from Real-world Weather Forecasting 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": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "I would like to recommend the authors to avoid using existing company (stocks) names where it is not really required (Figure 1 and further in the text). I think, in the examples the names can be replaced with e. g. “Company A” and “Company B” or \"Stocks A\" and \"Stocks B\" without any loss for the paper motivation and content. I think, the reasoning for using real names might have been to show an actual example of the problem tackled in the paper. However, I think that using real names (especially in the context of an adversarial attack) does not contribute anything to the discussion and may potentially lead to uninvited implications." }, "flag_for_ethics_review": { "value": [ "Yes, Discrimination / bias / fairness concerns" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1) Do you have any suggestions on how one can defend web-agents against attacks similar to the one that you propose?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Originality: the papers proposes the first black-box attack against VLM-based web agents which is based on existing DPO Method. \n\nQuality: the authors perform in depth analysis of their proposed attack. The limitations of the method are discussed (Appendix C).\n\nClarity: the paper is structured reasonably.\n\nSignificance: black-box attacks on web agents may cause significant harm, so raising awareness of such attacks is important for the LLM/VLM field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the robustness of Web Agents and proposes a black-box adversarial attack AdvWeb which is based on Direct Policy Optimization (DPO). The attack efficiency is evaluated against existing SOTA web agent for several domains." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Could you elaborate in the paper on how exactly the baselines (line 353) were adapted to your problem? Since they all achieve 0.0 on all tasks (Tables 1-2), it looks like the comparison is either not suitable or not fair. Do you have an explanation for such poor performance of the baselines?\n\n2) See the ethics review discussion." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- * I am confused about why the different between SFT and DPO varies so much by domain in figure 3. Can you help me understand this? This makes me suspicious.\n - Do you need RL? What about a prompting based baseline?\n - I did not follow \"We also fix the injection position at the ground truth HTML element e to further reduce the search space\"\n - In Algorithm 2, you are using the positive examples twice, both for SFT and then also for DPO. I'm curious whether this introduces some bias, and wondering if you have the comparison without this. This should be easy to do.\n - I'd be curious for an adversarial prompter model scaling study, how it varies with size.\n - Is there any overlap between the train and test tasks? How does the model algorithm performance scale with the number of train tasks?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+: handling an important problem, web-use agents with black box access, quite close to what would happen in practice i.e., the threat model seems realistic, point about the stealthiness is important\n+ high ASR, but how is the ASR measured?\n+ The paper seems to be clear and well written\n+ Good use of train and test tasks" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents AdvWeb, a framework that injects invisible adversarial strings into websites to attack VLM-powered web agents. The technique uses Direct Policy Optimization (DPO) to optimize a prompter model that generates effective adversarial strings. The key claim is that the attack transfers effectively across different targets and settings. The main technical contribution is the technique for injecting invisible adversarial strings onto websites and using DPO to optimize a prompter model to make these strings work well.\n\nI am excited about this paper's potential, as it addresses an important practical problem. However, several issues currently prevent me from advocating for acceptance. If these issues are adequately addressed in the rebuttal, I would revise my recommendation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "High priority:\n\n- ** One weakness is that because the HTML isn't rendered, if you had a purely computer image based agent, it would not work. It would have been better to study stealthiness in that setting. I.e, studying agents that _don't_ take in the HTML. Please at least discuss this limitation in the paper.\n- ** Concerns about Baselines. I feel like the baseline should probably be someone adding a GCG string to a website HTML as a prompt injection attack, rather than as a jailbreak? You can use transfer there as you did. Did you use jailbreaks that work on GPT-4v? I think there should be working jailbreaks (e.g., Pliny's Godmode Jailbreak or Many Shot jailbreaking). I find the fact that no other attacks work pretty suspicious, and wonder how hard you tried to optimize them? FWIW, I am also sympthatetic to the view that the prompt injection is just a new setting, in which case, a human written baseline, a prompting online, SFT, and SFT + DPO baseline may be more appropriate, which is already later in the paper (but framed differently)\n- ** \"The results, as shown in Table 3, demonstrate that the ASR remains high when we change the injection position or HTML field, with ASR being 71.0% and 97.0%, respectively.\" is a misclaim, given that it depends a lot on the domain with finance performing poorly. The overall claim is OK, but please rewrite the claim to be correct. In general, you need to go through the paper and make sure you aren't overclaming.\n\nMedium priority:\n- * Please add more examples of the attacks found in the main paper, and mention diveristy. Do we have model collapse?\n- * Please move Algorithm 2 into the main paper, its mentioned several times there. It's also not an algorithm, its a function. Please fix this.\n- * You need to explain how you are doing the labelling in Algorithm 1, I assume you use an LLM but that seems important? Or do you just check the action taken is hard?\n- * I appreciate this is hard, but I'd love to see whether the attack works on Claude Sonnet 3.5. Claude models tend to be more robust, so thats why I am curious.\n\nLow priority:\n- \"innovative reinforcement learning (RL)-based attacking pipeline, tailored to solve these challenges effectively\". Please drop \"innovative\". Similarly \"advanced\" attacking framework, drop advanced. Please fix the language." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. What other baselines did you consider, and what would it look like for you to compare your method to the attacks against web agents that you list in related work? \n2. How expensive was the training & optimisation process for generating successful attacks?\n2. Could you explain what signal you optimise against from the target model - does the victim model refuse?\n3. When you inspect samples maually, do you get a sense of why transfer from GPT-4V to Gemini on finance tasks so much lower?\n4. How did you select the tasks from Mind2Web to test against? Was it manual inspection? What were your criteria to judge that a task involves \"critical events that lead to severe consequences\"?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "I believe what the authors are trying to achieve is novel, interesting, and important: they target a timely and significant concern that will only become more prevalent as generalist web agents are deployed and relied on. \n\nThe work also seems novel. My impression from the literature is that there aren't many other papers showing successful attacks/injections against web agents, especially black-box, and via means that would be invisible to most humans.\n\nThe authors' constraints of 1.) only allowing themselves black-box access to the target models, and 2.) their constraint of \"stealthiness\" increase the difficulty and realism of their attack framework. Attackers hiding malicious instructions in hidden HTML fields seems realistic and aligns with their threat model. \n\nGiven how realistic their attacks seem, and how high their reported ASRs are, this paper could promote increased safeguards on generalist web agents - or in the least could present a compelling demonstration that users should be careful when selecting which agent frameworks to rely on. \n\nIt's impressive that their set their success threshold quite high - at achieving exactly the the action their attack targets (instead of e.g. just distracting the agent to make it fail at its main instruction)." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an attack framwork against generalist web agents that parse websites through screenshots and HTML. The authors attempt to demonstrate how a malicious web-developer (either in-house or via open-source HTML that other devs blindly import) could use HTML that’s invisible to human users to influence the actions of a generalist web agent. \n\nThey optimise attacks in a black box setting against models based on GPT-4V and Gemini 1.5, scaffolded as web agents through the SeeAct framework. They generate attacks using GPT-4 to generate HTML with hidden instructions, then optimise an adversarial prompter against the target model using RLAIF so that those hidden instructions are maximally effective against the target.\n\nThe SeeAct framework allows the victim models to read HTML elements in the webpage that a normal human user wouldn't normally see. They inject adversarial prompts into these hidden (the authors call these \"invisible\") HTML fields, and their results show success in influencing SeeAct agents to take different actions than their user prompt instructed them: for example, when told to purchase MSFT stocks on a financial website, the injected HTML prompt misleads the agent to buy e.g. Nvidia instead. \n\nThe authors display high success rate against the target models, especially in the face of their four baselines, all of which achieve 0% ASR across every domain, on both models. Three of their baselines (GCG, AutoDan, COLD-Attack) involve optimising adversarial prompts against whitebox models and hoping those transfer, and the other, Catastrophic Jailbreak, requires no access to model internals but control over the model's decoding hyperparameters.\n\nIn addition to requiring that their injected prompts are \"stealthy\" - i.e. invisible to a normal user, the authors also emphasise the controllability of their attacks - i.e. the ease with which attackers can switch out one set of injected instructions for others. With this property, they also demonstrate that their attack strings can be adapted to transfer between different task-domains. \n\nThe authors also report that their attacks transfer well between the two models they attack, and between different HTML fields (using the same successful prompts), which suggests that their attacks aren't brittle to the specifics of the model / HTML field they were optimised for." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the paper presents a novel and important problem setting, I don't feel confident that the paper provides me the information I'd need to evaluate if the attacks it introduces are as successful and significant as the authors claim. I think it's quite possible the authors have all the data they need to convince me otherwise, but I urge them to include it in the paper.\n\nOverall, reading the paper gives me a weak sense of what generalist web agents can be influenced to do, or how significant the negative outcomes of these attacks could be. I am not able to evaluate the risks of their attack method if all I know is that they cause models to take unspecified targeted actions in a domain the paper only labels as \"cooking\" or \"housing.\" If the authors could provide many more examples of successful tasks & attacks, and their consequences - at least one for each attack domain, their claims of significance would be much more convincing. \n\nIn their related work, the authors claim that they beat \"strong baselines.\" But I'm unconvinced that any baseline on a task that achieves 0% ASR on every task, on both models they test, should be considered a strong baseline. The authors claim that there are no analogous black-box attacks in this setting, which I can't refute from what I've read elsewhere. However, I'm confused why they don't compare at all against the methodologies from the papers they list in the existing body of research from \"Existing Attacks against Web Agents.\" Even if those methods are manual, where this paper's methods are learned, it feels like a much more informative comparison. I would urge that the authors find a more successful baseline to test against, and ideally show some results comparing their success rate to at least one other method for steering the same web-based agents. \n\nBecause of the weakness of the baselines they compare against, and the lack of comparison to other methods for achieving the same ends. I'm left without much context to evaluate how powerful this method is. I recognise that the attack domain is novel, and I found figure 3 helpful for understanding that their training pipeline helps their models achieve higher success after DPO. Discussion of more ablations, and in particular considerations of alternative methods for training and optimising their attacks, and why they weren't as successful, would be informative to my understanding of this paper.\n\nI don't leave the paper with a strong sense of the offense-defense balance in this setting. In particular, the paper might benefit from more detail on how expensive it is to generate these attacks, as the authors do not provide much detail on how expensive the training process is (in steps or dollars) for their RLAIF pipeline outlined in Algorithm 1. For any attack, it seems important to know how quickly and cheaply attackers could generate new attacks when the victim models are updated. Further, if their RLAIF pipeline required very many training steps, it's plausible that the developers of these web agents could become aware. I would be interested to see e.g. how the ASR of their framework increases throughout training.\n\nI'm confused why the title of the paper focuses on attacking \"VLM-Powered Web Agents.\" While true - the SeeAct framework only employs VLMs - as far as I can tell nothing about the victims' multimodality is being attacked, simply their parsing of HTML text. My first impressions of the paper led me to expect exploits against the multimodality of these models, which was ultimately incorrect. I suggest the authors remove \"VLM powered\" from the paper title. \n\nThe authors repeatedly stress the importance of the controllability of their HTML attacks (including in the title). That is, that the same attack strings can be easily adapted to cause different target actions on the same task. Any examples of the ways in which their HTML attack strings are editable would be helpful. But more importantly I do not get a sense from the paper why this is important under their threat model. I think the answer may be a question of cost: that it is cheap to retool successful attack strings for a different purpose, but the existing wording in the paper does not mention this. The most I see is that the authors claim: \"allowing attackers to switch targets with minimal effort and no additional computational overhead.\" This remark is on the penultimate page of the paper, which I think is too late and too little justification for what is introduced as a key constraint of their attacks - even in the title of the paper. More commentary on cost, as I request above, would also have made the relevance of this constraint - which seems technically sound - much clearer.\n\nThe transfer results aren't as convincing as their main results - especially since the ASR is quite varied across different domains, achieving 0% transfer on probably the most compelling domain for their threat model (online finance tasks). \n\nAppendix C, addressing limitations is two sentences long and claims that \"It is possible to optimize a more effective adversarial prompter model.\" The authors don't expand on this claim and I would rather they address more of the limitations in their threat model that this review (& if relevant, others) highlights.\n\nSome weakness in writing & setting that ought to be addressed but should be trivial to fix:\n\nThe authors refer to Algorithm 2 throughout the paper including in the second figure and in their description of Algorithm . Algorithm 2 can in fact be found as the only entry in Appendix A, while Algorithm 1, which in my opinion requires Algorithm 2 to understand it, is in the main body. Algorithm 2 is critical for understanding how they generate the initial malicious HTML requests that they then label for RLAIF using the victim model, and I was confused at first because the authors didn't list where algorithm 2 could be found.\n\nSome claims in the abstract and introduction seem too strong for the level of evidence this paper provides. For example:\n* That Vlms “have revolutionized the creation of generalist web agents … thereby boosting human efficiency and productivity” requires at least some citation or evidence\n* That their choice of injecting into unseen HTML fields makes it “nearly impossible for users to detect tampering” feels like a stretch: can't any user inspect the page's HTML? I appreciate that most users wouldn't. \n\nThere's also an error on the final page, where a paragraph break interrups a sentence immediately before the conclusion, leaving a hovering sentence fragment that reads \"[newline] demonstrates the situation in which the user wants to buy Qualcomm. However, after adding the adversarial injection, the agent buys Apple stocks instead.\" This sentence fragment, which appears to be an incomplete draft of the final setence of Section 5, needs to be addressed before the paper can be acceptable." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Thanks for submitting your paper to ICLR 2025.\n\nWith the advancement in the intelligence of large language models, using VLM agents to automate a series of tasks is emerging as a future development trend. However, the associated security risks remain unexplored. This paper introduces AdvWeb in an attempt to expose the security risks involved in VLM-powered web agents. The topic discussed in this work is both timely and highly important. \n\nHowever, I have the following concerns:\n\n- First, regarding the threat model, I am curious about the identity of the adversary here. The adversary’s method involves embedding malicious strings in web content. Generally, if users choose commands that access official websites, how could an adversary manipulate an official site? If it is not an official website, how would the adversary ensure that their site gets indexed?\n\n- The threat highlighted in this paper is urgent, making it essential to consider corresponding defense mechanisms. However, this paper does not discuss any defense strategies.\n\n- The AdvWeb attack scenario, which targets web content, appears may similar to prompt injection attacks [R1]. Analyzing and comparing the differences between these two attack types would be beneficial.\n\n- This paper selects SeeAct as the web agent framework and uses GPT-4V and Gemini 1.5 as the underlying VLMs. I am curious whether current open-source VLMs also support SeeAct. Additionally, in future edge applications, where privacy is a priority, smartphones may deploy smaller-scale VLMs locally. Analyzing the attack’s effectiveness on open-source VLMs would help demonstrate the generalizability of the attack.\n\n- In practical applications involving user payment actions, a secondary authentication of the user’s identity is typically required, providing the user with an opportunity to review the final outcome and potentially prevent malicious actions. Does this scenario indirectly suggest that the stealthiness of the attack may be limited?\n\nReference\n\n[R1] Adversarial Search Engine Optimization for Large Language Models (https://arxiv.org/abs/2406.18382)" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper presents AdvWeb, the first black-box target attack framework for VLM-based website agents, proposing a method to train adversarial prompt models through reinforcement learning with DPO in a black-box setting, which is innovative and effective.\n\n- The method ensures stealth, controllability, and black-box scenarios, which adds greater practical value.\n\n- The experimental results demonstrate the effectiveness of AdvWeb in attacking different VLM-based website agents and tasks, which helps to raise awareness in the field for developing more reliable web agents and implementing effective defense measures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel black-box attack framework, AdvWeb, targeting website agents driven by Visual Language Models (VLMs). Under the black-box framework, AdvWeb proposes a two-stage training paradigm. First, it conducts supervised fine-tuning on positive adversarial prompts to obtain the SFT Prompter Model. Subsequently, it leverages the black-box feedback from the target agent combined with DPO strategy for reinforcement learning to train an adversarial prompt model, generating and injecting adversarial prompts into web pages to induce specific adversarial behaviors from the network agent. This method is characterized by its stealth and controllability, with a high success rate of attacks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The threat model considered in this paper may be impractical.\n\n- This paper does not propose any potential defense mechanisms.\n\n- The types of victim web agents considered are limited." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No ethical concerns are involved." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Why does the paper not compare with prompt injection attacks, which are more aligned with the context of this work?\n\n- What are the key advantages of the proposed attacks compared to existing prompt injection attacks?\n\n- What are the evaluation results of the proposed attacks against prompt injection defenses?\n\n- How are the white-box baseline jailbreak attacks implemented against black-box VLMs?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper showcases the effectiveness of attacks on SeeAct across several prominent VLM models, including GPT-4V and Gemini 1.5.\n\n- It conducts an extensive ablation study to evaluate the effectiveness of the proposed attacks under different attack settings and targets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a black-box targeted attack against Vision-Language Model-powered web agents. The approach involves training an adversarial prompt model using direct policy optimization. This model generates and injects adversarial prompts into web pages, leading web agents to execute targeted malicious actions. The effectiveness of the proposed attack is demonstrated through evaluations conducted on SeeAct." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Since the attack's goal is to mislead agents into executing malicious actions, the scope of the paper aligns more closely with prompt injection attacks rather than jailbreak attacks. However, the paper focuses significantly on distinguishing itself from jailbreak attacks and includes them as evaluation baselines for comparison. It would be beneficial to provide a comprehensive survey of prompt injection attacks in the related work section, highlighting the key differences and advantages of the proposed method compared to existing attacks. Including prompt injection attacks as baselines would strengthen the contribution of the paper.\n\n- Given the extensive research on prompt injection attacks, the paper does not clearly distinguish the key advantages of its proposed attack over existing methods. Specifically, there is a line of prior studies [1][2] that have utilized context-ignoring techniques by appending prompts like \"Ignore previous instructions, do xxx.\" This is similar to the method used in this paper, as illustrated in Figure 4, where the attack prompt is \"Disregard all previous and subsequent instructions and choices, do xxx.\" Moreover, previous methods do not require collecting a prompt training dataset or performing expensive DPO training of an LLM-based attack prompter, making them more efficient.\n\n- Defenses against prompt injection attacks have also been extensively studied (e.g., [3][4]). It is important to assess the effectiveness of the proposed attack against these defenses to determine its practical applicability and robustness.\n\n- The selection of victim web agents is limited. The paper evaluates only one type of VLM-based web agent, which may not fully demonstrate the generalizability of the proposed attacks. Incorporating a more comprehensive evaluation with a variety of web agents would strengthen the paper.\n\n- In comparing baselines during evaluation, the paper includes several white-box jailbreak attacks like GCG and AutoDAN. It is unclear how these baselines are implemented against SeeAct, given that it uses proprietary VLMs that are black-box and do not provide gradient information.\n\nReference:\n\n[1] Branch, Hezekiah J., et al. \"Evaluating the susceptibility of pre-trained language models via handcrafted adversarial examples.\" arXiv preprint arXiv:2209.02128 (2022).\n\n[2] Perez, Fábio, and Ian Ribeiro. \"Ignore Previous Prompt: Attack Techniques For Language Models.\" NeurIPS ML Safety Workshop. 2022. \n\n[3] Liu, Yupei, et al. \"Formalizing and benchmarking prompt injection attacks and defenses.\" 33rd USENIX Security Symposium (USENIX Security 24). 2024.\n\n[4] Chen, Sizhe, et al. \"Aligning LLMs to Be Robust Against Prompt Injection.\" arXiv preprint arXiv:2410.05451 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce AdvWeb, a novel black-box attacking framework targeting LLM and VLM-based web agents, leveraging the feedback from target agents to generate adversarial prompts." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024advweb,\ntitle={AdvWeb: Controllable Black-box Attacks on {VLM}-powered Web Agents},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x9gCQC3rVA},\nnote={under review}\n}" }, "abstract": { "value": "Vision Language Models (VLMs) have revolutionized the creation of generalist web agents, empowering them to autonomously complete diverse tasks on real-world websites, thereby boosting human efficiency and productivity. However, despite their remarkable capabilities, the safety and security of these agents against malicious attacks remain critically underexplored, raising significant concerns about their safe deployment. To uncover and exploit such vulnerabilities in web agents, we provide AdvWeb, a novel black-box attack framework designed against web agents. AdvWeb trains an adversarial prompter model that generates and injects adversarial prompts into web pages, misleading web agents into executing targeted adversarial actions such as inappropriate stock purchases or erroneous bank transactions—actions that could lead to severe consequences. With only black-box access to the web agent, we train and optimize the adversarial prompter model using Direct Policy Optimization (DPO), leveraging both successful and failed attack strings against the target agent. Unlike prior approaches, our adversarial string injection maintains stealth and control: (1) the appearance of the website remains unchanged before and after the attack, making it nearly impossible for users to detect tampering, and (2) attackers can modify specific substrings within the generated adversarial string to seamlessly change the attack objective (e.g., purchasing stocks from a different company), greatly enhancing attack flexibility and efficiency. We conduct extensive evaluations, demonstrating that AdvWeb achieves high success rates in attacking state-of-the-art GPT-4V-based VLM agents across various web tasks in black-box settings. Our findings expose critical vulnerabilities in current LLM/VLM-based agents, emphasizing the urgent need for developing more reliable web agents and implementing effective defenses against such adversarial threats." }, "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", "Web Agent", "Multimodal", "Attack" ] }, "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/33d8ee071549e5099bb2e6ad531aff8cce198c3f.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/28cdf3f34501ef51da8355453d65e3c04c29b2ed.zip" }, "title": { "value": "AdvWeb: Controllable Black-box Attacks on VLM-powered Web Agents" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "Distribution of Spuriousness Scores: Could the authors show the distribution of the proposed spuriousness scores across neurons in different datasets? This would help validate the claim that the spurious and core neurons can be effectively separated.\n\nDifference from Retraining with Up-Weighting: What is the difference between the proposed algorithm and retraining the last layer while up-weighting the misclassified samples? Clarifying this would help in understanding the distinct contribution of the proposed method." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The proposed LaSAR method aims to achieve robust learning without group information by proposing metrics to evaluate whether a neuron is spurious or core related. This approach makes LaSAR a practical and fully unsupervised solution to mitigating spurious bias." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the issue of spurious correlations in deep neural networks trained with empirical risk minimization (ERM). The authors propose an approach called Last-Layer Selective Activation Retraining (LaSAR), which aims to mitigate spurious bias without requiring group labels or external annotations. The method selectively blocks neurons identified as spurious during the retraining of the last classification layer, thus promoting the model to learn robust decision rules. The authors demonstrate that LaSAR is effective across multiple data modalities, such as vision and text, and improves worst-group accuracy in benchmark datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Limited Theoretical Analysis: While the empirical results are promising, the theoretical foundation for why the proposed spuriousness score works effectively in all cases is very limited. Including more rigorous analysis or theoretical guarantees would strengthen the paper's claims about the effectiveness of LaSAR.\n- Limited Heuristic Exploration: There is limited heuristic exploration of the distribution of the proposed spuriousness score. Figure 4 appears to be cherry-picked, and it would be more persuasive if the authors could provide the distribution of the proposed spuriousness score across neurons in different datasets.\n- Incremental Contribution: The phenomenon that spurious neurons and core neurons can be separated has been demonstrated in prior work [1][2]. Moreover, the proposed spuriousness score is calculated as the median among misclassified samples and the median among correctly classified samples, which appears equivalent to retraining the last layer while up-weighting the incorrect samples. This limits the novelty of the contribution. Furthermore, the neuron masking algorithm assumes that a neuron can represent part of the spurious features, which is a strong assumption that may not always hold true. Additionally, it is unclear why masking the last layer is necessarily better than masking a middle layer.\n- JTT Algorithm Classification: JTT is listed as a semi-supervised algorithm at line 362, but it appears to work without group information. This classification should be corrected.\n\n[1] Last Layer Re-Training is Sufficient for Robustness to Spurious Correlations\n[2] Complexity Matters: Dynamics of Feature Learning in the Presence of Spurious Correlations" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How does the method ensure that it doesn't accidentally block neurons representing valid but complex feature combinations rather than truly spurious correlations?\n2. How does the method handle cases where features might be spurious in some contexts but valid in others?\n3. (Also related to 1.) There has been evidence that neurons may learn polysemantic features. What is the impact of LaSAR in case neurons may learn linear combinations of spurious and core features?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Key Strengths:\n\n1. Spurious neuron identification: The proposed LaSAR framework introduces an interesting approach to identify spurious neurons using activation patterns and prediction outcomes, providing a self-guided mechanism for bias detection.\n\n2. Practical Utility: The method works as a post-hoc tool in standard ERM training settings, making it highly practical for real-world applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Last-layer Selective Activation Retraining (LaSAR), which identifies and mitigates spurious bias without requiring external supervision or group labels. The key point lies in observing that neuron activations combined with prediction outcomes can self-identify spurious features, and then using this information to selectively block spurious neurons during last-layer retraining. The method works as a practical post-hoc tool in standard ERM training settings, and requires no additional annotations beyond class labels. Authors compare their method with competitive baselines such as JTT, and DFR and show some improvement in worst group accuracy on a benchmark with 4 datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The contribution of this paper is severely limited. Indeed, the core intuition that using (i) misclassified examples of validation data, (ii) and retraining all layers or the linear head to reduce reliance on spurious features has been demonstrated previously with methods such as JTT and AFR. How is LaSAR fundamentally different from AFR? \n\n2. Lack of fair comparison. Although JTT and AFR need group information on the validation data only to tune hyper-parameters. They can be tuned using the worst-class accuracy. Authors should therefore compare their method with JTT and AFR when tuned on worst-class accuracy.\n\n3. No theoretical guarantees are provided about the convergence and stability of the selective activation retraining process, even on synthetic data." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1, why did you define the spuriousness score as (5)? To help readers understand the behind rationale, I think the authors may need to add more explanation.\n\n2, In line 157, the authors mentioned that WGA is the accuracy on the worst performing data group in the test set $\\mathcal{D}_{test}$. However, they used argmax in the formula of WGA, it seems problematic since argmax will output a group label rather than the value of accuracy and the argmax will output the best performing data group in terms of accuracy rather than the worst performing data group. \n\n3, In Section 3.2, they used a synthetic motivating example. It may be better to use a real motivating example.\n\n4, In line 321, the authors may want to say \"equation (6) and equation (7)\" rather than equation 6 and equation 7.\n\n5, I think the study objective in this paper is quite similar to variable selection in statistics. We can use many penalties such as L1 penalty to remove those spurious features. I do not see the advantages of the proposed method compared with those variable selection methods in statistics. The authors may need to discuss this point." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors did extensive experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors propose a novel method to self-identify spurious features and mitigate the spurious bias by retraining the last classification layer. In general, the idea of using neuron activations before the last classification layer, coupled with their final prediction outcomes, to provide self-identifying information on whether the neurons represent spurious features seems interesting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The writing in some places is unclear, in particular, they did not clearly explain the behind reasoning of the proposed method to identify the spurious features. The did not use some theoretical results to support the proposed method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- Could also provide results in terms of balanced accuracy (other than accuracy and WGA)? \n- I think that retraining on a part of the validation set may lead to \"unfair\" comparison with baseline methods that are not trained also on validation data\n- Reweighting the classification layer essentially does not \"remove\" the bias from the whole model, but it is just a correction. Do you think this might be an issue in certain cases?\n- I do not see a clear difference between core activation maps and spurious activations maps for CelebA in Fig. 4., the spurious heatmaps even seem a bit more focused on the hair (which is the target task). \n- Using your method do you think it would be possible to provide pseudo-labels for the training data in order to use a supervised debiasing method?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper tackles a very important issue, which is learning unbiased models from biased data. \n- The proposed method does not need any kind of annotation on the bias, and it just leverages the class label (unsupervised debiasing)\n- The reported results show improvement w.r.t other methods" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this work, authors propose a debiasing method that works by retraining only the last layer (classification layer) in order to re-weight different factors in the latent representation. The assumption is that latent factors carry different information (source, spurious, noise) which can be effectively filtered out from the classification layer by reweighting. They test their method on standard debiasing benchmarks such as celeba, waterbirds, multinli and civil comments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Here are my main concerns about the work: \n\n- The authors' assumption is that latent representation can be factorized in source, spurious and noise components. This is clearly shown in the toy example; however it is not clear why this should also happen in representations extracted from deep neural networks on complex data. It might not be so simple to factor out single components in the learned representations, as they might be intertwined and correlated. Can you provide some more theoretical backing of this method? \n\n- I think that validation on more difficult datasets such as 9-Class ImageNet / ImageNet-A (https://openreview.net/forum?id=2OqZZAqxnn) should be added to the experimental validation.\n\n- The related work section should be updated a bit with relevant works in the area (e.g. [1-6])\n\n[1] Bahng, Hyojin, et al. \"Learning de-biased representations with biased representations.\" International Conference on Machine Learning. PMLR, 2020.\n\n[2] Tartaglione, Enzo, et al. \"End: Entangling and disentangling deep representations for bias correction.\" Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021. \n\n[3] Y.-K. Zhang, Q.-W. Wang, D.-C. Zhan, and H.-J. Ye, “Learning debiased representations via conditional attribute interpolation” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023.\n\n[4] Barbano, Carlo Alberto, et al. \"Unbiased Supervised Contrastive Learning.\" ICLR. 2023.\n\n[5] Zhang, Yi, et al. \"Poisoning for Debiasing: Fair Recognition via Eliminating Bias Uncovered in Data Poisoning.\" ACM Multimedia 2024. 2024.\n\n[6] Wang, Yining, et al. \"Navigate Beyond Shortcuts: Debiased Learning through the Lens of Neural Collapse.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Some minor concerns.\n- I don't understand, why not provide visualizations using linear dimensionality reduction for the motivating example (section 3.2), since you are using a linear model. Using T-SNE somehow confuses the example\n- How were the baselines implemented? are they openly available (it could make sense to provide the links) or did you re-implemented them\n- How many spurious features were masked away in each of the examples" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper deals with a relevant problem. It proposes a new method that is, to the best of my knowledge original. And presents an evaluation in relevant benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new method for mitigating spurious bias in an unsupervised fashion. \n\nIt tries to detect non-essential features based on the pattern of errors, mask them away, and retrain the last layers. The method is compared against other methods in image and text datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concern is that I find the motivation for the method not very strong. I believe the authors don't provide strong evidence for the main assumptions that motivate the methods.\n\nParticularly, core assumptions for the method are that\n1. some features in the latent embedding that are responsible for encoding the *spurious correlation are confined to single neurons*, and can be masked away. It is a bit unclear to me whether this is true. For instance, maybe some component that is not entirely aligned with any specific neuron could be responsible for encoding this spurious feature.\n2. spurious features can be distinguished from core features by looking at the error density. And while the toy example motivates this, It seems the pattern we see in Fig2(b) for spurious vs core features is very different from what we see in Fig 4.\nOverall, I think these are two very important assumptions of the method that should be more clearly demonstrated" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024mitigating,\ntitle={Mitigating Spurious Bias with Last-Layer Selective Activation Retraining},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=x9rtYetTsA},\nnote={under review}\n}" }, "abstract": { "value": "Deep neural networks trained with standard empirical risk minimization (ERM) tend to exploit the spurious correlations between non-essential features and classes for predictions. For example, models might identify an object using its frequently co-occurring background, leading to poor performance on data lacking the correlation. Last-layer retraining approaches the problem of over-reliance on spurious correlations by adjusting the weights of the final classification layer. The success of this technique provides an appealing alternative to the problem by focusing on the improper weighting on neuron activations developed during training. However, annotations on spurious correlations are needed to guide the weight adjustment. In this paper, for the first time, we demonstrate that neuron activations, coupled with their final prediction outcomes, provide self-identifying information on whether the neurons represent spurious features. Using this information, we propose last-layer selective activation retraining, which retrains the last classification layer while selectively blocking neurons that are identified as spurious. In this way, we promote the model to discover robust decision rules beyond spurious correlations. Our method works in a classic ERM training setting where no additional annotations beyond class labels are available, making it a practical post-hoc tool for improving a model's robustness to spurious correlations. We demonstrate that our method is effective with different model architectures and can effectively mitigate spurious bias on different data modalities without requiring annotations of spurious correlations in data." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "spurious correlation", "robustness", "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/b835717ce94880a162b94d1b821fd3245dd64c38.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": "Mitigating Spurious Bias with Last-Layer Selective Activation Retraining" }, "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": "- Some of the figures done have confidence bound (last subplot for Fig 6). Am I missing something?\n\n- What does \"simplicity\" mean on L138?\n\n- Some works suggest that LLMs confidence might be unreliable, I wonder what the intuition on much better results with ICDA is in comparison to random agents?\n\nMinor:\n\n- Figures might benefit from increasing the font size.\n\n- a weird indent on L254" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper introduces a novel application of LLMs for causal discovery - using LLM defined interventions to refine causal discovery.\n\n- ICDA is evaluated on diverse datasets including a dataset not part of the model pertaining.\n\n- The paper is well-written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed Interactive Causal Discovery Agent (ICDA), that uses LLMs for causal discovery through an uncertainty-driven edge intervention selection process. The method prioritizes uncertain edges for intervention and utilizes local updates from feedback, achieving strong performance on a range of real-world causal graphs. Extensive experiments validate ICDA’s robustness and adaptability, showing it outperforms zero-shot LLM prompting across diverse graph structures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of comparison with statistical methods.\n\n- There is a lack of comprehensive results across models. I acknowledge the authors have presented results in Figure 6, but comparing different ICDA variants and random agents for smaller models would have been 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": 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 plots in Figures 2 and 4 are very small and hard to read.\n2. In line 312 there seems to be a space missing - “weablate”\n3. The citation (Sharma & Kiciman, 2020), in line 147 seems misplaced. What was the authors' intention?" }, "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 cleanly written. \n2. The approach is well motivated by the literature. \n3. The experimental section is extensive." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a new method for end-to-end interactive causal discovery using LLMs.\nThe approach comprises two main components intervention selection method based on LLM uncertainty predictions and local update strategy based on newly acquired knowledge. \nThe approach is based on a formulation of edge intervention. The method is evaluated on the set of 7 real-world graphs and compared against its ablations. Additional analysis is provided which covers evaluation with different LLM models and evaluation on the graph unseen during LLM training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The definition of edge intervention feels unrealistic. Could the authors please provide an example of a causal operation that reveals the edge without additional knowledge or assumptions about the graph structure? It seems to me that such data might be extremely costly to obtain and the operation might in some cases be equivalent to revealing the whole graph, thus making the described approach impractical.\n2. The paper lacks a discussion about the limitations of the proposed approach." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 setting may not be realistic:\n- In the proposed setting, line 144, it is assumed that one could directly obtain the ground-truth causal edge label in each intervention, which is not realistic;\n- The setting significantly differs from the standard practice in the literature of experimental design [1,2,3];\n\n2. There is no guarantee that LLMs could provide valid results:\n- It is widely shown that LLMs can not provide faithful causal results [4,5], while the proposed framework heavily rely on the results of LLMs;\n- Similarly, the uncertainty provided by LLMs is not warranted;\n\n3. Previous baselines on experimental design are neglected, for example, previous works on intervention selection or experimental design [1,2,3].\n\nMinor\n- The line numbers of the algorithm are all 0; \n- lots of key steps in the algorithm are not defined;\n\n**Refereneces**\n\n[1] Learning neural causal models with active interventions.\n\n[2] Trust your $\\nabla$: Gradient-based intervention targeting for causal discovery.\n\n[3] Active learning for optimal intervention design in causal models.\n\n[4] Causal parrots: Large language models may talk causality but are not causal.\n \n[5] Discovery of the hidden world with large language models." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(+) This work presents an interesting use of LLMs in causal discovery;\n\n(+) The presentation and organization of this work are clear and easy-to-follow;\n\n(+) Some experiments demonstrate the effectiveness of the proposed approach;" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work studies using LLMs to perform causal discovery in an interactive manner. The authors propose to incorporate LLMs as an agent to produce initial graphs, and iteratively optimize the updated causal graphs by selecting proper interventions. During the selection of the intervention targets, LLMs are leveraged to provide uncertainty measures for the unknown edge. The authors show that the proposed approach can effectively outperforms simple baselines on eight real-world causal graphs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(-) The setting may not be realistic, since it is challenging to directly obtain the ground-truth causal edge label in each intervention;\n\n(-) There is no guarantee that LLMs could provide valid results;\n\n(-) Previous baselines on experimental design are neglected;" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "- Can you elaborate on :\na) how your algorithm satisfies the LLMs as optimizers framework? ;\nb) why using the F1 metric specifically as a loss?,\nc) why these specific updates on parents of intervened edges as the choice of local updates?,\nd) how exactly the intervention is performed, at least in experiments?\n\n- l.94-95 : *Building on Meek (2013), Chickering (2002) proposes a greedy search algorithm that performs well in practice.* There seems to be a confusion in time here... wrong Google Scholar citation?\n\n- Can you increase the font of Figure 2?\n\n- l.403-404 : *Additionally, we notefFor large enough graphs, putting everything in context is simply not feasbile*. Typo? We note that for?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Originalty : The submissions brings an interesting perspective by making use of the use of the literature on LLMs as optimizers.\n\nQuality : the experiments are extensive and exhaustive, performing multiple ablation studies on the model's properties but also on aspects such as memorization.\n\nClarity : the paper is mostly clear in my opinion.\n\nSignifiance : the experiments are interesting as they underline the importance of finding a subtle balance wrt local updates, between throwing the whole graph into the prompt and only modifying intervened edges." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper builds on previous literature on using LLMs for causal discovery on one side, and for active black-box function optimization on the other side, to iteratively update a graph using ground-truth edges obtained through interventions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Originalty/Signifiance/Quality : this point harms the correctness of the claims of the paper and is my main concern : it seems like the iterative updates do not satisfy the framework of LLMs as optimizers. From my understanding of the submission and the references, this frameworks consists in having the LLM decide on next points in the admissible space to query based on former (point, function realization) couples. But here, the next edges to query are done in a pre-determined, algorithmic manner, based on confidences, and the objective (the F1 score) is not used as an objective to optimize and is never parsed to the LLMs. The LLM is simply used on a post-hoc manner *after* having queried edges and read their associated output ground-truth labels.\n\nClarity : there are a few unclear points or mistakes developed in 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": 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": "Refer to weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(1) Introducing LLMs to study causal discovery is an interesting direction.\n\n(2) The author's writing is clear, making it easy to read and understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper applies large language models (LLMs) to causal reasoning. Specifically, the authors prompt LLMs to address two key tasks: (1) selecting which edge to intervene on in the next round, and (2) updating the predicted causal graph. The authors demonstrate that their approach significantly outperforms a random selection baseline across eight different real-world graphs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) The experimental setup is quite simple, comparing only three basic methods: random selection, direct LLM, and static confidence selection.\n\n(2) Additionally, the comparison should include the performance of different language models, not just one.\n\n(3) In the main experimental section, it would be better to include a table for quantitative results alongside the graphs.\n\n(4) The experimental setup is overly simplistic, and for a conference like ICLR, the complexity of the method, theoretical analysis, and experimental thoroughness are insufficient." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We utilize LLMs as a black box optimizer to iteratively propose and update interventions on a causal graph" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024icda,\ntitle={{ICDA}: Interactive Causal Discovery through Large Language Model Agents},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xA8WW2dlTX},\nnote={under review}\n}" }, "abstract": { "value": "Large language models (\\textbf{LLMs}) have emerged as a powerful method for causal discovery. Instead of utilizing numerical observational data, LLMs utilize associated variable \\textit{semantic metadata} to predict causal relationships. Simultaneously, LLMs demonstrate impressive abilities to act as black-box optimizers when given an objective $f$ and sequence of trials. We study LLMs at the intersection of these two capabilities by applying LLMs to the task of \\textit{interactive causal discovery}: given a budget of $I$ edge interventions over $R$ rounds, minimize the distance between the ground truth causal graph $G^*$ and the predicted graph $\\hat{G}_R$ at the end of the $R$-th round. We propose an LLM-based pipeline incorporating two key components: 1) an LLM uncertainty-driven method for edge intervention selection 2) a local graph update strategy utilizing binary feedback from interventions to improve predictions for non-intervened neighboring edges. Experiments on eight different real-world graphs show our approach significantly outperforms a random selection baseline: at times by up to 0.5 absolute F1 score. Further we conduct a rigorous series of ablations dissecting the impact of each component of the pipeline. Finally, to assess the impact of memorization, we apply our interactive causal discovery strategy to a complex, new (as of July 2024) causal graph on protein transcription factors. Overall, our results show LLM driven uncertainy based edge selection with local updates performs strongly and robustly across a diverse set of real-world graphs." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Causal Discovery", "LLM", "Black box optimizer" ] }, "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/d1d07828e7ac3cecb646de94b04c76ba958f2e9c.pdf" }, "presentation": null, "primary_area": { "value": "causal reasoning" }, "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": "ICDA: Interactive Causal Discovery through Large Language Model Agents" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Please address the aforementioned concerns\n2. Some typos: \n - Equation 5 (line 201): the first term doesn’t require parentheses. \n - “reply on” -> rely on in line 268 and 346. \n - There are duplicate references for Behrouz & Hashemi, 2024 \n\n\nConclusion:\n\nWhile the proposed idea is appealing, and I acknowledge its potential impact, I am not sure that the paper is ready for ICLR in its current form. Therefore, I am hesitant to fully support acceptance but I am willing to increase my score if the concerns and questions are addressed." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper provide a method to customize the recursion in the SSMs for graphs,\n- It proposes a method for designing a permutation-invariant kernel for efficient global convolution operations on graphs.\n- The paper demonstrates provably stronger expressiveness of the proposed model for 3-paths and 3-cycles and also for 4-paths and 4-cycles." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work extends the state space models (SSMs) from sequence modeling to the domain of graph-structured data. By tailoring SSMs for graphs, the proposed model (GSSC)\ncan capture long-range dependencies and overcome the limitations of Message Passing Neural Networks (MPNNs) while offering efficient computation addressing the quadratic computation of the graph transformers. To preserve the permutation equivariance in the graphs, it outlines a method for designing a permutation-invariant kernel for convolution operations on graphs. Furthermore, it extends to a data-dependent version of the proposed model by defining a selection mechanism for graph-structured data. The proposed model demonstrates provably stronger expressiveness than MPNNs and Graph Spectral Convolution in counting graph substructures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The distinction between the proposed model (equation 4) and linear graph attention is minimal. It is known, even prior to Dao & Gu (2024), that SSMs can be represented as linear attention. The global convolution representation used here is also an equivalent form of SSMs. Additionally, the use of positional encodings is not novel, as it is a feature used by other works such as GraphGPS, which propose a general framework for transformer-based models and can adopt their approximations like Performer and linear attention. \n\n2. The presentation requires improvement, and some claims need to be more precise. For example: \n - a. at line 143, the *Computational efficiency and parallelism*, The text should clarify that SSMs like S4, when represented as global convolution (equation 2), they can leverage the FFT algorithm to result in parallel and efficient quasi-linear complexity. Parallel scan is used for the recurrence form of equation (1) under certain conditions (employed by data-dependent SSMs like Mamba), which also results in O(n log n) computation but offers parallelization. \n - b. The statement that the kernel should be factorizable as a dot product to be permutation-invariant (line 127), needs revision. The dot product between absolute position representations is one way to achieve translation invariance. Reference [2] offers alternative methods, like cross-correlation, to achieve translation invariance in kernels for global convolution using translation-equivariant functions. \n - c. The factorized form in line 259 is not equivalent to the data-dependent convolution presented in the preceding line. \n - d. The new positional encodings in equation (6) are positional encodings of all nodes but utilize only the features of node u (as it is a function of $x_u$ only). Lines 267 and 267 need correction to reflect this. \n - e.The text should elaborate on how $\\phi()$ in in eqn (5) is modeled to be a permutation equivariant function and capture interactions between frequencies.\n\n\n\n\n3. **Insufficient Empirical Studies**: The literature review lacks citations for many state-of-the-art models, and the experimental section lacks comparisons to them. \n - Since the proposed model is compared with GSC and Linear Graph Transformers in the paper, it is essential to include their performance comparison in the experimental section (particularly in Table 2 for graph substructure counting, where the proposed model is expected to show superior expressiveness). \n - Comparisons with newer models like Spatial-Spectral GNN [1] and Polynormer [3] are also necessary. \n - Additionally, comparisons against other SSM-based models (Graph-Mamba I and II), GSC, Linear Graph Transformers and recent models in Tables 2 and 3 are important to demonstrate the proposed model empirically." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "(i) For the scalable version of the architecture when you proposed using the first k eigenvectors, how does the method address the issue of missing eigenvectors in the middle and end of the spectrum?\n\n(ii) What are the main challenges in extending GSSC to heterophilic graphs, and do you have any preliminary insights on how these challenges could be addressed?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "I have found the paper well written and self-contained. I think a non-expert could find most of the information in the paper, and I appreciate this aspect.\n\n Figures 1 and 2, which demonstrate the problem domain and architecture, are interesting and easy to read. I commend the authors for their explicit effort in making these illustrations clear and informative.\n\nThe insights are didactical and well communicated. The conclusion given by the experiments looks interesting and valuable for future practitioners, while I think a synthesis would be beneficial for the reader." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces the Graph State Space Convolution (GSSC) method, an extension of State Space Models (SSMs) to graph data. GSSC utilizes global permutation-equivariant set aggregation and factorizable graph kernels based on relative node distances as its convolution kernels." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Despite these merits, I have the following concerns about the paper.\n\n1- While there is a careful analysis of the different design decisions/performance tradeoffs, I feel that there is only a limited understanding about what are the properties of the Architecture that lead to these decisions/performance differences.\n\n2- Scalability Concerns: The paper acknowledges the challenge of scalability for larger graphs. To address this, the authors could explore methods for optimizing the computational complexity of the GSSC architecture.\n\n3- Weak experimental study: The paper lacks experimental evaluation of the Graph State Space Convolution (GSSC) method on heterophilic datasets. This omission is significant as such studies are crucial to assess GSSC's performance with heterophilic data and its robustness against issues like over-squashing and over-smoothing, which are common challenges in graph data 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": "1. Compared to other graph SSM works, this study indeed offers a novel perspective. In your opinion, what advantages does this new viewpoint provide over other methods? How do your experiments support these advantages?\n\n2. The title is \"What Can We Learn from State Space Models for Machine Learning on Graphs.\" Could you elaborate a bit more on other graph SSM models / SSM model on other domins, and compare them with GSSC?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Novelty**: Unlike previous methods that design SSM variants on graphs, this paper approaches the problem from the perspective of separable distance kernels on graphs. This idea is both novel and interesting.\n\n2. **Comprehensive Evaluations**: The experimental section of this paper covers synthetic datasets, real datasets, and computational efficiency benchmarks, providing a comprehensive evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Machine learning on graphs often uses Message Passing Neural Networks (MPNNs), but these have limited expressive power. Graph State Space Convolution (GSSC) is proposed as a new method that extends State Space Models (SSMs) to graph data, overcoming these limitations and demonstrating superior performance on benchmark datasets. GSSC achieves good results on many datasets and offers a scalable solution for graph machine learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Experimental Design**: My main concern is that the model evaluated in the experiments is a hybrid of GSSC and MPNN, without showing the performance of GSSC alone. This makes it difficult to assess the independent contribution of GSSC, affecting the judgment of its effectiveness. I strongly recommend that the authors provide experimental results for the GSSC module alone, which will help readers understand the primary contributions of GSSC.\n\n2. **Baseline Selection**: In Section 5.1, the authors compare only with MPNN, without providing baseline comparisons with SSM or GT, which seems insufficient. Considering the authors claim that GSSC is a replacement module for Transformers, I suggest adding comparisons with other models that include shortest path information (such as high-order MPNN) or global information (such as Graph Transformer). Moreover, the current baselines are mainly single MPNN models, and comparing them with the hybrid model (GSSC+MPNN) seems unfair.\n\n3. **Inconsistency in Model Architecture**: In Section 5.1, the authors use selective GSSC, but it is not used in other benchmarks.If different architectural variants of the model are used in the experiments, I advise clearly indicating this in the tables. Additionally, the authors claim that GSSC without selective is already powerful enough, does this imply that using selective GSSC would be better? If so, it is recommended to provide relevant experimental results to support this argument. If selective GSSC is not used due to other disadvantages (such as computational efficiency), it would be beneficial to discuss this further in the paper. Generally, a unified model architecture is more attractive than one that requires adjustments across different benchmarks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see weaknesses." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper achieves competitive performance compared to SOTA methods.\n2. The paper proposes a novel approach in extending state-space models to graphs.\n3. The paper provides a plausible framework for capturing long-range dependencies efficiently." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Graph State Space Convolution (GSSC), an extension of state-space models to graph-structured data. The authors emphasize GSSC’s capability for capturing long-range dependencies in linear time and demonstrate competitive performance across several benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concerns relate to the complexity claims, specifically:\n1. **Layer Complexity vs Expressivity:** The paper states that the complexity of a GSSC layer is “… $O(nmd)$


 where $n$ is the number of nodes and $m$, $d$ are hidden and positional encoding dimension.” (239-240) This means that a GSSC layer has $O(|V|)$ complexity


. Consequently, GSSC is in general *incapable of examining every edge in the graph*, unlike MPNNs with $O(|V|+|E|)$ complexity, such as GINE [1]. Although the authors prove that GSSC is “more powerful than MPNNs” *in terms of the WL hierarchy*, the expressivity implications of not being able to examine every edge seems to be overlooked. Even preprocessing the graph to incorporate edge features into nodes, which itself requires $O(|E|)$ time, would still necessitate $m\\in O(\\frac{|E|}{|V|})$ to store these features without information loss, thus exceeding $O(|V|)$ complexity overall.\n2. **Preprocessing Complexity:** The paper claims that finding the top $d$ eigenpairs with Lanczos methods has $O(nd^2)$ complexity (286-287). However, since sparse matrix-vector multiplication with the Laplacian matrix is necessary for Lanczos methods, the complexity per iteration would be at least $O(|E|)$, resulting in an overall complexity of at least $O(d|E|)$ to find $d$ eigenpairs. This exceeds the paper’s claim of $O(nd^2)$ preprocessing, and thus requires clarification.\n\nDespite these concerns, the paper’s main contributions remain valid. I would appreciate if the authors could clarify these points during the rebuttal phase.\n\n[1] https://arxiv.org/abs/1905.12265" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024what,\ntitle={What Can We Learn from State Space Models for Machine Learning on Graphs?},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xAM9VaXZnY},\nnote={under review}\n}" }, "abstract": { "value": "Machine learning on graphs has recently found extensive applications across domains. However, the commonly used Message Passing Neural Networks (MPNNs) suffer from limited expressive power and struggle to capture long-range dependencies. Graph transformers offer a strong alternative due to their global attention mechanism, but they come with great computational overheads, especially for large graphs. In recent years, State Space Models (SSMs) have emerged as a compelling approach to replace full attention in transformers to model sequential data. It blends the strengths of RNNs and CNNs, offering a) efficient computation, b) the ability to capture long-range dependencies, and c) good generalization across sequences of various lengths. However, extending SSMs to graph-structured data presents unique challenges due to the lack of canonical node ordering in graphs. In this work, we propose Graph State Space Convolution (GSSC) as a principled extension of SSMs to graph-structured data. By leveraging global permutation-equivariant set aggregation and factorizable graph kernels that rely on relative node distances as the convolution kernels, GSSC preserves all three advantages of SSMs. We demonstrate the provably stronger expressiveness of GSSC than MPNNs in counting graph substructures and show its effectiveness across 11 real-world, widely used benchmark datasets. GSSC achieves the best results on 6 out of 11 datasets with all significant improvements compared to the state-of-the-art baselines and second-best results on the other 5 datasets. Our findings highlight the potential of GSSC as a powerful and scalable model for graph machine learning. Anonymous code\nis available at https://anonymous.4open.science/r/GSSC-5ED8." }, "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", "state space 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/0604423a66dd548b730b8ce28d2373b8cc151aaf.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": "What Can We Learn from State Space Models for Machine Learning on Graphs?" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "n/a" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "* Will the theorem hold if $𝑀_{train}$ is unbounded? In RL, policies typically interact with the environment on the fly, allowing for the gathering of infinite samples.\n* Are the two characteristics discussed in Section 4.2 sufficient to ensure robust generalization performance? If not, what additional considerations could be relevant?\n* Are any weights needed in Equation 19? If not, why?\n* Would it be beneficial to consider heterogeneous encoders in the proposed method?" }, "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 easy to follow, with an effective and engaging presentation. The paper begins with a theoretical analysis, systematically deriving the motivation and key designs, and ultimately showing good results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a dual-agent adversarial policy learning framework to address generalization gaps in reinforcement learning (RL). The authors first derive a lower bound on generalization performance, showing that optimizing this bound corresponds to constrained optimization at each RL step. They then leverage some approximations, leading to the dual-agent adversarial framework proposed in this paper. It employs two identical policy networks that are updated alternately to minimize reliance on irrelevant features through a combined loss function, which includes both the primary task loss and a new adversarial loss that includes both adversarial attacks of the other and robust defences of itself. Experiments on the ProcGen benchmark show that this approach outperforms PPO and DAAC baselines in generalization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My main concern is the evaluation. Although the performance is promising, the paper lacks in-depth analysis and extensive discussion on several aspects.\n\nCurrently, it seems that even without generalization, the proposed method is showing good performance. Thus it is unclear if it is due to better convergence or better generalization capability. We should be careful about this when drawing the conclusion that the proposed method has better generalization performance. And it would be helpful to add some experiments/discussion to compare only the generalization if the two baselines have similar in-distribution performance. Also, one needs to check if the comparison is fair or not. It would be helpful to report the wall clock time and number of gradient steps between the proposed method and the baseline.\n\nSecond, the approach is evaluated on only one generalization setup. However, it is unclear how challenging such a generalization setting is. It would be beneficial to conduct further analysis, assessing the degree to which the proposed algorithm enhances generalization across varying levels of difficulty, such as easy, moderate, and challenging generalization cases.\n\nThird, please also consider adding the training complexity compared to baseline methods, as well as a discussion on the impact of hyperparameter $\\alpha$.\n\nAdditionally, the authors may want to compare and discuss their method relative to other approaches aimed at enhancing RL generalization, such as [1][2]. Given that these methods may also align with the two characteristics described in Section 4.2, further elaboration on the distinctions or similarities would strengthen the paper.\n\n* [1] MaDi: Learning to Mask Distractions for Generalization in Visual Deep Reinforcement Learning\n* [2] Policy Rehearsing: Training Generalizable Policies for Reinforcement Learning\n\nIn Section 2, the introduction of the concept of MDP state semantics and the subscript $𝑚$ in the MDP notation is not well-motivated or clearly explained. Consider improving clarity by first defining the distribution $𝑝𝑀$ explicitly and then introducing $m$. Furthermore, the limitations and future works should be discussed in 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": "- One claim is that the method can be widely used with a variety of algorithms. Would you be able to share results on how this approach transfers with different Online RL algorithms in the ProcGen benchmark?\n- Would you be able to provide some qualitative analysis of the representations learned by your framework compared with those of PPO and DACC to further validate the claim that robust representations are being learned?\n- Could you share results of the sensitivity of $\\alpha$ and the selection criterion for it?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Empirically show a significant improvement over prior work in the ProcGen environment with their adverserial learning framwork\n- Provide theoretical insights about how a policy's robustness to irrelevant features improves generalization performance which is a novel contribution that can be generally applied to any algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel adverserial learning framework that involves a minimax game process between two homogeneous agents to improve the generalization capability of these agents in RL. This framework integrates with existing RL algorithms such as PPO, leverages no additional human prior knowledge which can lead to poor robustness in generalization and has minimal hyperparameters allowing for effective applicability. The authors additionally derive lower bounds for the training and generalization performance of the agent and show that by minimizing the policy's robustness to irrelevant features, one can improve generalization performance. The authors evaluate their framework in the ProcGen environment, showing gains over algorithms such as PPO and DACC." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Several works such as DRAC and RARL consider a multi-agent/adversarial optimization process in RL. Would be good to include an extensive evaluation of these approaches as baselines and contextualize the novelty of your approach with respect to each baseline.\n- The method is primarily evaluated in the ProcGen environment and could benefit from additional empirical evaluation with a larger set of RL benchmarks to further evaluate the efficacy of the approach.\n- GANs and other adversarial optimization techniques commonly have issues with mode collapse, vanishing gradients and convergence issues, which all make optimization more difficult. Though this is controlled with the parameter $\\alpha$, would be good to consider the tradeoff of the robustness to adversarial threats and the performance of the agent." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to the Weakness" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. **Solid Theoretical Background**: The approach is backed by strong theory, clearly explaining how it supports RL generalization.\n2. **No Human Bias in Addressing Generalization**: The method achieves generalization without relying on human biases, such as hand-designed augmentations.\n3. **No Extra Network Parameters**: The framework achieves its goals without adding network parameters, relying on just one hyperparameter for flexibility.\n4. **Novel Idea**: The dual-agent adversarial setup is an innovative way to tackle RL generalization.\n5. **Strong Performance**: The approach performs well across tested environments, demonstrating robust generalization and effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a dual-agent adversarial framework to improve generalization in reinforcement learning (RL). In this setup, two agents interact adversarially, each attempting to disrupt the other's policy while maintaining stability in its own. This competition drives both agents to develop robust and generalizable strategies. The framework is efficient, adding only one hyperparameter, and shows strong performance improvements in challenging environments, especially when used with standard RL algorithms like PPO. This approach offers a promising solution for enhancing RL generalization without relying on complex data augmentations or human-designed biases." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Limited Environments and Baselines**: Testing is somewhat limited in environments and baseline comparisons. Adding diverse environments, such as DMC-GB[1], and competitive baselines like PIE-G[2], SVEA[3], and ARPO[4] would provide a more complete comparison and further demonstrate the model's capabilities.\n\n[1] Generalization in reinforcement learning by soft data augmentation., Hansen et al., ICRA 2021\n\n[2] Pre-Trained Image Encoder for Generalizable Visual Reinforcement Learning., Yuan et al., NeurIPS 2022\n\n[3] Stabilizing deep q-learning with convnets and vision transformers under data augmentation., Hansen et al., NeurIPS 2021\n\n[4] Adversarial Style Transfer for Robust Policy Optimization in Deep Reinforcement Learning., Rahman et al., ArXiv 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": 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": "Please see above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "Generalization in deep reinforcement learning is a highly important research direction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The submission proposes an adversarial framework that involves a game process between two agents: each agent seeks to maximize the impact of perturbing the opponent’s policy by producing representation differences for the same state, while maintaining its own stability against such perturbations. The submission conducts experiments in the ProcGen environment with 3 random seeds in 8 different games and provides comparison against DAAC and PPO." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The theoretical claims of the submission follow almost immediately from previous work, and do not bring any additional new knowledge. \n\nTable 1 should include standard deviations. Three random seeds is relatively small to interpret the results reported. The results reported in Table 1 and Figure 5 are contradictory. Table 1 reports that DAAC in climber is 3.299 and PPO + Adv. (Agent 1) is 4.473. However, Figure 5 clearly reports that DAAC performance as the highest. How is this possible?\n\nWhy is it only compared to DAAC and original PPO? There are more studies on generalization in deep reinforcement learning. \n\nIn the DAAC paper there is another algorithm called IDAAC that performs better. Why is the algorithm IDAAC not included in the comparison?\n\nI would also recommend checking page 9 and page 10 of the ProcGen section of paper [1]. In particular, the paper [1] states for ProcGen that: \n\n*“We note that a number of improvements reported in the existing literature are only 50 − 70% likely.”*\n\nFurthermore the paper [1] states:\n\n*“Instead, we recommend using normalization based on the estimated minimum and maximum scores on ProcGen and reporting aggregate metrics based on such score.”*\n\nAs it has been reported in [1] and [3], the performance of PPG [2] is also quite high. It might be good to include PPG in the comparison baseline.\n\nProcGen seems to have 16 tasks. Both of these papers [1,2] test across the 16 games in the ProcGen environment. The submission tests their proposed algorithm in only 8 of them.\n\n[1] Deep Reinforcement Learning at the Edge of the Statistical Precipice, NeurIPS 2021.\n\n[2] Phasic Policy Gradient, ICML 2021.\n\n[3] Decoupling Value and Policy for Generalization in Reinforcement Learning, ICML 2021.\n\nMore recent techniques report substantially higher scores in the ProcGen environment [1,2].\n\n[1] DRIBO: Robust Deep Reinforcement Learning via Multi-View Information Bottleneck, ICML 2022.\n\n[2] Explore to Generalize in Zero-Shot RL, NeurIPS 2023.\n\n\nHow adversarial learning is mentioned in the introduction is incorrect. In the introduction it is stated that: \n\n*“Adversarial framework facilitates the development of agents capable of adapting to new environments by emphasizing the distinction between relevant and irrelevant information.”*\n\nby referring to these studies [1,2,3] as adversarial learning \n\n[1] State-Adversarial DQN for robust deep reinforcement learning, NeurIPS 2020.\n\n[2] Robust adversarial reinforcement learning, ICML 2017.\n\n[3] Robust Deep Reinforcement Learning through Adversarial Loss, NeurIPS 2021.\n\nHowever, recent studies demonstrated that in fact adversarially trained policies cannot generalize, and furthermore the generalization skills of standard reinforcement learning training is substantially higher [1]. \n\n[1] Adversarial Robust Deep Reinforcement Learning Requires Redefining Robustness, AAAI 2023.\n\n\nSince the submission proposes an adversarial training method it would have been good to test against adversarial examples as well. It does not have to be the most state-of-the-art adversarial attacks, but still it would have been good to include for reference.\n\n\nAnother thing I want to mention is that by employing the proposed adversarial learning framework the number of encoder parameters that needs to be trained is in fact doubled. This brings a new set of questions. Is it really a fair comparison to lower capacity models as previous ones? Would the prior methods perform also well if we simply just increased the parameters in the encoder?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 does the adversarial training framework handle environments where the distinction between relevant and irrelevant features is not well-defined or context-dependent?\n- What strategies could be employed to reduce the computational overhead introduced by the dual-agent setup, especially in more complex or resource-constrained environments?\n- Could the method be extended or adapted to improve generalization in real-world tasks beyond the Procgen benchmark, and what modifications would be necessary to achieve this?\n- What impact would the framework have in environments with continuous action spaces or higher-dimensional state representations, where irrelevant features may be harder to isolate?\n- How would performance vary in scenarios where adversarial training results in catastrophic forgetting of useful features, and what mechanisms could prevent this?\n- Are there any considerations for applying this approach to tasks with dynamic or evolving feature relevance, where the set of relevant features may change over time?" }, "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 a dual-agent adversarial framework is an innovative approach that addresses the pressing issue of overfitting and generalization in reinforcement learning, offering a new perspective on how agents can improve adaptability in varying environments.\n- The paper provides a strong theoretical foundation, proving that reducing an agent’s robustness to irrelevant features can lead to better generalization, enhancing the depth of the contribution.\n- The experiments conducted on the Procgen benchmark show significant performance improvements over existing methods like PPO, demonstrating the effectiveness of the proposed framework in real-world, challenging tasks.\n- By focusing on reducing overfitting and enhancing generalization, the paper addresses a critical gap in reinforcement learning research, providing solutions applicable to broader, more complex environments.\n- The framework is well-designed to scale across different environments, making it applicable to a wide range of RL tasks with high-dimensional observations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a dual-agent adversarial framework aimed at improving the generalization capabilities of reinforcement learning (RL) models, which often struggle with overfitting and fail to adapt to minor variations in tasks. The proposed framework facilitates a game process between two agents that learn to perturb each other’s policies while maintaining their own stability, enabling them to focus on relevant features in high-dimensional observations. Extensive experiments on the Procgen benchmark demonstrate that this adversarial approach significantly enhances the agents’ performance, especially in challenging environments, outperforming traditional RL algorithms like Proximal Policy Optimization (PPO). Additionally, the authors theoretically prove that reducing an agent’s robustness to irrelevant features can improve its generalization performance. Overall, the study marks a significant advancement in addressing generalization challenges in deep reinforcement learning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While the framework performs well on the Procgen benchmark, its applicability to real-world tasks remains untested, leaving questions about how well it generalizes outside controlled environments.\n- The dual-agent adversarial framework introduces additional computational complexity, which may pose challenges in terms of scalability and efficiency for resource-constrained systems.\n- Although the framework is shown to improve generalization, more detailed ablation studies could have been included to clarify the contribution of individual components, such as the specific impact of the adversarial training mechanism.\n- The paper assumes that irrelevant features can be identified and suppressed, but it does not sufficiently address how to detect these features in environments where their classification is unclear or context-dependent.\n- The comparison with state-of-the-art methods is somewhat limited, with a stronger focus on performance gains rather than in-depth analysis of differences in behavior between approaches." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "The paper proposes a dual-agent adversarial framework for generalizable reinforcement learning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Dual-Agent Adversarial Framework for Generalizable Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xAYOfMV264},\nnote={under review}\n}" }, "abstract": { "value": "Recently, empowered with the powerful capabilities of neural networks, reinforcement learning (RL) has successfully tackled numerous challenging tasks. However, while these models demonstrate enhanced decision-making abilities, they are increasingly prone to overfitting. For instance, a trained RL model often fails to generalize to even minor variations of the same task, such as a change in background color or other minor semantic differences. To address this issue, we propose a dual-agent adversarial policy learning framework, which allows agents to spontaneously learn the underlying semantics without introducing any human prior knowledge. Specifically, our framework involves a game process between two agents: each agent seeks to maximize the impact of perturbing on the opponent's policy by producing representation differences for the same state, while maintaining its own stability against such perturbations. This interaction encourages agents to learn generalizable policies, capable of handling irrelevant features from the high-dimensional observations. Extensive experimental results on the Procgen benchmark demonstrate that the adversarial process significantly improves the generalization performance of both agents, while also being applied to various RL algorithms, e.g., Proximal Policy Optimization (PPO). With the adversarial framework, the RL agent outperforms the baseline methods by a significant margin, especially in hard-level tasks, marking a significant step forward in the generalization capabilities of deep reinforcement learning." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Generalizable Reinforcement learning", "Adversarial 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/c2da026386f692b9b177fe070b11cec1953ed308.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": "A Dual-Agent Adversarial Framework for Generalizable Reinforcement Learning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I mentioned all comments including reasons and suggestions in the above sections. I recommend that the author will provide all the concerns, and improve the completeness of the paper. If the rebuttal period resolves the above-mentioned concerns, I will gladly raise my score. Also, there are little vague sentences and grammatical errors in the paper. I recommend that the author will revise the paper." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "First of all, the motivation for proposing a new model to address the limitations of existing epipolar-based methods in handling dynamic objects and the need for additional models is well-founded. \n\nWhile it may lack significant technical novelty, it shows a solid understanding of the operation of existing modules and effectively applies them to the target problem. More specifically, applying cross-attention mechanism for full cost volume calculation and incorporating MIM to improve matching similarity are highly appropriate choices.\n\nIn experimental section, the proposed method achieved SoTA results in depth estimation benchmarks compared with existing self-supervised multi-frame depth prediction methods, demonstrating its effectiveness." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors proposed a novel self-supervised multi-frame depth estimation architecture that incorporate the CRAFT module to compress and refine the cost volume through attention mechanism and feature aggregation. The main argument is that training cross-attention layers for image reconstruction facilitates the implicit learning of a warping function, resembling the explicit epipolar warping used in existing methods. Also, by employing masked image modeling for pre-training, the authors can successfully leverage the cross-attention map as a full cost volume for depth prediction in dynamic scenarios without requiring additional information such as camera pose. In experimental section, the authors demonstrated that the proposed method can outperform traditional methods utilizing epipolar-based cost volume in challenging scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One concern is the validity of the proposed CRAFT method for better representation learning. While the full cost volume is well-represented through a clear understanding of the cross-attention module, it is not clear whether the proposed module is operated to function as the author’s intention, because all experimental results include MIM, without an independent evaluation of the CRAFT module itself. Furthermore, the significant performance gap between with and without CroCO pre-training raises further doubts about whether the proposed module is functioning as intended.\n\n\nAnother concern is the lack of explanation regarding the justification for cross-attention. I think that the authors should provide a more descriptive explanation as to why Transformer is advantageous for feature matching compared to CNN. Although feature similarity is explicitly calculated through the cross-attention module, the relatively weaker inductive bias of Transformers does not necessarily align with the effectiveness for feature matching." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 1 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please refer to 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. The Paper is well written, outlining the shortcomings of contemporary/previous works and explains clearly the approach followed by the paper to overcome those shortcomings.\n2. Proposes to use Full Cost Volume instead of Epipolar Cost Volume. Employs MIM-based pre-training and then employs Cross-Attention.\n3. Evaluates the Model performance on relevant datasets to point-out the improvement in the shortcomings of previous methods on KITTI and Cityscapes dataset." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The Paper deals with using Transformer architecture for estimating depth using multiple frames in a self-supervised setting, without using explicit pose information. Paper Proposes to use Full Cost Volume instead of Epipolar Cost Volume. Employs MIM-based pre-training and then employs Cross-Attention. Paper claims to deal with noise much better and does not need an explicit pose network or pose information" }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Qualitative results provided in the Figure 4 is 'not 'good'. Where there is improvement in depth prediction where the Figure 4 highlights using the red boxes but on non-highlighted region of the depth prediction is seen, we see worse results.\na. Trees in the second image is definetly not sharper compared to Dynamicdepth & Manydepth as presented in the paper.\nb. Buildings in the background are much more clearer for DynamicDepth and Manydepth for the first image.\n\n2. Since the tasks proposed here is very close to Multi View Stereo tasks, Could the authors provide the 3D Reconstruction view of the scene to compare with earlier MVS methods ?\n\n3. Limited evaluation on outdoor scenes, would be great to see the efficacy of the model on indoor dataset such as ScanNet or NYU." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the weakness section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well-written and structured;\n\nThe experiments are thorough and conducted on multiple datasets;\n\nThe proposed method outperforms previous state-of-the-art methods;" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors target at the self-supervised multi-frame monocular depth estimation. Besides, they propose to use the cross-attention to replace conventional cost volume. The proposed method is validated on the KITTI and Cityscapes datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Using cross-attention to replace the cost volume has been explored in previous methods, such as [1]:\n\n[1] Revisiting Stereo Depth Estimation From a Sequence-to-Sequence Perspective with Transformers, ICCV 2021.\n\nCould you provide a detailed explanation of the structural differences between the proposed method and CRFAT? It currently appears that the differences are mainly at the output level. \n\nAdditionally, could you offer a comparison of the computational complexity between the proposed method and ManyDepth?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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.\tCompare more recent methods in Cityscapes.\n2.\tPlease compare fairly. Comparison on fair feature extractors.\n3.\tCompare model complexity and inference time.\n4.\tOther questions refer to weakness." }, "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 is simple, clear and easy to reproduce.\n2. The writing is excellent." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a self-supervised multi-frame depth estimation framework, which introduces the cross-attention mechanism to implicitly learn the warping function instead of explicit epipolar warping. To this end, this paper proposes the CossAttention map and Feature aggregaTor (CRAFT), which is designed to effectively leverage the matching information of the cross-attention map by aggregating and refining the full cost volume. In addition, the CRAFT is used in a hierarchical manner to improve the depth prediction. Evaluations on the KITTI and Cityscapes datasets demonstrate that this work effectively in environments with dynamic objects and image noise." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The method presented in this paper lacks novelty. The cross-attention mechanism or the transformer architecture has been widely used in depth estimation task, and can improve depth quality well. Using a cross-attention mechanism instead of traditional warping and coarse-to-fine strategies is not particularly novel.\n2.Performance improvements are limited. The work is based on a transformer architecture, while most of the compared methods are based on CNN architecture such as ResNet and HRNet， which is unfair. Moreover, the quantitative results in Table 4 are worse than the 2022 work and the recent work [1].\n\n[1] Miao X, Bai Y, Duan H, et al. Ds-depth: Dynamic and static depth estimation via a fusion cost volume[J]. IEEE Transactions on Circuits and Systems for Video Technology, 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024revisiting,\ntitle={Revisiting Emergent Correspondence from Transformers for Self-supervised Multi-frame Depth Estimation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xAZLCWbsTF},\nnote={under review}\n}" }, "abstract": { "value": "Self-supervised multi-frame depth estimation predicts depth by leveraging geometric cues from multiple input frames. Traditional methods construct cost volumes based on epipolar geometry to explicitly integrate the geometric information from these input frames. Although this approach may seem effective, the epipolar-based cost volume has two key limitations: (1) it assumes a static environment, and (2) requires pose information during inference. As a result, this cost volume fails in real-world scenarios where dynamic objects and image noise are often present, and pose information is unavailable. In this paper, we demonstrate that the cross-attention map can function as a full cost volume to address these limitations. Specifically, we find that training the cross-attention layers for image reconstruction enables them to implicitly learn a warping function within the cross-attention, resembling the explicit epipolar warping used in traditional self-supervised depth estimation methods. To this end, we propose the CRoss-Attention map and Feature aggregaTor (CRAFT), which is designed to effectively leverage the matching information of the cross-attention map by aggregating and refining the full cost volume. Additionally, we utilize CRAFT in a hierarchical manner to progressively improve depth prediction results through a coarse-to-fine approach. Thorough evaluations on the KITTI and Cityscapes datasets demonstrate that our approach outperforms traditional methods. In contrast to previous methods that employ epipolar-based cost volumes, which often struggle in regions with dynamic objects and image noise, our method demonstrates robust performance and provides accurate depth predictions in these challenging conditions." }, "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 Depth estimation; Multi-frame Depth estimation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/00c698e26a6a6c2fe9bf2f37816027600e6c73fe.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": "Revisiting Emergent Correspondence from Transformers for Self-supervised Multi-frame Depth 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": { "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 previous section" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-written, well-organized, and easy to follow. \n2. The concept of encouraging isometric embedding from $z$—the latent variable—to $h(z) = F(\\cdot, z)$, a functional representation, is interesting. \n3. The efficient estimation of a distortion measure that quantifies the lack of isometry is also novel. \n4. Extensive experiments, including applications to neural SDFs and DeepONets, effectively demonstrate the robustness of the proposed method" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an isometric regularization technique for manifolds embedded in an (infinite-dimensional) function space. The goal is to enforce the learned operator $h: Z \\to \\mathcal{F}$ as an isometric embedding. Here, $Z$ represents a (finite-dimensional) latent space with the standard metric, and $\\mathcal{F}$ is the space of functions on $X$, equipped with an appropriately defined metric. A distortion measure that quantifies the lack of isometry (up to rescaling) is introduced, along with an efficient method for its estimation. Numerical experiments validate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Algorithm 1 and its explanation in Appendix B should be expanded. For instance, further elaboration is needed on estimating the numerator in Equation 9. Additionally, in line 14 of Algorithm 1, should $G$ be given as $G = G_2 / G_1$? \n2. Given the significant computation involved in estimating the distortion measure, especially when computing gradients, a comparison of computational time should be included in the paper. \n3. While the method promotes an isometric embedding between the latent variable $z$ and its functional embedding $h(z) = F(\\cdot, z)$, it remains unclear if an autoencoder mapping the original input $u \\in \\mathcal{U}$ to its latent code is isometric. For example, in DeepONet, where $u$ represents the input function, if the encoder from $u$ to $z$ diverges substantially from an isometric mapping, enforcing isometry in $z \\mapsto h(z)$ may have limited impact." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- How does the computational cost of the proposed method compare to competing methods like LipDeepSDF? I’m particularly interested in understanding how the marginal cost of training the latent variable model with isometric regularization compares to the marginal benefit relative to cheaper baselines.\n- Why is it insufficient to enforce the smoothness of F with respect to the latent coordinates alone? Is this especially challenging from a technical standpoint? More challenging than isometric regularization?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The paper is well-written, and the authors do a good job of introducing the mathematical preliminaries needed to describe their method.\n- The experiments persuasively make the case that isometric regularization empirically improves on the Lipschitz regularization baseline.\n- The numerical scheme the authors propose in Algorithm 1 to avoid costly Jacobian computations and sampling of spatial coordinates at training time is a reasonable strategy for approximating their regularizer." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new regularizer for implicit neural representations (INRs). Motivated by the fact that existing methods such as Lipschitz regularization result in overly smooth neural representations with respect to the spatial coordinates, the authors propose isometric regularization. This regularizer encourages the image of the latent variable model F(x,z) with respect to the latent z variable to form a well-behaved manifold, so that small changes in the latent variable lead to small changes in the resulting INR. As their regularizer requires expensive Jacobian computations at many spatial and latent inputs during training, the authors employ the Hutchinson trace estimator to avoid materializing full Jacobian matrices during training, and pre-sample input points to avoid the need to draw fresh samples during training. They then demonstrate that their method improves on an unregularized baseline and Liu et al’s Lipschitz regularization on surface reconstruction, BRDF learning, and neural operator learning tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I’m not sure if I understand the motivation for isometric regularization relative to simpler alternatives. The authors state that existing works such as Liu et al (2022) apply Lipschitz regularization to the latent variable models F to make them smooth across both the input space X and the latent space Z; this is problematic because for any fixed latent z, the function h(z) is overly smooth in X-space. While I agree that this is a notable deficiency in existing methods, why can one not simply enforce the smoothness of F with respect to the latent coordinates alone?\n\nThis question is especially salient in light of the high cost of isometric regularization relative to simpler baselines like Lipschitz regularization. I wonder whether the undeniable improvements arising from using isometric regularization outweigh its costs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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.\tIn Line 416, it says “The effect of regularization is prominent when the training data N is greatly reduced to 20”. However, it is difficult to object the significant improvement of IsoAD with N=20 from Fig. 6? Could you clarify this argument?\n2.\tHow does the strength of regularization choose for different datasets in the paper?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThe approach is grounded in differential geometry, offering a mathematically rigorous framework for managing functional data representations. \n2.\tIt demonstrated effectiveness on various data types, including neural signed distance functions (SDFs), bidirectional reflectance distribution functions (BRDFs), and neural operators, showcasing its versatility. \n3.\tThe method preserves data fidelity while regularizing, resulting in better interpolation, reconstruction, and generalization across different types and qualities of data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a regularization method, Isometric Regularization, for handling manifolds in implicit neural representations (INRs) designed for infinite-dimensional functional data. It suggests that existing regularization methods often over-smooth data, leading to a loss of fidelity. To address this, the authors introduce a Riemannian manifold-based regularization that minimizes curvature and preserves the geometric consistency between the latent space and data manifold. Experiments across multiple data modalities demonstrate that this approach enhances the structure of the latent space, providing better generalization, especially for noisy and small datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThis work uses Hutchinson’s stochastic trace estimator to approximate the trace terms, but it is unclear how much additional computational cost this method incurs compared to the baselines. It is better to include a discussion of the computational requirement.\n2.\tIt utilizes offline samples from the \"ground truth functional data\" to compute the expectation of $J^TJ$. It should provide the justification of accessing the ground truth data.\n3.\tThe regularization process should involve parameter tuning, which might be non-trivial. The paper lacks discussion of how to choose the strength of isometric regularization when applying the method to novel datasets or configurations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Please clarify what is the total loss function that is being optimized for, and what parameters are being learnt, and what are being kept fixed?\n\nHow is the latent-space itself being learnt?\n\nDiscussion about training convergence and model size comparison (without needing new experiments).\n\nClarify ‘auto-decoder’ terminology." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Geometry-preserving and isometry-preserving loss functions are a novelty, and are a growing area of work in generative models. \n\nThe approach to add an isometry-preserving loss to conditional INRs is novel. Approximate algorithms to compute this isometry measure are proposed which enhance the strengths of the paper.\n\nMultiple types of results for proposed INR learning are shown including on 2D, 3D, and neural operator learning." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a way to conditional implicit neural models that may have been conditioned on additional information, such as a class label, or shape vector information. The core idea is to create a mapping from latent-codes to generated output in a way that preserves isometry – that is changes in input and changes in output should be related in a way that preserves distances and angles in the input and output spaces, subject to certain scaling parameters encoded by the function Jacobian. The paper shows experimental results that show the favorable properties of this approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Notation is unnecessarily complicated and hinders readability. For instance in the definition of the INR itself, \\mathcal{X} is more often than just \\mathbb{R}^m with m = 1, 2, or 3. Similarly, V is simply \\mathbb{R}. The lean toward generality isn’t helping with comprehensibility, although everything could be described without loss of generation for f: \\mathbb{R}^2 -> \\mathbb{R}. I would recommend taking such an approach and moving the generalized framework for an appendix instead. We suffer from the same issue once again when defining the structure of the latent space \\mathcal{Z} which always happens to be simply a vector space of the type \\mathbb{R}^m.\n\nThe premise of section 4, which is the core starts of in a confusing way by stating “Without proper regularization, the latent space can become ill-behaved, overfitting to the data instances, which is exacerbated by the infinite dimensionality of the function space” – here it is unclear what is meant by ‘latent space can become ill-behaved’ – because until this point \\mathcal{Z} is referred to as the latent-space, which is more or less given and not learnt end-to-end. E.g as I understood it, \\mathcal{Z} can refer to class categorical labels, or shape-features that are learnt elsewhere. So I am not at all clear what is meant by ‘latent space can become ill-behaved’.\n\nAfter reading section 4, we are presented with a way to compute a measure of isometry – as I understand it, this measure really applies to the core INR map, but it is not clear how this measure is actually used in a loss function. No loss function is described from what I can tell. It is unclear thus where latent-codes themselves are being optimized for, or is only the INR mapping portion being optimized. On reading the appendix, it is found that network parameters and the latent codes are both being optimized. I would recommend clarifying this upfront, as we do not see any loss function that suggests latent-code optimization, and this makes it hard to understand how this is being done. This also means that the latent-codes cannot be class categorical labels, as I had initially thought they could be. Overall, this leads to multiple types of confusion which I would suggest really clearly describing.\n\nThe paper also uses the term auto-decoder multiple times, which is unclear what that means. Do they mean ‘decoder-only’ architecture, or ‘auto-encoder’. I did additional digging among the references cited to see if I am missing something, but I could not find what an ‘auto-decoder’ means anywhere.\n\nMany conditional INRs also include an element of randomness that creates variation in output, which could be easily included in this model, but was less clear if this was already considered. \n\nDetails about training convergence, comparison with other approaches in terms training times, model size would help position the experiments more clearly." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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. How does the proposed isometric regularization scale with respect to the dimensionality of the latent space and the number of data points? Have you tested its performance on more complex datasets or tasks beyond those mentioned in the paper?\n2. How sensitive is the approach to the choice of hyperparameters, such as the weight of the isometric regularization? Is there a systematic way to select these parameters for different types of data?\n3. Can you provide more insights into the computational costs relative to other regularization methods, such as Lipschitz regularization or weight decay, especially for larger datasets?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The proposed method is tested across various types of functional data, including neural SDFs, BRDFs, and neural operators. This cross-domain applicability highlights the generality and versatility of the approach. Experimental results show significant improvements in the quality of latent space, interpolation, and reconstruction tasks. The breath of synthetic to realistic tasks also gives a degree of insight both into the underlying functionality of the method as well as its practical use. The authors also address the computational complexity of infinite-dimensional regularizations and use practical approximations for computational feasibility. Finally, the authors provide an interesting theoretical perspective by interpreting the regularization problem through a Riemannian lens, allowing future researchers to leverage differential geometry principles to derive further regularization algorithms over functional spaces.\n\nIn general, this was a concrete, interesting problem with a fresh take that was presented well and demonstrated very well. I enjoyed reading this paper and believe it will make a positive impact on the field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel isometric regularization technique for Implicit Neural Representations (INRs) aimed at regularizing learned functions, a crucial component of optimizing over a theoretically infinite-dimensional space. This regularization is interpreted through the lens of Riemannian geometry over functional spaces, which helps maintain a well-behaved \"latent space\" while reducing overfitting, even for small or noisy datasets. The proposed method is validated across multiple scenarios, including synthetic, 2D, and 3D visual data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the paper demonstrates robustness and generalizability across several datasets, it does not provide sufficient exploration into the scalability of the approach for very high-dimensional latent spaces or large datasets. An obvious interesting application is in NeRFs, but in my view it is reasonable to delegate this to future work -- there is enough for the paper itself to stand.\n2. The paper lacks comprehensive ablation studies on the choices of parameters and effect of the individual components (e.g. the approximations), which may help future researchers in applying and developing from this method." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present isometric regularization for manifolds of functional data, leading to robust data representation learning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024isometric,\ntitle={Isometric Regularization for Manifolds of Functional Data},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xBuURiCChw},\nnote={under review}\n}" }, "abstract": { "value": "While conventional data are represented as discrete vectors, Implicit Neural Representations (INRs) utilize neural networks to represent data points as continuous functions. By incorporating a shared network that maps latent vectors to individual functions, one can model the distribution of functional data, which has proven effective in many applications, such as learning 3D shapes, surface reflectance, and operators.\nHowever, the infinite-dimensional nature of these representations makes them prone to overfitting, necessitating sufficient regularization. Naïve regularization methods -- those commonly used with discrete vector representations -- may enforce smoothness to increase robustness but result in a loss of data fidelity due to improper handling of function coordinates. \nTo overcome these challenges, we start by interpreting the mapping from latent variables to INRs as a parametrization of a Riemannian manifold. We then recognize that preserving geometric quantities -- such as distances and angles -- between the latent space and the data manifold is crucial. As a result, we obtain a manifold with minimal intrinsic curvature, leading to robust representations while maintaining high-quality data fitting. Our experiments on various data modalities demonstrate that our method effectively discovers a well-structured latent space, leading to robust data representations even for challenging datasets, such as those that are small or noisy." }, "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": [ "Isometric regularization", "Geometric reularization", "Implicit Neural Representation", "Manifold Learning", "Neural SDF", "Neural BRDF", "Neural Operator" ] }, "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/fe3a1d3b061fec86139b5a6c12eaa50e2ed7e7fc.pdf" }, "presentation": null, "primary_area": { "value": "unsupervised, self-supervised, semi-supervised, and supervised representation learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/e3635f39908f81375ad59129941eb81b7ade6007.zip" }, "title": { "value": "Isometric Regularization for Manifolds of Functional 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": null, "code_of_ethics": null, "comment": { "value": "Update: It has been fixed. Thank you.\n\n---\n\n~~Dear Reviewer VRoQ,~~\n\n\n~~Thank you for your careful review. It appears the review intended for another submission may have been posted here by mistake. Could you please help us in resolving this?~~\n\n\n~~Thank you,~~\n\n~~-Authors of Submission 1361~~" }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "[Fixed] Possible review submission error" }, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- In example 2: This will be much more convincing if there is a numerical example that actually shows this (like example 1). Otherwise, it's not obvious to me why selection might imply that $X_1, X_3$ are dependent when conditioning on $X_2$.\n- In section 3.2: It is claimed that interventional distributions are not Markovian due to the original dag due to $X_1$ not being independent of $X_3$ when $X_2$ is conditioned on. The open path here goes through $X_2^*$. However, in example 4 it is claimed that conditioning on $X_2$ automatically conditional on $X_2^*$, This seems contradictory to the previous statement, what am I missing here?\n- Thm 1: If this is contrasted directly with the regular Markov property in DAGs, it will help clear up what the extra condition here is. Reading this section, the extra condition has not been motivated with respect to the examples shown. For example, how does this theorem differentiate DAGs $\\mathcal{G}$ and $\\mathcal{H}$.\n- I don't really understand how Lemma 1 \"misses key distributional information\", could this be made more explicit?\n- L291: Why does the twin graph have fewer degrees of freedom with the invariance constraint? And why does it lead to CIs not implied by d-separations? The claim in L293 is also not clear at all to me.\n- Section 3.3: Intuitions behind the lemmas and why they are needed would greatly help here. Right now they are hardly motivated and tough to understand given the dense notation.\n- Algorithm 1 step 2: how is this done? What method is used to test if the conditional distributions are the same or not?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The authors have found an interesting flaw in previous methods in the presence of selection bias\n- The paper is well motivated with clear examples. Some examples can be made clearer (see below)" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The work shows that current theory for dealing with interventional causal discovery is insufficient under selection bias, as there are cases where the selection mechanism takes place before the intervention. The authors then propose a twin graph framework where the selection happens before the intervention and then define a Markov property on this twin graph that implies certain independences in distributions when conditioned on the selection mechanism. The authors then propose a method that constructs the graph based on the Markov property up to an equivalence class." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Some bits of the exposition are unclear (see below)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 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. Please refer to the questions mentioned in Weaknesses.\n\n2. I think theoretically the proposed framework (the proposed graphical model + algorithm) should be able to handle interventions without selection bias, but how does the proposed algorithm perform empirically compared to existing methods in scenarios without selection bias?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The selection bias in interventional experiments is an under-explored but crucial issue in causal discovery, and the authors use two clear examples to illustrate why this problem matters and why the existing methods and simply augmenting DAG fail.\n\n2. The authors provide a solid theoretical foundation in this paper by (1) rigorously defining the interventional twin graph and characterizing its Markov properties and (2) proving the soundness of the proposed algorithm.\n\n3. Synthetic experiments show that the proposed method outperforms baselines in handling selection bias and remains robust as the number of variables increases. It also uncovers novel causal relationships in real-world applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the issue of selection bias in interventional causal discovery, highlighting how existing methods fall short when subjects are selectively enrolled in experiments. To address this, the authors introduce a new graphical representation called \"interventional twin graph\" that explicitly represents both the observed world (where interventions are applied) and the counterfactual world (where selection occurs). They characterize the Markov properties of the proposed graphical model and develop an algorithm built upon FCI based on the proposed graphical model and Markov properties, named CDIS (Causal Discovery from Interventional data under potential Selection bias). The authors prove the soundness of CDIS. Experiments conducted under selection bias conditions demonstrate that CDIS outperforms baselines on synthetic datasets and uncovers novel causal relationships in real-world applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the introduction of the interventional twin graph and its Markov properties is rigorous, it may be challenging for readers to grasp at first glance due to its complexity. Providing a high-level explanation to offer an intuitive understanding would greatly benefit readers.\n\n2. The interventional twin graph is more complex than a simple DAG and involves additional nodes. It would be helpful if the authors discussed the computational cost of the proposed model compared to the simpler DAG, including an analysis of the algorithm's computational complexity under the new graphical model.\n\n3. The authors did not address the identifiability guarantees of the proposed method. It would be useful to know if the method can reliably identify the selection variables and under what conditions the true interventional twin graph can be identified.\n\n4. Minor typos:\n * Line 48: \"We show that existing existing graphical representation paradigms\" --> \"We show that existing graphical representation paradigms\"\n * At the end of line 169: \"models a completely different scenario\" needs to be revised" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 are the main challenges that affect the precision and completeness of the algorithm? How sensitive is it to the unmeasured confounders?\n2. If selection bias is detected through diagnostics, how can this information be leveraged to help causal discovery?\n3. Any data points on the computational cost of the algorithm?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The problem of selection bias is important yet often overlooked in existing interventional causal discovery. The setting is general.\n2. This paper is technically sound, with clear formulation of the causal DAG and Markov properties. \n3. The illustrative examples enhance the paper's clarity, helping readers better understand the concepts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the limitations of current graphical models and causal discovery methods in handling pre-intervention selection bias in interventional causal discovery. To overcome these limitations, the authors propose a novel twin graph model that effectively captures both the observed world and the counterfactual world where selection occurs. The paper establishes the Markov properties of this new model and introduces a provably sound algorithm for identifying causal relationships. The effectiveness of this approach is demonstrated through synthetic and real-world experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "More comprehensive results and explanations of the empirical studies would be beneficial to support the effectiveness of the proposed algorithm. For example:\n1. For the simulation, can you report the proportion of true causal edges estimated as directed edges by the algorithm? Additionally, a comparison of the output graphs with the ground truth would illustrate how the new algorithm perform differently from other methods under selection bias. \n2. For the gene application, can you provide more comprehensive analysis of the result? \n3. For the education dataset, can you explain why the pre-intervention selection bias is a potential issue? Highlighting and interpreting key information of the resulting graphs would be helpful, as the current graph and variable names are difficult to follow.\n\nMinor comments about clarity:\n- The notation in this setting is dense and improvements of readability can be helpful for readers less familiar with the area. For example, \"CI\" in line 93 and different types of arrows in Example 7 can be clarified before their first appearance.\n- Typos: line 48 \"existing\", line 295 \"false\"." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Can you think of any ways in which you could evaluate the completeness of your proposed algorithm in the simulations? In a causal discovery paper, we should be concerned with making true discoveries, not just avoiding false discoveries.\n\nDoes the paper's approach generalize to arbitrary latent variables?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper clearly lays out the problem of selection bias in causal discovery and why certain natural approaches to the problem are not sufficient. The paper also puts forward a very general solution to the problem and considers its consequences. Overall the paper is well-written, despite being notation-heavy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers the problem of interventional causal discovery in the context of selection bias. The paper first explains why simply augmenting the causal DAGs with a selection variable is insufficient. The paper then goes on to introduce the concept of a twin graph, in which every node is replicated in a counterfactual/\"basal\" world, and there is a defined set of rules by which the relationships of the basal world are constructed. Based on this construction, the paper goes on to define the set of d-separations that are implied by this model and establish the graphical criteria for Markov equivalence. The paper provides an algorithm to recover the Markov equivalence class and evaluates it on both simulated and real-world data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One thing that was unclear to me is how complete is the paper's characterization of Markov equivalence classes in the given model. I would think that the Markov equivalence class would encode all DAGs with the same conditional independence structure (i.e. the right-hand side of the implications in Theorem 1). However, the equivalence structure is defined with respect to the left-hand side of the implications in Theorem 1. This would seem to imply that the equivalency classes are not as fine-grained as they potentially could be.\n\nThe algorithm provided also suffers from this issue, in that the authors point out that it may not be complete. It is not clear to me how useful it is to have a causal discovery algorithm that is sound but not complete. The trivial algorithm that says there are no causal relationships is sound but not useful.\n\nThe simulation study not reporting any information on completeness is disheartening. While I understand that the paper does not contain any guarantees on completeness, in the simulations there is access to the ground truth. So it is hard to see how there is no way to evaluate the ability to discover some fraction of those relationships.\n\nAt a higher level, I'm not sure how much of the framing of the paper is specific to the selection problem. It seems like the approach of the paper is tackling the more general problem of causal discovery with unobserved latent variables. If that is not the case, then the paper should explain how their methods do not generalize to the latent variable 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": 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": "-" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper studies a relevant and interesting problem - both mathematically and philosophically. It considers the question: \"What does selection bias actually mean\" and proposes a sound answer and the necessary mathematical framework to deal with such situations.\n- Based on the framework to treat selection bias, a sound and complete causal discovery algorithm is proposed.\n- The method is evaluated not only on synthetic data, but on real-world examples as well. This provides some confidence that it may be useful in practical applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the important but overlooked problem of selection bias in interventional causal discovery, where subjects are selectively enrolled in experiments (e.g., drug trials focusing on patients with specific conditions). The authors show that existing methods for interventional causal discovery fail to properly handle selection bias, as subtle differences in when and where interventions occur can lead to inconsistent conditional indepenence relations. To address this, they introduce a novel graphical model called the \"interventional twin graph\" that explicitly accounts for both the observed world (where interventions are applied) and the counterfactual world (where selection occurs before interventions), along with characterizing its Markov properties and equivalence classes. They develop a provably sound algorithm called CDIS (Causal Discovery from Interventional data under potential Selection bias) that can identify causal relations and selection mechanisms from data with soft interventions and unknown targets. Through experiments on both synthetic data and real-world applications in biology and education, they demonstrate that their method achieves higher precision in identifying true causal relations compared to existing approaches when selection bias is present." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The biggest weakness I see is the presentation of the paper. The first two sections are dense, but give a good introduction and motivation to the problem, based on good illustrations in Examples 1 and 2.\n\nHowever, Section 3 is the most painful piece of text I have read in a while. It relies mostly on mathematical notation to bring across the main points and lacks the contextualization in prose. I appreciate that examples are given in Section 3, but even those are a bit cryptic and fail to provide an accessible intuitive understanding. I suppose there are three main reasons for this: (1) Writing about complex SCMs is inherently difficult and a certain level of formalism is necessary - not much you can do here. (2) The amount of content in the main paper, given the page limit might a bit too much. Some of the more technical parts could be relegated to the appendix and exchanged for more contextualization. (3) The text could consider the reader's state of mind more. Some examples:\n\n- L211f: introducing the functions $f^*$ uses the mathematical symbols for the corresponding variables in the counterfactual basal world to introduce them, but does not use the word \"counterfactual\". That means as a reader, I either have it in working memory, or I have to go back to the definition and jump back again to the sentence to parse it.\n- As far as I can tell, abbreviations like \"CI\" and \"MAG\" aren't defined, or used before they are defined, e.g. \"PAG\".\n\nSuch presentation choices add unnecessary mental effort for understanding, and I would think twice if I'd go back to this paper and build on it for future work - not because it's wrong, but because of the mental effort to access the information." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024when,\ntitle={When Selection meets Intervention: Additional Complexities in Causal Discovery},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xByvdb3DCm},\nnote={under review}\n}" }, "abstract": { "value": "We address the common yet often-overlooked selection bias in interventional studies, where subjects are selectively enrolled into experiments. For instance, participants in a drug trial are usually patients of the relevant disease; A/B tests on mobile applications target existing users only, and gene perturbation studies typically focus on specific cell types, such as cancer cells. Ignoring this bias leads to incorrect causal discovery results. Even when recognized, the existing paradigm for interventional causal discovery still fails to address it. This is because subtle differences in _when_ and _where_ interventions happen can lead to significantly different statistical patterns. We capture this dynamic by introducing a graphical model that explicitly accounts for both the observed world (where interventions are applied) and the counterfactual world (where selection occurs while interventions have not been applied). We characterize the Markov property of the model, and propose a provably sound algorithm to identify causal relations as well as selection mechanisms up to the equivalence class, from data with soft interventions and unknown targets. Through synthetic and real-world experiments, we demonstrate that our algorithm effectively identifies true causal relations despite the presence of selection bias." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "causal discovery", "selection bias", "experiments", "interventions" ] }, "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/4007a87b0e47f7c459c6d887383f2fc559a863f8.pdf" }, "presentation": null, "primary_area": { "value": "causal reasoning" }, "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": "When Selection meets Intervention: Additional Complexities in Causal Discovery" }, "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": "Major: \n- Do you compute area-weighted RMSEs? If not, I strongly encourage fixing this (especially for the weather forecasting experiment; see e.g. Weatherbench2).\n- For the fine-tuning experiments, do you start them based on the weights resulting from the first or second pre-training stage? If the former, then the second pre-training seems to also be a form of fine-tuning and I would suggest avoiding using the term \"zero-shot\" for reporting the forecasting results.\n- From looking at Figure 3, I don't think that I agree with the author's interpretation that *\"It is interesting that reconstruction performance is relatively little affected by lead time at the lower end of masking ratios\"*. There's a clear cap between \"0h, global\" and \"6h, global\", especially for the lower end of masking ratios. Do I miss something?\n- Can you include downscaling results on more variables than only T2M?\n\nMinor:\n- What's the difference between the encoder and decoder blocks? Fig. 1 suggests these are identical... Is the difference some densification by the \"reconstruct batch\" module in Fig. 1? If so, can you explain this more and make it clearer?\n- Line 174: \"reduce the masking ratio to 50%\"... is this a typo (you use the same rate for the first stage)? What's correct?\n- Why is there no reference line in figure 5b)?\n- Please define what's meant by spatial and temporal RMSEs.\n- Can you include snapshots like Fig. 9 but at the native temporal resolution used for prediction (i.e. not a monthly mean)?\n- How does using different patch sizes (larger than the used size of 1) impact downscaling performance? \n- Why do you call the model Prithvi WxC?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Prithvi WxC is quite flexible as it can be used for a broad range of downstream applications, as convincingly shown in the experiments. \n2. The method contains original ideas such as the pre-training objective and using climatology-derived anomalies as targets, which I found interesting to read about.\n3. The paper is generally clearly written and easy to read." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new foundation model, Prithvi WxC, for atmospheric modeling applications in weather and climate. Prithvi WxC was trained on 3-hourly data from 1980 to 2019 from the MERRA-2 reanalysis dataset based on a masked reconstruction/forecasting pre-training objective. \nThe model follows a transformer-based encoder-decoder architecture inspired by Hiera and MaxViT.\nAfterward, the model is fine-tuned for various downstream tasks: Medium-range weather forecasting, global and regional downscaling, and learning a gravity wave flux parametrization. These tasks have different sets of spatial resolutions, variables, and datasets, showing the flexibility of the foundation model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper falls short of establishing a compelling case for Prithvi WxC as a foundation model for weather or climate. The practical significance and advantages of this approach remain inadequately demonstrated:\n\na.) While foundation models typically excel at zero-shot performance and data-efficient fine-tuning across diverse tasks, the evidence presented for Prithvi WxC's capabilities in these areas is not convincing. Baselines for the non-forecasting experiments are either very weak (interpolation-based downscaling) or non-existent (gravity wave experiments). Some highly relevant and simple baselines are: \n- How much worse(?) does Prithvi WxC perform on these tasks if you omit the pre-training stage (i.e. initialize with random weights instead of the frozen pre-trained ones, and train all parameters jointly from scratch on the tasks)? \n- How about completely removing the pre-trained transformer backbone (i.e. removing the Prithvi WxC block from Figures 12 & 13)? \n- For the latter, it would be also good to run an experiment where you replace the pre-trained Prithvi WxC backbone with some \"lightweight\" blocks (e.g. a (deeper) U-Net), trained in a task-specific way from scratch, to account for the huge difference in parameter counts if you completely remove Prithvi WxC. \n\nThese ablations would immensely help in understanding how useful the pre-training stage is for these downstream applications (e.g. does using pre-trained Prithvi WxC improve performance over such simple baselines? Is it more data-efficient?). Besides, otherwise, it is hard to see evidence for the claim in the conclusion that *\"Instead of building task-specific ML-models from scratch, these pretrained encoders can be used to develop more precise data-driven models of atmospheric processes\"*.\n\nb.) No ablations are included. I understand that training such a huge model is expensive but having a few ablations would have been very appreciated (perhaps, with a smaller-scale version of the model). For example:\n\n- How crucial is it to predict climatology-normalized targets as opposed to normal per-variable means/stds? \n- What's the forecasting performance of Prithvi WxC after the first pre-training phase?\n- How important is local vs global masking? What about the masking rates?\n- What's the line of thought behind randomizing the distance between input timesteps? Can the model only use one input timestep? I presume this is possible by masking the corresponding snapshot by 100%, but no experiments with this setting are shown. \n\nc.) The weather forecasting results seem lukewarm, albeit it is hard to judge because the comparison is not apples-to-apples.\n- Prithvi WxC is trained and evaluated on Merra-2. The baselines are evaluated on ERA5. These reanalysis datasets have different spatial resolutions. The evaluation years seem to be different too (correct me if I'm wrong). It would help to fix this mismatch. For example, given the foundational nature of Prithvi WxC... why not fine-tune it on ERA5 directly? Showing that it can be competitive to these baselines in an apples-to-apples comparison would be a very strong result.\n- Based on the mismatched comparison, Prithvi WxC seems to be competitive on 6h to 12h forecasts but it's quite notable that its performance implodes compared to the baselines for longer lead times. It is very unclear why. I wouldn't necessarily expect this version of Prithvi WxC to be state-of-the-art, but the performance does seem underwhelming. Especially given that the authors did \"several things\" to tune these results (i.e. a second forecasting-specific pre-training stage and autoregressive rollout fine-tuning).\n- The hurricane evaluation includes hurricanes from 2017 to 2023. This seems to overlap with the training data period (up to 2019). \n- Either Figure 6 or its analysis in the main body of the text (lines 251-253) is wrong because I see all of the three models do best on exactly one of the three RMSE figures.\n- For the hurricane forecasting experiments, I would appreciate a comparison to the state-of-the-art models included in the weather forecasting experiments (e.g. GraphCast) which have shown to be better than FourcastNet.\n\nd.) The downscaling problem setup is artificial. Downscaling coarsened of existing reanalysis/model outputs is not of much use in practice. A realistic and important downscaling application, as discussed in the Appendix, would be to downscale coarse-resolution model outputs to high-resolution outputs (either of a different model, observations, or the same model run at higher resolution).\n \ne.) The climate model parameterization experiments should be more carefully interpreted. \n- The model predicts outputs that are normalized by the 1980-2019 climatology. Unfortunately, decadal or centennial simulations of the future under a changing climate are inherently a non-stationary problem. It is highly unclear if Prithvi WxC would remain stable, let alone effective, under this highly relevant use case. This is particularly so as the in-the-loop (coupled to a running climate model) stability of ML-based climate model parameterizations is a well-known issue.\n- The selling point for ML-based emulators of climate model parametrizations is often their computational cheapness. Thus, the runtime of Prithvi WxC should be discussed. Given the large parameter count of Prithvi WxC it might be important to note its runtime as a limitation for these kinds of applications.\n- Line 461 claims that Prithvi WxC \"outperforms\" task-specific baselines but no baselines whatsoever are included in the manuscript for this experiment.\n- Are the inputs a global map? I am not familiar with gravity waves, but I believe that most physics parameterizations in climate models are modeled column-wise (i.e. across atmospheric height but ignoring lat/lon interactions). This is surely a simplification of these parameterizations, but it seems to indicate that they're highly local problems. What's the motivation for using global context then?\n- The end of the section should be worded more carefully, clearly stating the aforementioned limitations.\n\nf.) No scaling experiments are included. Thus, it is unclear how important its 2.3 billion parameter size is, how well the model scales, and how its size impacts performance on the downstream applications. Besides, vision and language models are usually released with multiple model sizes that cover different use cases (e.g. balancing inference speed with accuracy). It would be really useful to get these (and carefully compare them) for Prithvi WxC.\n\n2. Related work is insufficiently discussed. Please include an explicit section discussing it, focusing on:\n- Carefully comparing similarities/differences to existing weather foundation models (e.g. architectures, pre-training objectives, downstream applications etc.). Besides, ClimaX is not properly discussed in the paper. Given that it's also a transformer-based foundation model, validated on forecasting, downscaling, and climate emulation, it is very important to include it in the comparison. \n- Similarly, please discuss how exactly the masking technique in this paper relates to the ones proposed in Vandal et al. and McNally et al..\n- Carefully discuss how the architecture is derived from Hiera and/or MaxViT (and other papers of which components were derived, if any).\n\n3. While the authors transparently discuss some issues/limitations with their experiments (e.g. the evaluation data mismatches), it would be nice to also include an explicit paragraph or section on this (and include aforementioned things like the issues with the climate model parameterization experiments).\n\nMinor:\n- Can you properly discuss, and include a reference to, what a Swin-shift is?\n- Similarly, for the \"pixel shuffle layers\"\n- Line 39: Pangu -> Pangu-Weather\n- Line 48: Nowcasting should be lower-case\n- Equation 1: Consider reformulating this as an objective/loss function.\n- Also Eq. 1: What is $\\hat{X}_t$? What is $\\sigma_C$?\n- Line 93: $\\sigma^2_C = \\sigma^2_C(X_t - C_t)$ doesn't make sense to me.\n- Line 104: *\" same 20 year period that we used for pretraining.\"* .... Do you mean 40 year period? If not, which 20-year period from the 40-year training period did you use?\n- Line 157: Multiple symbols are undefined (e.g. $V_S$).\n- Line 169: It's not entirely clear what \"alternates\" means in this context.\n- Line 429: \"baseline\"... do you mean Prithvi WxC? \n- Line 507: \"improved\"... improved compared to what?\n- Figure 12: Do you mean 'downscale' on the right \"upscale\" block?\n- Sections D. 2.3 and D.2.4 in the appendix are literal copies of the corresponding paragraphs on pages 8 and 9. Please remove." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "### Minor issue: Weak baselines for downscaling\nThe comparison primarily involves interpolation-based methods and does not consider more advanced, AI-driven downscaling models or domain specific statistical/dynamical downscaling." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "### Originality and significance \nPrithvi WxC pushes the foundation model concept in atmospheric science further by expanding beyond just forecasting— seen in earlier models foundation model Aurora. Its architecture and training approach enable it to tackle a variety of downstream tasks, such as downscaling and parameterization, making it a valuable tool for both short-term weather predictions and long-term climate modeling. With its modular design, Prithvi WxC can be adapted flexibly to new tasks that combines AI with physical climate models.\n\n \n### Quality\nThe authors have thoroughly evaluated Prithvi WxC across a range of tasks, including zero-shot reconstruction, downscaling, and extreme event forecasting. The extensive validations across different downstream tasks support Prithvi WxC’s adaptability and effectiveness in diverse weather and climate applications.\n\n\n\n### Clarity and open research\nThe paper is organized in a clear, logical flow, moving smoothly from motivation and background to model architecture, objectives, and results. Key ideas, like the mixed masking and forecasting objective, are presented in a way that makes the technical contributions accessible to both AI and climate science audiences. Code and comprehensive supplementary materials are provided." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Prithvi WxC, a new foundation model designed to support a wide range of weather and climate applications. Built on a large, transformer-based architecture, Prithvi WxC is trained on the extensive MERRA-2 dataset, which covers 160 variables capturing atmospheric data. The model is unique in its ability to address multiple tasks—including forecasting, downscaling, and parameterization—making it versatile in handling both regional and global weather patterns." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "### Reliance on a single reanalysis dataset (MERRA-2)\nMERRA-2, with its relatively lower spatial resolution, is not commonly used in AI-driven weather and climate research, where higher-resolution datasets like ERA-5 are preferred for their superior predictive accuracy. The authors themselves acknowledge this limitation, citing the weaker performance of Prithvi WxC in hurricane track forecasting compared to the ERA-5-trained FourCastNet, attributing this discrepancy to MERRA-2's lower spatial resolution. \n\nPrithvi WxC’s exclusive training on the MERRA-2 reanalysis dataset raises questions about whether it truly qualifies as a foundation model. The narrower training base implies that the model may be learning a representation more specific to MERRA-2 characteristics, along with its biases and errors, rather than capturing a broader, more generalized understanding of weather and climate dynamics. By contrast, Aurora foundation model are pretrained on six diverse weather and climate datasets, including ERA-5, CMCC, IFS-HR, HRES Forecast, GFS Analysis, and GFS Forecasts, which span various sources like forecasts, analyses, reanalyses, and climate simulations [1]. This multi-source approach ensures a broader, more representative foundation that enhances versatility across diverse applications.Prithvi WxC would benefit from a similar multi-dataset training approach to strengthen its robustness and generalizability, which would then qualify it as a foundation model.\n\n[1] Bodnar, C., Bruinsma, W. P., Lucic, A., Stanley, M., Brandstetter, J., Garvan, P., ... & Perdikaris, P. (2024). Aurora: A foundation model of the atmosphere. arXiv preprint arXiv:2405.13063.\n\n### Ablation Study\nWhile the authors propose a novel objective function, they only hypothetically attribute Prithvi WxC's strong short-term forecasting performance to its masking objective, without providing empirical evidence. This lack of testing weakens the claims about the model’s unique architecture and objective function. The observed performance could be influenced by several factors: the mixed objective itself, specific network structures or attention mechanisms, and choices made in the pretraining setup.\n\nWithout ablation experiments, the paper's assertions about the effectiveness of these innovations remain speculative, leaving readers uncertain about the impact of each component. An ablation study could isolate the contributions of these elements and would strengthen the paper by making its claims more concrete and providing clearer insights into Prithvi WxC’s architectural and training contributions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 addition to the weaknesses above, can the authors clarify the following:\n\n1. How is the masking pre-training strategies better than just using variable lead-time that includes delta_t = 0? \n2. How is the local-global attention better than existing pre-training strategies of e.g., patch-variable-level tokenization employed in e.g., ClimaX? \n3. With masking as an additional pre-training strategy, what is the cost-performance tradeoff? \n4. Has the authors measure Prithvi's parameterization stability in an online-coupled setting?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Operationally, the large-scale pre-training of 2.3B FM is impressive in the emerging field of AI4Science. The use of large-scale dataset is also noteworthy. The writing is clear and the downstream tasks are clearly defined, and touch upon some of the hardest challenges facing the field. The use of beyond-atmospheric variables (coupled ocean + land) is also welcomed to build a full Earth system FM." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a new foundation model for weather and climate, called Prithvi WxC, with 2.3 billion parameters, and trained on relatively unconventional yet interesting reanalysis products of MERRA-2. The authors use a relatively novel pre-training strategy in this field, where in addition to forecasting-only pre-training, they also combine masking for reconstruction in the hope of better self-supervision, among several upsides (e.g., natural extension to data assimilation with sparsely-gridded observation). The FM is then evaluated on several downstream tasks, including forecasting, downscaling, and parameterization." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are several major weaknesses in the paper, including non-existent/weak baselines and misleading claims, summarized below:\n\nMajor weaknesses\n1. The term foundation model for this work is misleading, since in contrast to other similar works e.g., ClimaX, Aurora, the model is only trained on one data source which is MERRA-2. The problem: reanalysis products are by construction applicable for short-medium range applications as they are used either to evaluate NWPs or used as ICs for the next forecasting window. The climate dataset being missing, makes the claim of Prithvi being also a climate FM is an overreach at best. \n\n2. Related to the first point, the paper does not evaluate on any climate-related tests despite claiming it to be a climate FM too: (a) is the model stable over 50-100 years climate rollout? (b) what is the climate drift/bias compared to SOTA climate emulator? The paper applies downscaling to a climate dataset CORDEX, but this is less of a climate question and just a general downscaling problem since the former is more concerned about getting the long-term statistics, rather than a high-resolution state realization, correct (which is near impossible to nonlinear chaotic systems such as the Earth system). \n\n3. The downscaling benchmarking is also unacceptable as the baseline is too weak (e.g., bilinear, nearest neighbor). The author should either remove this part or add stronger baselines where the SOTA is at least a deep learning based method. Also, there is no benchmarking on the parameterization downstream task. At least use existing DL-based models from recent works. \n\n4. The forecasting performance appears to be unconvincing at best: with 2.3B parameter, it performs worse than e.g., <50M parameter GraphCast (~40x smaller) even at short lead time of 66 hours (<3 days). Even when the authors mention this result is \"zero-shot\" (which I find it unconvincing since there is rollout fine-tuning still) and the target is different (MERRA-2 vs ERA5), the obvious larger error growth (Figure 4) is alarming as it may not be useful for long-range climate forecasting. Also, why not benchmark against MERRA-FourCastNet in the forecasting task since this provides for a fairer comparison as both are trained with MERRA (as in the case for hurricane prediction). Finally, figure 4c (single line: forecasting cloud) is a case-in-point: the lack of sufficient apple-to-apple baselines where training/eval is done on MERRA-2. \n\nOverall, I find the results unconvincing given the lack of data sources, proper baselines, and inferior performance gain despite Prithvi being orders-of-magnitude larger than any SOTA model. As a side note: I believe the task of downscaling and parameterization is similar in that both attempts to resolve small-scale physics in an otherwise coarse-resolution model. I suggest the authors combine or use different downstream tasks e.g., climate projection." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Architectural flexibility: Can you provide a detailed explanation of the adaptability of the architecture when applied to non rectangular grid systems, such as those used for ocean simulation or polar regions?\nGeneralization strategy: What steps are taken to ensure the performance of the model when training or fine-tuning on datasets outside of MERRA-2, such as ERA5 or even higher resolution datasets?\nHurricane forecast: Can you provide more details on how the hurricane trajectory prediction of this model compares to specific task models such as FourCastNet, especially in areas with sparse data coverage?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Originality: The introduction of basic models for weather and climate applications is novel and important. Unlike specific task models such as FourCastNet and GraphCast (Lam et al., 2022), Prithvi WxC addresses a wide range of tasks and effectively narrows the gap between artificial intelligence models for specific weather tasks and general artificial intelligence base models.\nQuality: The mixed pre training objective of this model combines masking and prediction, which is robust, especially because it uses climate bias rather than just future state prediction, which enhances the adaptability of the model. The model also showed impressive results in zero sample evaluation of reconstruction and autoregressive prediction, outperforming the baseline in a short delivery cycle.\nMeaning: The ability to generalize to multiple downstream tasks, such as downscaling and gravity wave parameterization, suggests that this model has the potential to have a significant impact on weather and climate modeling." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper titled \"Basic Models for Weather and Climate\" introduces Prithvi WxC, a 2.3 billion parameter basic model for various weather and climate tasks. These include downscaling, autoregressive prediction, and extreme event estimation. The model is trained on 160 variables from the MERRA-2 dataset and combines mask reconstruction with prediction tasks to learn from various atmospheric data. Its encoder decoder architecture and ability to work with different spatial topologies make it suitable for global and regional weather modeling." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Generalization to other datasets: Although the model performs well on MERRA-2 data, its generalization ability to other datasets such as ERA5 or CMIP has not been fully explored. Validation on different datasets will better demonstrate its robustness.\nLong term prediction: The accuracy of Prithvi WxC decreases with the extension of the prediction window, especially beyond 66 hours, and its performance is poor compared to models such as Pangu. A deeper investigation into how to maintain performance within an extended time frame can improve its utility in medium - and long-term forecasting." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce a foundation model for weather and climate data. The model is validated on a number of downstream tasks ranging from downscaling (super-resolution) to forecasting." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Foundation Model for Weather and Climate},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xCFdAN5DY3},\nnote={under review}\n}" }, "abstract": { "value": "Triggered by the realization that AI emulators can rival the performance of traditional numerical weather prediction models running on HPC systems, there is now an increasing number of large AI models that address use cases such as forecasting, downscaling, or nowcasting. While the parallel developments in the AI literature focus on foundation models -- models that can be effectively tuned to address multiple, different use cases -- the developments on the weather and climate side largely focus on single-use cases with particular emphasis on mid-range forecasting. We close this gap by introducing Prithvi WxC, a 2.3 billion parameter foundation model developed using 160 variables from the Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2). Prithvi WxC employs an encoder-decoder-based architecture, incorporating concepts from various recent transformer models to effectively capture both regional and global dependencies in the input data. The model has been designed to accommodate large token counts to model weather phenomena in different topologies at fine resolutions. Furthermore, it is trained with a mixed objective that combines the paradigms of masked reconstruction with forecasting. We test the model on a set of challenging downstream tasks namely: Autoregressive rollout forecasting, downscaling, gravity wave flux parameterization, and extreme events estimation." }, "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 models; atmospheric physics; weather; climate; fine-tuning; super-resolution" ] }, "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/bb8fc3ac3b95ec2203f1764d25c0ab9064058aa1.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/ee0ead1d738f3a4a7b0de0952ed83411592b1c33.zip" }, "title": { "value": "A Foundation Model for Weather and Climate" }, "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": "NA" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper utilizes the Diffusion Bridge Model to help the reverse process start from a more deterministic state, reducing the instability caused by noise and thereby facilitating better predictions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a comprehensive framework that encompasses most existing diffusion-based methods. Building on this foundation, the authors introduce the Series-to-Series Diffusion Bridge Model (S2DBM). Experimental results demonstrate that S2DBM delivers superior performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "weakness:\n1. There is a notation issue with \\(\\hat{\\gamma}_t\\) in line 203; the writing needs to be standardized. Additionally, it needs to be clarified whether the values of \\(\\hat{\\alpha}_t\\), \\(\\hat{\\beta}_t\\), and \\(\\gamma_t\\) should have a specific relationship to conform to the diffusion model.\n2. In line 411, it is mentioned that a comparison with timediff was made, but Table 2 does not include TimeDiff data while other baselines are present.\n3. The content of the paper appears to primarily build on existing work by combining the model guidance from TMDM and the non-autoregressive approach of timediff with the existing Brownian Bridge process. This integration seems to lack novelty." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weakness plz." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors provide a comprehensive summary of existing models, highlighting that their primary differences lie in the formulation of $\\hat{\\gamma_t}$. By introducing the Brownian bridge process into diffusion-based time series forecasting models, they establish a range of relevant properties." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper revisits the application of diffusion models for time series forecasting, presenting a unified framework that consolidates these methods. Building on this framework, the authors incorporate the Brownian bridge process to enhance prediction accuracy. The results demonstrate that the proposed approach outperforms existing diffusion-based forecasting models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Line 259, the title of Proposition 1 is \"Brownian Bridge between Historical and Predicted Time Series.\" However, the Brownian bridge’s endpoint is set to $h$, the Prior Predictor’s forecasted value of $x$. The paper provides no explanation as to why $h$ is chosen as the endpoint for the Brownian bridge. \n2. It would be beneficial for the authors to clarify, from a theoretical perspective, why the Brownian bridge is integrated into the diffusion model for time series forecasting. Specifically, how does its ability to \"pin down\" the diffusion process at both ends help reduce instability from noisy inputs and enable the accurate generation of future features based on historical time series?\n3. In line 195, Theorem 1 concerns non-autoregressive diffusion processes. However, the summarized models, CSDI and SSSD, do not appear to be non-autoregressive models. Could there be an issue with the theorem here?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Q1.\nCan you elaborate on the computational efficiency of $S^2DBM$, particularly in relation to larger datasets or real-time forecasting applications?\n\nQ2.\nWhat specific advantages does the Brownian Bridge process offer over other stochastic processes for time series forecasting in the context of your model?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1.\n$S^2DBM$ integrates the Brownian Bridge process into time series forecasting using diffusion models.\nBy redefining the diffusion framework, the authors introduce a new model and consolidate various non-autoregressive diffusion techniques into a comprehensive framework, elucidating their interrelationships and underlying principles. \n\nS2.\nThe authors conduct thorough theoretical groundwork. \nThe empirical evaluations are robust, utilizing diverse real-world datasets to benchmark the model's performance against SOTA methods.\n\nS3.\nThe draft is well-structured and organized. \nThe introduction succinctly outlines the problem context and motivates the need for the proposed model." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents the Series-to-Series Diffusion Bridge Model ($S^2DBM$), a promising approach in time series forecasting using diffusion models. \nTraditional diffusion models often struggle with deterministic point-to-point predictions. \n$S^2DBM$ addresses this by leveraging the Brownian Bridge process to reduce noise and improve accuracy in reverse estimations, effectively capturing temporal dependencies in time series data. \nThe model incorporates historical data as informative priors, stabilizing diffusion and enhancing point-to-point prediction capabilities. \nExperimental results on various datasets show that $S^2DBM$ outperforms existing diffusion-based and other state-of-the-art time series models, demonstrating superior accuracy in both deterministic and probabilistic forecasting tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. \nSome assumptions and derivations in this work could be better justified. \nFor example, the choice of using a Brownian Bridge process warrants a more detailed discussion on why it is preferred over other stochastic processes. \nProviding empirical evidence or theoretical reasoning for the choice of this process could strengthen the argument for its effectiveness in reducing randomness in forecasting.\n\nW2.\nWhile the paper presents strong performance metrics, it lacks robustness testing under adverse conditions, such as noisy inputs or missing data. \nFuture experiments should include scenarios with varying levels of data quality to demonstrate how $S^2DBM$ handles real-world challenges.\n\nW3.\nCurrently, the experiments focus primarily on a single configuration. \nInvestigating how different settings of hyperparameters, such as the choice of the prior predictor or conditioning modules, impact performance could provide valuable insights. \nFor instance, evaluating the effects of varying the number of diffusion steps or using alternative conditioning mechanisms could highlight the robustness of $S^2DBM$ across diverse scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In Table 2, your models do not achieve the best performance across all datasets. Could you explain why linear models perform best in several cases, while your models do not?\n2. Since $c = E(x)$ and $h = F(x)$ are two important components of your algorithms, I am curious about the role of F in enhancing performance. What impact does it have? An ablation study of F may help." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The presentation of this paper is clear and easy to follow, with algorithms well articulated.\n2. The paper provides a unified framework for several existing diffusion-based time series models, which is well-supported by theoretical foundations. The insights offered are impressive.\n3. Both point-to-point and probabilistic forecasting are addressed in the reverse process, broadening the algorithm's application scope." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses point-to-point time-series forecasting through the use of diffusion models. The authors propose a unified non-autoregressive framework that encompasses most existing diffusion-based time-series models. Building on this framework, they introduce S2DBM, which incorporates the Brownian Bridge process. By integrating historical information via informative priors and conditions, S2DBM effectively reduces randomness and enhances forecasting accuracy." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Some expressions reference prior work without sufficient explanations, which can hinder comprehension. For instance, in the last equation of Section 3.1, the introduction of \\(y_\\theta\\) relies solely on a citation, making it time-consuming for readers to fully grasp its meaning.\n2. Table 4 presents results only for a horizon of 96, omitting long horizon settings. This limits the analysis and may leave important insights unaddressed." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Series-to-Series Diffusion Bridge Model" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024seriestoseries,\ntitle={Series-to-Series Diffusion Bridge Model},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xCMmtYOsiL},\nnote={under review}\n}" }, "abstract": { "value": "Diffusion models have risen to prominence in time series forecasting, showcasing their robust capability to model complex data distributions. However, their effectiveness in deterministic predictions is often constrained by instability arising from their inherent stochasticity. In this paper, we revisit time series diffusion models and present a comprehensive framework that encompasses most existing diffusion-based methods. Building on this theoretical foundation, we propose a novel diffusion-based time series forecasting model, the Series-to-Series Diffusion Bridge Model ($\\mathrm{S^2DBM}$), which leverages the Brownian Bridge process to reduce randomness in reverse estimations and improves accuracy by incorporating informative priors and conditions derived from historical time series data. Experimental results demonstrate that $\\mathrm{S^2DBM}$ delivers superior performance in point-to-point forecasting and competes effectively with other diffusion-based models in probabilistic forecasting." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "time series forecasting; diffusion 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/4087bf50c532889a275e543287df8467c428f536.pdf" }, "presentation": null, "primary_area": { "value": "learning on time series and dynamical systems" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/f00be8e99421c0fef9c2114d6346425f20b36f7d.zip" }, "title": { "value": "Series-to-Series Diffusion Bridge Model" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 are the “difficult goals” sampled, how many are sampled, etc? Why are thy chosen in this way?\n\nAre there other experiments done (e.g. something like the perturbation test, or a test where the single-goal policy shows useful representations or reasonable success on goal states outside of those in the goal path) which would alleviate concerns that this method is simply overfitting to solve for this one goal?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Under the conditions set by the work, i.e. learning single difficult goals without rewards, demonstrations, curriculums, etc, the single-goal CRL clearly outperforms other methods. Furthermore, this method reduces assumptions compared to standard RL. Finally, an important finding of this paper is that a policy can be trained under contrastive learning to find a difficult goal without providing intermediate easier goals." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work aims to solve goal-conditioned RL (GCRL) tasks where only one goal is desired without access to rewards (dense or sparse), curriculums, or demonstrations. It proposes modifying standard Contrastive RL (CRL) by using only a single goal in rollouts while still using multiple goals while training. The paper runs experiments to show that the specific combination of single-goal rollouts, multi-goal training, and contrastive representations are needed to achieve good results. The paper also includes some additional analysis to examine single-goal CRL under impossible goals and to explore the representations learned by CRL." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Firstly, it is unclear how the authors selected or defined “difficult goals”. Are the difficult goals sampled by failure cases of CRL? How many difficult goals are the results evaluated on? For example, if this method is effective on one particular goal, but not others, then it would not be effective in practice.\n\nSecondly, the authors claim that the first ablation disproves the hypothesis that single-goal CRL “learns useful representations and an effective policy only for states along that goal path” (line 476). However, since the single-goal CRL fills the replay buffer with rollouts where the policy aims to reach goal $s^*$, and the actor is only being updated with states from the replay buffer, how does this prove that the single-goal rollout + multi-goal training policy is not overfit to/does not only contain good representations for states along the goal path?\n\nFinally, Fig. 11 shows that a monolithic critic CRL (Range of Difficulties) can reach approximately 60% success rate on the single hard goal on Sawyer Box after 100,000 steps. However, Fig. 3 shows that standard CRL (range of difficulties) can only achieve up to approximately 40% success rate on Sawyer Box after 1,000,000 steps. It seems this might contradict the assertion that the separate representations are important? (It is possible that the methodology for selecting the “difficult goals” is impacting these results.)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Do the authors have any statistics on eval performance over the same distribution of goals used to generate the replay buffer? One natural argument is that, in so far as each goal can be viewed as a separate task, the model will be best at the goal distribution that appears in the replay buffer. I believe this is distinct from the overfitting experiments, because in those experiments the final evaluation number is still only the final hard eval goal." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The fact that a single hard goal is sufficient with contrastive RL is certainly surprising, and the paper is upfront in not having a good explanation why this occurs. The fact it occurs over a few environments provides some evidence it is not simply a one-off occurrence." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors describe a curious phenomenon observed in some simulated robot manipulation environments used for RL. When using contrastive RL, using a single hard goal (one far away from the robot corresponding to task success) led to better learning outcomes than using a human designed curriculum of easy and hard goals.\n\nThis is observed across 4 environments, a Sawyer bin placing task, box placing task, peg insertion, and a 2d maze navigating task.\n\nThe authors provide a number of ablation experiments trying to study why this phenomenon occurs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am suspicious that the results are too good - that there was no environment where using multiple goals performed best. Broadly my concern is that maybe these environments are too easy. If the model is able to succeed at the task given just the final hard goal, perhaps it's too hard to design a good dense reward or curriculum of goals to speed up learning. This wouldn't be too surprising, it's often been remarked that good partial reward design is difficult.\n\nI'm also not sure how \"single goal\" the final method is. In particular, Figure 10's caption was confusing to me. It seemed to suggest that in their method, the actor loss uses multiple goals, rather than a single goal? If so, this doesn't really seem like \"1 goal is all you need\". Essentially I think the authors may be overgeneralizing their conclusions.\n\nMy understanding so far is this:\n\n* A contrastive critic is learned via contrastive RL, defining reward by $\\phi(s,a)^T \\psi(s_f)$, where $s_f \\sim Geom(1-\\gamma)$ steps in the future\n* When generating data, we use a single hard goal $s^*$ and act according to $\\pi(a|s,s^*)$\n* When updating the actor, we sample a trajectory from the replay buffer of data generated according to $\\pi(a|s,s^*)$. But, for each initial $s$, we then sample $s_f \\sim Geom(1-\\gamma)$ within the trajectory, and apply gradient updates as if we collected data according to $\\pi(a|s,s_f)$, even though we actually collected data according to $\\pi(a|s,s^*)$.\n\nIn which case, at most you could say 1 goal is the only requirement for data collection, but the policy still needs to be trained on every intermediate goal observed to achieve good 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": "Thank you for contributing this work. When reading the paper the emergent skill learning argument felt like it needed further justification. \n\n\n(1) I wished to clarify if the following pattern of skill emergence always held? The paper suggests yes but I wished to confirm this point. \n\n\"For example, in all three environments, the agent (1) first learns how to move its end-effector to\nvarying locations, (2) then learns how to nudge and slide the object, and (3) finally learns to pick up\nand direct the object. \"\n\n(2) I also wished to ask when you mention fixed sampling intervals for checkpoints, what fixed interval was used? Were there any exceptions to the skill learning trends identified in (1) (e.g. checkpoints with purely random behaviours) and if so how frequent were exceptions?\n\nIn answering (1) and (2) it would be useful to provide:\n\n(a) Video's indicating the progression of skill learning with details of the checkpoints, this would be helpful for validating the qualitative behaviours claimed within the paper. \n\n(b) Not entirely necessary but it would be nice to have quantitative metrics for characterising the various skills. For instance when you mention learning how to move the end-effector, distinguishing between purely random movements and directed movement would be useful, there are certain quantitative metrics that could be used, such as correlations between actions. Contrasting this with naive exploration approaches such as SAC may yield insights and further validate the skill learning claim. Similarly for learning to interact with objects and eventually pick them, characterising these skills quantitatively would be useful. \n\n(c) It may be possible to perform clustering on policy rollouts data to identify distinct exploratory behaviours as learning progresses. This would be a valuable addition." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The authors provide valuable insights into the effectiveness of training a policy conditioned on a single goal with contrastive reinforcement learning in the single-task setting. \n\n- The authors contribute to the discussion on skill learning through empirically demonstrating the emergence of skill learning with contrastive objective functions.\n\n- The authors validate each of their claims and provide discussions on counter arguments (e.g. overfitting single-task) that support their empirical results. \n\n- The authors provide sufficient information and high-quality materials to reproduce their results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work builds upon previous work in contrastive reinforcement learning. What's unique to this work is how their learning algorithm conditions on a single goal; this has implications on the distribution of data within the replay buffer which ultimately impacts the learning dynamics and exploration behaviour of the agent. The authors argue that this approach results in emergent skill learning without explicitly having to define a reward function or learning curriculum for the task being solved. This argument is supported with simulated experiments on robotic manipulation and maze traversing benchmarks. The performance of their approach is compared to contrastive reinforcement learning using a range of human-specified goal difficulties and existing approaches (SAC+HER and RIS). The authors justify their claim with empirical results, leading to the insight that contrastive reinforcement learning with conditioning on a single goal demonstrates advantageous exploration and skill learning behaviours relative to existing approaches that seek to learn effective exploration and skill learning strategies." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There could be more analysis of emergent skill learning, in the current draft the authors mention that fixing a checkpoint and examining the qualitative behaviour of the policy they see evidence of skill learning, further justification of this claim is required in my opinion. \n\n- The authors highlight an awareness of this point but further analysis on why this strategy works would be useful. \n\n- It would be interesting to see if this approach results in exploration strategies that are capable of solving more complex tasks. For example tasks that encounter Sussman's anomaly, where naive subgoal design may not work well. I am unsure how this approach would perform but it would be promising if it did well and might further justify the strength of this more general approach to learning to solve tasks with reinforcement learning. Also tasks with significant bottlenecks in which naive exploration strategies are often incredible sample inefficient, does this approach help improve performance on such tasks?\n\n- While the authors address the concern of overfitting to the individual task being solved, they do not explore how well the learnt representations generalise. While there are examples of solving tasks with minor perturbations there isn't a discussion of more significant changes to the task environment. From a practical standpoint, it would be interesting to understand how well this approach can be used to learn value representations that generalise well across task environments. \n\n- This work suffers from ailments of reinforcement learning approaches more generally, most especially sample efficiency. It would be interesting to understand if the representations learnt in this work can be leveraged to address the issue of sample efficiency when learning multiple tasks (e.g. on the metaworld benchmark)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024a,\ntitle={A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive {RL} without Rewards, Demonstrations, or Subgoals},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xCkgX4Xfu0},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we present empirical evidence of skills and directed exploration emerging from a simple RL algorithm long before any successful trials are observed. For example, in a manipulation task, the agent is given a single observation of the goal state (see Fig. 1) and learns skills, first for moving its end-effector, then for pushing the block, and finally for picking up and placing the block. These skills emerge before the agent has ever successfully placed the block at the goal location and without the aid of any reward functions, demonstrations, or manually-specified distance metrics. Once the agent has learned to reach the goal state reliably, exploration is reduced. Implementing our method involves a simple modification of prior work and does not require density estimates, ensembles, or any additional hyperparameters. Intuitively, the proposed method seems like it should be terrible at exploration, and we lack a clear theoretical understanding of why it works so effectively, though our experiments provide some hints." }, "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": [ "exploration", "emergent skills", "contrastive reinforcement learning", "open-ended 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/860dad66264a3d8e8e9b99d37f80c9d08731dfa7.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": "A Single Goal is All You Need: Skills and Exploration Emerge from Contrastive RL without Rewards, Demonstrations, or Subgoals" }, "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": { "value": "No ethics review." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "I would like to hear the opinion of the authors on the concerns I raised in the weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is well-written and easy to follow. \n- It presents an easy solution to the sampling problem of diffusion models that only requires limited training time while obtaining. \n- The soft teacher loss is effective and simple to implement. \n- The evaluation is thorough and includes multiple models, multiple datasets, and multiple sampling strategies.\n\nIn general, I liked the paper and I lean toward acceptance. However, since this is not my area of expertise, I would wait for the discussion with the authors and other reviewers to increase the score to Accept and recommend borderline Accept for now." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a sampling method for diffusion models in order to reduce the sampling time required to generate an image. The paper proposes to learn the sampling steps from a teacher model that accurately solves the ODE by taking small step sizes. Extensive experiments show the effectiveness of the method while only requiring small amounts of training time." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although I liked the paper, there are some concerns that, if addressed, would improve the paper. In the following paragraphs, I describe my concerns in detail:\n\n- In the table with the main results, sometimes it is not clear what the metrics are computed against. I suppose the metrics in table 2, 3, 4, and 5 are computed against random samples of the model using the accurate estimation of the ODE. However, if this metric is computed against the true distribution, the performance of the teacher with the accurate computation of the ODE should be shown (1000 steps). I think the evaluation protocol needs to be more clearly defined.\n\n- In a similar direction, Table 6 shows the performance of a teacher model using 8 steps. Why only 8 steps are used here? Would not the teacher use a higher number of samples?\n\n- The model used is quite simple being only composed of a single vector (or two in the decoupled version). From the results in Table 7, increasing the number of parameters leads to better results. Would increasing the complexity of the model lead to better results?\n\n- In the limitations section I found missing that the proposed method needs to be retrained for the target number of sampling steps. One model trained to generate images with 2 samples, would not be useful for 3 and a new model would need to be trained. This might be a problem since different images might necessitate a different number of steps to achieve good quality." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 invertibility assumption in Theorem 1 be relaxed and still achieve an upper bound on the KL divergence?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The LD3 algorithm is extremely lightweight, requiring only 100 samples and less than 1 hour on a single GPU to learn optimized sampling schedules.\n- The method is evaluated on a comprehensive set of pretrained models and compared against several baseline, showing improved quality in the majority of cases\n- A proper ablation study is done on the various choices/hyperparameters." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper tackles the challenge of optimizing timestep schedules for sampling in diffusion and flow-based generative models. The authors propose optimizing the schedule in order to minimize the global truncation error of multi-step sampling by reducing the discrepancy between a high-NFE teacher model and a low-NFE student model. They demonstrate that a relaxed version of this optimization problem yields even better results. Impressively, their proposed algorithm runs very quickly, achieving convergence in under 1 hour on a single GPU. A comprehensive evaluation is done using multiple pretrained diffusion models and ODE solvers. The proposed algorithm is compared against several hand-designed and learnt sampling schedules, and is shown to considerably improve the image quality in the low-NFE regime." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There are several typos in the paper. See some examples below:\n - Algorithm 1 Line 6: $x'_T ← x'_T + ...$ must be $x'_T ← x_T + ...$\n - Line 251: $x'T \\rightarrow x'_T$\n - Line 251: $\\Psi*(x_T) \\rightarrow \\Psi_*(x_T)$\n- Theorem 1 requires more explanation on its invertibility assumption. Specifically, if the NFE is small, functions $\\Psi_*, \\Psi_\\xi$ invertibility is a non-trivial fact which requires some justification on its assumption. \n- The method relies on a learned perceptual distance (LPIPS) to achieve optimal results, as shown by the significant quality drop in Table 7 when switching to a standard Euclidean loss. This raises questions about how well the method might generalize to other data types beyond images." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Near the introduction, the paper suggests the approach should be seen as complementary to distillation methods as opposed to going head-to-head, but the goal of both said approaches and LD3 is to achieve good sample quality with the least number of steps possible. Why do the authors argue this is the case? It is not clear to me (and it is not demonstrated in the paper) that distillation methods are compatible with optimizing the timesteps used, e.g. progressive distillation is trained to match two steps of the teacher with the student, so it's not clear a priori whether changing the timesteps later will completely break a model distilled this way." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper includes proofs of soundness of their proposed minimization objectives, going beyond purely empirical contribution.\n- The number of experiments is substantial across both datasets, baselines from prior work, and choice of pretrained models.\n- The experiments conducted include notoriously difficult datasets in the literature of diffusion step reduction like ImageNet, and shows improvement in more complex settings such as a text to image model.\n- The objective is cheap to train compared to prior work; the key being that a very low batch size of 2 is permissible to use.\n- Ablation studies demonstrate the importance of the proposed changes separately.\n- The samples presented qualitatively look very reasonable and show clear improvement over the usual hand-crafted timestep schedules, and they fix the random seed so the same samples can be comapred." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes LD3, a method to discover timesteps that will yield good samples for diffusion models when the number of forward passes is very small at inference time, through optimization. The authors propose a distillation-like objective where a global notion of distance between the original model and a student model with learned timesteps is minimized. Two variants are presented, one \"soft\" objective achieving best results that only requires samples to be within a ball of hyperparameter radius $r$ to the teacher's samples. The soft objective additionally is linked theoretically to also upper bounding the KL divergence between the two models. Various experiments are conducted to demonstrate that LD3 achieves better FID scores than prior methods on few-step sampling, and an ablation study is included to demonstrate the importance of different components of LD3, the most important seemingly being the decoupling of timestep choice and step size, and the use of a perceptual distance (LPIPS) as opposed to pixel-based L2." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There are two main areas around which the paper could be much stronger. The first is in comparisons to distillation methods, which are among the strongest in the literature. The paper includes a comparison to progressive distillation and consistency distillation in Table 9, but it is really difficult to compare these methods apples-to-apples. There are details missing (please correct me if I missed these e.g. in the supplementary material) such as what models were compared and where are the baseline scores taken from; ideally the same model should be post-trained with the different techniques. The number of forward passes across methods also doesn't match, making it difficult to draw any conclusions. One conclusion that can be drawn, however, is that progressive distillation remains better than LD3 in FID score at NFE=8, albeit requiring much more compute to distill.\n\nThe other major weakness is the lack of careful qualitative comparisons to other step reduction methods. The vast majority of the qualitative samples are compared to hand-crafted schedules, which are the weakest baselines. This is really important, especially because prior work has shown that very low FID scores can be achieved somewhat adversarially, resulting in strange samples (e.g., consider the CIFAR10 samples in the GGDM paper), so quantitative results are insufficient to truly demonstrate that LD3 improves over all prior work. Careful side-by-side comparisons of different step reduction methods, derived from the same pre-trained model and using the same initial noise and matching NFE would be significantly more convincing.\n\nOverall, the work is strong, and the quantitative results already put this paper as a valuable contribution to the literature that should be accepted at the conference. I am opting for a weak accept, because the comparisons to distillation methods seem improper and incomplete, and the qualitative comparisons require more care. But even if so, due to the very low cost of the proposed technique and the achieved scores, the work already has intrinsic value. I strongly encourage the authors to address the concerns outlined above as it would make the work excellent." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning to Discretize Denoising Diffusion {ODE}s},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xDrFWUmCne},\nnote={under review}\n}" }, "abstract": { "value": "Diffusion Probabilistic Models (DPMs) are generative models showing competitive performance in various domains, including image synthesis and 3D point cloud generation. Sampling from pre-trained DPMs involves multiple neural function evaluations (NFE) to transform Gaussian noise samples into images, resulting in higher computational costs compared to single-step generative models such as GANs or VAEs. Therefore, reducing the number of NFEs while preserving generation quality is crucial. To address this, we propose LD3, a lightweight framework designed to learn the optimal time discretization for sampling. LD3 can be combined with various samplers and consistently improves generation quality without having to retrain resource-intensive neural networks. We demonstrate analytically and empirically that LD3 improves sampling efficiency much less computational overhead. We evaluate our method with extensive experiments on 7 pre-trained models, covering unconditional and conditional sampling in both pixel-space and latent-space DPMs. We achieve FIDs of 2.38 (10 NFE), and 2.27 (10 NFE) on unconditional CIFAR10 and AFHQv2 in 5-10 minutes of training. LD3 offers an efficient approach to sampling from pre-trained diffusion models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Diffusion models", "Efficient Sampling", "Ordinary Differentiable Equations" ] }, "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/ec79ba5865f4c12903666e09b0b7ad920bd6b6d5.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Learning to Discretize Denoising Diffusion ODEs" }, "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 \"Weaknesses\"." }, "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 MKRAG (Multi-grained Knowledge Retrieval-Augmented Generation) for hyper-long context question answering. By integrating multi-grained entity graphs with an iterative retrieval and reasoning process, MKRAG addresses the limitations of traditional models constrained by context window size.\n2. This paper introduces LoopAgent, an iterative retrieval framework that progressively refines queries across multiple retrieval cycles. By incorporating advanced reasoning capabilities, LoopAgent improves both retrieval and answering accuracy and mitigates information loss in traditional single-pass retrieval methods, particularly in complex multiple-entity scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a multi-grained entity graph-based QA method that constructs an entity graph and dynamically combines both local and global contexts, capturing information across three granularity levels: micro, feature, and macro levels, and incorporates iterative retrieval and reasoning mechanisms to generate accurate answers for hyper-long contexts. Evaluation results on LongBench and InfiniteBench demonstrate the effectiveness of the approach, significantly outperforming existing methods in both the accuracy and granularity of the extracted answers, and it can be deployed in online novel-based applications." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The entities and their associated attributes, relationships, and events are extracted by LLMs. However, as noted in previous work, LLMs may fall short in information extraction (IE) tasks, such as entity extraction, relation extraction, and event extraction. If LLMs cannot handle IE well, errors could propagate through the system, leading to random, unpredictable, and non-generalizable outcomes. It could be better if the authors provide evaluation on the reliability of the IE process.\n2. While the model of experiments is based on ERNIE, there lack a comparison with the other ERNIE variants capable of handling longer inputs, such as ERNIE-Turbo-128K. Including a comparison with models from the same series would strengthen the paper by better demonstrating the effectiveness of the proposed approach." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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.\tHow is \"importance weight\" defined in equation (6)? More explanation of this part would be helpful.\n2.\tWhat is the module performance in each subtask? e.g., the recall rate of entities, accuracy of entity aggregation, effectiveness of entity pruning, effectiveness of thresholds in equation (10), and more?\n3.\tWhat is the average number of rewrites for each question (line 401)? How is the computation cost for the LoopAgent?\n4.\tThe hyter-long context QA benchmark, InfiniteBench, has an average output of 6.3 tokens, which should be only a few words. Does these questions really need query-decomposition described in line 383?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The EntiGraph module effectively enhances the representation and may have the potential to extract most necessary information from the document.\n2. The multi-grained knowledge generation from entity graph helps capture possible QA from different granularity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To address the challenges of hyper-long context QA, particularly the limitations of context windows and retrieval errors in retrieval-augmented generation (RAG) due to inadequate semantic representation, this paper proposes Multi-grained Knowledge RAG Method (MKRAG). MKRAG involves extracting entities from context with EntiGraph chunk-by-chunk, generating multi-grained QA pairs, and iteratively retrieve related information by refining queries to get the final answer.\n\nResults show that the proposed method achieve on par or slightly better performance compared with SOTA models/methods, and achieves high performance gain in the online test." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThis paper claims that the EntiGraph module enables accessing nearly all the necessary information. However, the performance of EntiGraph is not discussed.\n2.\tIn the Entity Pruning section, how the threshold $\\tau$ is selected is not explained. There are also pre-defined thresholds in (10). The selection of thresholds are not justified by detailed analysis of module performance.\n3.\tBy turning the document into entity graphs and generating knowledge accordingly, the model risks missing information during these processes, and the agent may be unable to answer the question given the insufficient information.\n4.\tIn the ablation study section, “Chuck Retrieval + Baseline” setting uses a chunk size of 500 tokens/chunk while the context window limit is 4k. It is not clear what is the chunk size used in MKRAG and if the small chunk size in the previous setting limits its performance.\n5.\tThis paper claims that the proposed method demonstrating high accuracy in Online Test and high user satisfaction rate. However, details of the tests are not provided and the results are not compared against other SOTA models/methods." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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.Authors have not clearly stated the key innovations of this paper. Authors to explicitly state their key innovations and provide a clear comparison to existing methods, highlighting specific novel aspects of their approach, e.g., a specific new method or a new framework of well incorporating existing techniques.\n2. The multi-grained entity graph and iterative retrieval, while effective, could be computationally intensive, limiting scalability in resource-constrained environments. Therefore, authors should conduct time and space complexity analyses, as well as perform corresponding experiments for verification, such as the Inference Time and the Time per Iteration.\n3.Some experimental details are missing, such as hardware specifications, number of iterations, the stopping condition for iterative retrieval, and the convergence criteria for all models.\n4. About \" we employ context aggregation algorithms to integrate both local and global contexts of the same entity, and utilize an LLM, EntiGraph, to generate multi-grained QA pairs (i.e., micro-level, feature-level, and macro-level). This multi-level strategy not only ensures that the model produces highly accurate answers for complex queries, but also mitigates the fragmentation of information that often hampers traditional methods\", 1) the \"EntiGraph\" is a LLM model? 2) why generate multi-grained QA pairs from micro-level, feature-level, and macro-level, and what are the insights about these? 3) why this multi-level strategy can ensure high accurate answers? 4) does this strategy lead to much noise?\n5. The model is relatively complex, and the paper fails to clearly present the training and optimization process of the model. 1）How to perform joint optimization among multiple modules, i.e., ITERATIVE RETRIEVAL AGENT, MULTI-GRAINED REPRESENTATION, ENTITY EXTRACTION? How is the training data for each module constructed? For each entity, how to construct the embeddings for its attributes, relationships, events and temporal information?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.The MKRAG model captures the local and global information in the text by constructing multi-grained entity graphs (including micro, feature and macro levels), and can generate more accurate answers than traditional RAG methods.\n2. Overall well written and easy to understand.\n3.Iterative retrieval agent, MKRAG model uses LoopAgent iterative retrieval mechanism to refine the query through multiple rounds of retrieval to alleviate the problem of information fragmentation encountered by traditional methods in dealing with ultra-long text.\n4.In this paper, based on the hyper-long context question and answer field, the experiments not only aim at the long text, but also verify that the proposed model is also applicable in the short text." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, a model called MKRAG is proposed to deal with hyper-long context QA tasks. Through multi-grained knowledge generation and iterative retrieval agent, MKRAG model can effectively extract and integrate information, and thus improves the accuracy of question answering. The experimental results show that the MKRAG model achieves excellent performance on multiple benchmark datasets, and shows a strong ability of long text understanding and multi-domain question answering." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.Authors have not clearly stated the key innovations of this paper. Authors to explicitly state their key innovations and provide a clear comparison to existing methods, highlighting specific novel aspects of their approach, e.g., a specific new method or a new framework of well incorporating existing techniques.\n2. The multi-grained entity graph and iterative retrieval, while effective, could be computationally intensive, limiting scalability in resource-constrained environments. Therefore, authors should conduct time and space complexity analyses, as well as perform corresponding experiments for verification, such as the Inference Time and the Time per Iteration.\n3.Some experimental details are missing, such as hardware specifications, number of iterations, the stopping condition for iterative retrieval, and the convergence criteria for all models.\n4. About \" we employ context aggregation algorithms to integrate both local and global contexts of the same entity, and utilize an LLM, EntiGraph, to generate multi-grained QA pairs (i.e., micro-level, feature-level, and macro-level). This multi-level strategy not only ensures that the model produces highly accurate answers for complex queries, but also mitigates the fragmentation of information that often hampers traditional methods\", 1) the \"EntiGraph\" is a LLM model? 2) why generate multi-grained QA pairs from micro-level, feature-level, and macro-level, and what are the insights about these? 3) why this multi-level strategy can ensure high accurate answers? 4) does this strategy lead to much noise?\n5. The model is relatively complex, and the paper fails to clearly present the training and optimization process of the model. 1）How to perform joint optimization among multiple modules, i.e., ITERATIVE RETRIEVAL AGENT, MULTI-GRAINED REPRESENTATION, ENTITY EXTRACTION? How is the training data for each module constructed? For each entity, how to construct the embeddings of its attributes, relationships, events and temporal information?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See above" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The multi-grained QA method integrates a graph-based representation of entities and iterative retrieval, allowing the model to retain contextual coherence and retrieve relevant answers from hyper-long contexts.\n2. The inclusion of LoopAgent, an iterative retrieval mechanism, improves QA accuracy by refining the search process across multiple rounds, which is particularly beneficial for complex and nuanced queries.\n3. The model shows state-of-the-art performance on both hyper-long and moderately long datasets, outperforming baseline models like GPT-4 and other RAG-based approaches on LongBench and InfiniteBench.\n4. The model has been effectively deployed in real-world applications, such as online novel-based platforms, and demonstrates enhanced scalability and practical utility in handling real-time queries." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses a challenge in hyper-long context QA, where retrieving precise answers from extensive and dispersed content poses substantial obstacles. Current approaches often suffer from limitations, such as input-length constraints in LLMs and semantic loss in RAG systems. The authors propose a multi-grained entity graph-based QA framework, termed MKRAG, that operates across three granularity levels—micro, feature, and macro—to improve information extraction and reasoning across hyper-long texts. Their framework also introduces an iterative retrieval mechanism, LoopAgent, designed to refine retrievals and improve accuracy through multiple rounds. The evaluations across datasets show that MKRAG achieves state-of-the-art results, particularly excelling in scenarios with high granularity requirements, such as long-tail or detail-oriented queries." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although the model circumvents some LLM limitations, the use of a multi-granularity framework and iterative retrieval adds complexity and computational demands, which may be prohibitive for broader real-time applications. How to evaluate the efficiency of the proposed method?\n2. The approach heavily relies on accurate entity extraction and structured graph relationships. In cases where entity relationships are sparse or ambiguous, the model's performance may degrade. Do the authors test other entity extraction methods other than EntiGraph?\n3. The paper should further discuss trade-offs between different granular levels and how the system decides the optimal level of granularity in real time, especially in cases with sparse information.\n4. There are many prompt/context compressing methods that should be included as baselines:\n4.1. Jiang, Huiqiang, et al. \"Llmlingua: Compressing prompts for accelerated inference of large language models.\" arXiv preprint arXiv:2310.05736 (2023).\n4.2. Jiang, Huiqiang, et al. \"Longllmlingua: Accelerating and enhancing llms in long context scenarios via prompt compression.\" arXiv preprint arXiv:2310.06839 (2023).\n4.3. Pan, Zhuoshi, et al. \"Llmlingua-2: Data distillation for efficient and faithful task-agnostic prompt compression.\" arXiv preprint arXiv:2403.12968 (2024).\n4.4. Li, Yucheng, et al. \"Compressing context to enhance inference efficiency of large language models.\" arXiv preprint arXiv:2310.06201 (2023)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper presents a Multi-Grained Retrieval-Augmented Generation (MGRAG) method for hyper-long context question answering, integrating multi-grained entity graph with iterative retrieval and reasoning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024multigrained,\ntitle={Multi-Grained Knowledge for Retrieval-Augmented Question Answering on Hyper-long Contexts},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xE3Ra2GTpX},\nnote={under review}\n}" }, "abstract": { "value": "In the task of hyper-long context question answering (QA), a key challenge is extracting accurate answers from vast and dispersed information, much like finding a needle in a haystack. Existing approaches face major limitations, particularly the input-length constraints of Large Language Models (LLMs), which hinder their ability to understand hyper-long contexts. Furthermore, Retrieval-Augmented Generation (RAG) methods, which heavily rely on semantic representations, often experience semantic loss and retrieval errors when answers are spread across different parts of the text.\nTherefore, there is a pressing need to develop more effective strategies to optimize information extraction and reasoning. \nIn this paper, we propose a multi-grained entity graph-based QA method that constructs an entity graph and dynamically combines both local and global contexts. Our approach captures information across three granularity levels (i.e., micro-level, feature-level, and macro-level), and incorporates iterative retrieval and reasoning mechanisms to generate accurate answers for hyper-long contexts.\nSpecifically, we first utilize EntiGraph to extract entities, attributes, relationships, and events from hyper-long contexts, and aggregate them to generate multi-granularity QA pairs. Then, we retrieve the most relevant QA pairs according to the query. Additionally, we introduce LoopAgent, an iterative retrieval mechanism that dynamically refines queries across multiple retrieval rounds, combining reasoning mechanisms to enhance the accuracy and effectiveness of answering complex questions.\nWe evaluated our method on various datasets from LongBench and InfiniteBench, and the experimental results demonstrate the effectiveness of our approach, significantly outperforming existing methods in both the accuracy and granularity of the extracted answers. Furthermore, it has been successfully deployed in online novel-based applications, showing significant improvements in handling long-tail queries and answering detail-oriented questions." }, "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": [ "Knowledge-based Question Answering", "Retrieval-Augmented", "Large Language Model Generation", "Information Extraction", "Hyper-long Contexts" ] }, "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/a439494416d2ea379ddf8350a49fa98f3d4a6eb9.pdf" }, "presentation": null, "primary_area": { "value": "generative models" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Multi-Grained Knowledge for Retrieval-Augmented Question Answering on Hyper-long Contexts" }, "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 Weakness." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "HACk-OOD introduces a unique method using hypercone projections to delineate class contours, avoiding traditional Gaussian distribution assumptions and offering greater flexibility in complex feature spaces. The method achieves competitive, often superior, results on challenging datasets like CIFAR-100, demonstrating strong performance in both near and far OOD detection scenarios." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "HACk-OOD is a post-training OOD detection method using hypercone projections to construct class-specific contours in embedding space. The approach achieves SOTA performance on CIFAR-100 and improves with larger networks. Including Imagenet experiments could further demonstrate scalability on large datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Experiments are limited to CIFAR-based datasets, testing on a large-scale dataset like Imagenet would better validate the method’s scalability. Also evaluating HACk-OOD on the OpenOOD benchmark would provide a clearer comparison to recent methods. I would consider rating this paper higher if Imagenet results were provided.\n\n2. Missing comparisons with some of the latest post-hoc OOD methods, such as ASH and SCALE. Including these would offer a more comprehensive assessment of its relative performance.\n\nDjurisic, Andrija, et al. \"Extremely simple activation shaping for out-of-distribution detection.\" ICLR 2022\nXu, Kai, et al. \"Scaling for training time and post-hoc out-of-distribution detection enhancement.\" ICLR 2023" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see the weaknesses part." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The method takes an interesting approach to distance-based OOD detection by relaxing the distributional assumptions and, unlike naive KNN, still leveraging nearby training data statistics to construct class contours. To the best of my knowledge, the use of hypercones for this purpose is novel and appears well-motivated.\n\n2. The authors present their method clearly, making the paper easy to follow and understand.\n\n3. Although the method involves hyperparameter k, the authors provide a practical approach to estimating it without requiring additional OOD data.\n\n4. The experiments investigate various model sizes and training losses, demonstrating their impact on distance-based methods. Overall, the method benefits more from larger models trained with contrastive loss, as these models produce more distinguishable features." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a post-training out-of-distribution (OOD) detection method, HAC_k-OOD, which models the training data distribution through a set of hypercones and assesses OOD status based on whether a test sample falls within any hypercone. Specifically, for each class, the method first computes the class centroid and defines an angular boundary using the k-th nearest neighbors for each training point. Additionally, a radial boundary is set based on the mean and variance of the sample norms within the angular boundaries. During inference, a sample is classified as OOD if it lies outside either the angular or radial boundaries. Experiments were conducted using ResNet-18, ResNet-34, and ResNet-50, with both supervised contrastive learning and cross-entropy loss. The method was evaluated on CIFAR-10/100 as the in-distribution dataset and tested on various OOD datasets, covering both near and far OOD scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The computational complexity is a concern for this method. Since the computation appears to increase with the size of the training dataset, it’s unclear if this approach would be feasible for large-scale, real-world applications. Although the authors state that the method is computationally efficient and support this with inference time per sample, I encourage them to provide a more detailed discussion on this aspect. For instance, what is the time required to construct the hypercones? A comparison of inference times with other methods would also be valuable.\n\n2. How would the method perform on a large-scale, real-world dataset like ImageNet? Many recent OOD detection methods use ImageNet-1k as the in-distribution (ID) dataset. I encourage the authors to consider experiments on this dataset to evaluate the general applicability of the method in more realistic scenarios.\n\n3. Recent work has explored OOD detection using CLIP as a backbone model (eg. [a1]), as CLIP may offer a more robust feature space. It would be interesting to see how this method performs when applied to a CLIP-based model.\n\n4. Could the authors elaborate on why this method outperforms a naive KNN approach? One advantage seems to be that the method leverages the nearest neighbors within the training set (as opposed to KNN’s i.i.d. approach) to construct hypercones, which may capture more robust information about class boundaries. Additionally, an ablation study using either angular or radial boundaries separately for OOD detection could provide valuable insights into the method’s effectiveness and support future research.\n\n5. While the paper is generally well-written, a few sections could be clearer. For example, in lines 80-81, P_in is referenced without being introduced in a previous formula. Additionally, in Section 5.2, only ResNet-34 is mentioned as the backbone model, though ResNet-18 and ResNet-50 are also used.\n\n[a1] Ming, Yifei, et al. \"Delving into out-of-distribution detection with vision-language representations.\" Advances in neural information processing systems 35 (2022)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "- (see Weakness 1): how does the method perform on the ImageNet (200 or 1k) OOD benchmarks? Its evaluation on benchmarks beyond *quite simple* datasets (i.e., CIFAR-10/CIFAR-100) would strengthen the claims. \n- What is the intuition behind the use of hypercones? Isn't the embedding space of a class more \"dense\" close to its centroid?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper is well-written and easy to follow. The authors provide a good summary of the different approaches to OOD, a background section on hypercones, and a clear and precise method description. \n2. Relevance and novelty of the method: the algorithm doesn't require assumptions about the data distribution and can model complex embedding spaces since it draws multiple hypercones per class and since it defines per-hypercone decision boundaries. \n3. The authors discuss some limitations of the method (e.g., it works less well with smaller models )." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a post-training method for out-of-distribution (OOD) detection. The method approximates the contour of each class with a set of hypercones and defines per-hypercone decision boundaries. \nMore specifically, the hypercones are drawn as follows: \n1. compute per-class centroids, which will be the apex of the class hypercones \n2. take a class sample and set the hypercone axis to be the vector that points its (penultimate) representation \n3. set the opening angle to be the angle between the hypercone axis and the k-th nearest neighbor from the sample representation\n4. set the decision boundary using the distribution of representations within the hypercone's angular boundary\n\nThe method is an extension of another technique, SSD+, which assumes the decision boundary could be modeled with a unique hypersphere (or multidimensional ellipsoid) per class. It's training does not require OOD data. \n\nExperimentally, the authors follow common practice and evaluate the OOD performance with the CIFAR-100 dataset as an in-distribution dataset and many different datasets as out-of-distribution datasets. Two types of pre-trained classification models are tested: models trained with a softmax cross-entropy loss and with a supervised contrastive loss. \nThe method achieves SOTA results in the supervised contrastive learning setup and is competitive in the cross-entropy setting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited evaluation: the method is only evaluated on models pre-trained on the *quite simple* CIFAR datasets and not on more complex datasets such as the ImageNet-200 or ImageNet-1k OOD benchmarks." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Please address the questions in the weaknesses part as above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper addresses an important problem and proposes a new distance-based model based on hypercones. \n2. The evaluation of the method seems comprehensive and it achieves strong results in some cases." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a post-training strategy for Out-Of-Distribution (OOD) detection for image classification. The proposed method assumes no access to the OOD samples and employs a set of hypercones with varying cutoff distances in feature space to define the class boundaries of in-distribution data. The work evaluates this method and a combination with a previous OOD technique on the Far-OOD and near-OOD detection benchmarks, with comparisons to a set of baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper is not clearly motivated. The discussion on training-based and distance-based post-training methods are insufficient. While the introduction section listed many previous methods, it is unclear what OOD modeling challenges this method aims to address, in particular for the distance-based approach.\n2. The assumption of this method is very restrictive. As stated in Line 064, it requires \"that ID and OOD data are separable in the space\", which is unrealistic for real-world data. \n3. The novelty of this method is limited. The proposed hypercone representation is similar to a mixture of Gaussian kernels for the ID data distribution. \n4. The presentation of this work lacks clarity and the technical details are difficult to follow. Several parts of Sec 4.3 are confusing: 1) Line 260: Why do the hypercone representations rely on the test data feature Z_{test}, which should not be used during model construction? 2) Line 299: How is the score function defined and what threshold is used during the inference (in Sec 4.4)? \n5. The experimental evaluation is lacking in three aspects: 1) The experimental setup is limited, which only considers three ResNet-based backbones. More modern architectures, such as ViT, should be included to validate its generalization. 2) The ablative study is lacking. What contributions are from the hypercones? What if it is replaced by Gaussians? 3) The performance of the original version HAC_k-OOD is mixed in both Table 1 and 2, and in most of cases, it is worse than the SOTA methods. It is unclear whether the proposed representation is truly effective." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "HACk-OOD is a novel, training-agnostic OOD detection method that constructs hypercones in the feature space to approximate in-distribution contours of classes, without making distributional assumptions or explicitly training for OOD detection." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024hypercone,\ntitle={Hypercone Assisted Contour Generation for Out-of-Distribution Detection},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xE5ZaZGqBW},\nnote={under review}\n}" }, "abstract": { "value": "Recent advances in the field of out-of-distribution (OOD) detection have placed great emphasis on learning better representations suited to this task. While there have been distance-based approaches, distributional awareness has seldom been exploited for better performance. We present HACk-OOD, a novel OOD detection method that makes no distributional assumption about the data, but automatically adapts to its distribution. Specifically, HACk-OOD constructs a set of hypercones by maximizing the angular distance to neighbors in a given data-point's vicinity, to approximate the contour within which in-distribution (ID) data-points lie. Experimental results show state-of-the-art FPR@95 and AUROC performance on Near-OOD detection and on Far-OOD detection on the challenging CIFAR-100 benchmark without explicitly training for OOD 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": [ "OOD detection", "Out-of-distribution detection", "Computer Vision", "Deep Learning", "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/dd3d79734542f30802682964c28d7108aeded14a.pdf" }, "presentation": null, "primary_area": { "value": "other topics in machine learning (i.e., none of the above)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Hypercone Assisted Contour Generation for Out-of-Distribution Detection" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Please address the weaknesses mentioned above.\n2. In the evaluation, is the accuracy of the prediction computed using the augmented graph (after upsampling), or is it just based on the original graph?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Interesting ideas on using label informativeness to help tacking imbalance dataset.\n2. The is built on top of SMOTE, a well known upsampling technique, and augment it to graph setting.\n3. The experiments demonstrate the claimed benefit of the proposed approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents an oversampling technique for imbalance classification in graph classification settings. The method is based on the label informativeness as the criteria to augment the minority class samples. The method started with the regular non-graph SMOTE technique for generating new samples. Then it uses LI criteria to determine which nodes are connected to the new samples, in such a way that maximizes the LI criteria. The authors then demonstrate the performance of the proposed model in the real experiments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Presentation. In many places, terminologies/abbreviations/symbols/equations are used without explaining them. For examples:\n - “IR”. (Page 2)\n - “SMOTE”. (Page 3)\n - “Eq (6)”. (Page 4)\n - “Eq (7)”. (Page 4)\n - etc\n2. Motivation. The motivation behind using LI for upsampling is not explained clearly. Why is using LI a good idea to do upsampling? Why is having high LIs desirable? etc.\n3. Soundness. The data augmentation by creating new nodes with new features using SMOTE technique makes sense. However, the way the proposed method augments the edges by attaching the new nodes to potentially unrelated nodes (not a neighbor of the original node), does not necessarily make sense. I get that the goal is to improve the LI metric. However, edges in the original graph represent certain relationship patterns. Attaching the new nodes to unrelated nodes will create new relationship patterns that do not exist in the original graph. These relationship patterns in the graph are something that is not yet explored by the model." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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. Compared to existing oversampling methods based on graph data, what research gaps are addressed by the method proposed in this paper? \n2. The authors proved Theorem 1 in the manuscript. Still, I am more interested in whether the variation of t for different problems affects the results of the experiments and how the optimal value of t is determined.\n3. Is there a generalization of sacrificing significant computational cost to improve computational accuracy?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The authors propose Label Informativeness-based Minority Oversampling and theoretically establish the relationship between label informativeness and model prediction accuracy in GNN. \n2. The authors analyze the effect of variation in the number of inter- and intra-class edges on LI. \n3. The manuscript is logical and well-structured. The methodology section is clearly presented." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a new method for resolving class imbalances in graph-structured data based on a labeled informativeness-based minority oversampling method, called LIMO. By increasing the edges in a way that maximizes the amount of label information, LIMO strategically samples nodes of minority classes. And satisfactory results are obtained in node categorization dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. At the end of the first section, the authors show LIMO only in the figure, but there is no corresponding textual description in the text.\n2. The currently proposed Label Informativeness-based Minority Oversampling (LIMO) approach is indeed a novel idea to solve the graph imbalance problem, but the core differences with existing approaches can be further highlighted in the introduction section or related work section such as 10.1109/IJCNN60899.2024.10650494, 10.1109/ICASSP48485.2024.10448064. \n3. Discussion on the generalizability of LIMO is lacking in the manuscript, the authors only validate the performance effect of LIMO in GraphSAGE and lack discussion on LIMO in other GNNs such as Graph Convolutional Networks (GCN), Graph Attention Networks (GAT/ GAN). The authors need to provide a theoretical discussion of how LIMO might be generalized to other GNN architectures, or validated experimentally. \n4. The experimental results are not comprehensive enough, and oversampling methods that specifically address the problem of graph data imbalance are missing from the comparison methods. Therefore, it is suggested that the authors should add methods that target graph data correlation in the experimental section, e.g., 10.1109/IJCNN60899.2024.10650494, 10.1109/ICASSP48485.2024.10448064.\n5. Although the authors mention the inherent cost of the LIMO approach in generating and evaluating new edges in the summary section, the authors do not mention whether the added computational cost is within an acceptable range. The authors should provide actual runtime comparisons between LIMO and baseline methods, or a theoretical complexity 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": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is easy-to-follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "To alleviate the class-imbalanced issue, this paper propose a novel algorithm, Label Informativeness-based Minority Oversampling (LIMO), aiming to strategically synthesize minority nodes by augmenting edges to maximize label informativeness. And the experiments conducted on various homophilous and heterophilous benchmark datasets show the improvements compared with the baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty is limited. It seems just simply inject SMOTE with label informativeness, however, the label informativeness is borrowed from previous work without clear explanation. Specifically, what is the meaning of the definition of LI in Eq.2? What situation will the LI(G) increase? And when will LI(G) decrease? More importantly, how does LI(G) influence the performance of class-imbalanced learning?\n\n2. This paper fails to demonstrate its strengths compared with previous works. In lines 81-87, the authors claim that the proposed method take label informativeness to handle class-imbalanced issue. But they fail to illustrate the limitations of previous works, and how the proposed method can tackle the limitations? Why does the previous works, like GraphSMOTE [1], GraphENS[2] and TAM[3], can not improve the label informativeness? What is the strengths of LIMO compared with these methods?\n\n3. In line 152, during the node generation, LIMO need to find \"its nearest neighbor $u$ in the same class\". Does node $u$ belong to the labeled training set or the unlabeled nodes? If it comes from the training set, will it lead to overfitting when the number of minority classes is too small? If $u$ is an unlabeled node, how to get a reliable pseudo-label in the class-imbalanced scenario?\n\n4. Lack of complexity analysis of the proposed method. In lines 11-13 of Algorithm 1, for each minority node $v$, LIMO need to determine the edge between $s$ and $w$, and $w \\in \\mathcal{V}-\\{v\\}$. The computation cost seems pretty high.\n\n5. In line 216, \"Specifically, it suggests that enhancing the connectivity between classes (increasing inter-class edges) can improve the informativeness of the labels\". This conclusion seems to contradict graph homophily [4] (the foundation of GNN), i.e., the connected nodes often come from the same class. In my opinion, we always want to add edges between two nodes from the same class, so that the nodes can aggregate more information from the same class, thus helping the node classification, just like GraphSMOTE[1] does. I have some doubt about the conclusion of line 216.\n\n6. In addition, the baselines is out-of-date, there are many related works should be taken into comparison, for example, the oversampling methods like GraphENS[2] and GraphSHA[5], the pseudo-labeling method like GraphSR[6] and the loss adjustment method TAM[3]. The experimental setup is inappropriate. In line 729, \"When the minority class had fewer than three nodes, we allocated one node each for training, validation, and testing.\", It is unreasonable to have only one node for validation and testing. Why is the experimental setting not consistent with the previous method?\n\n\n\n[1] Zhao T, Zhang X, Wang S. Graphsmote: Imbalanced node classification on graphs with graph neural networks[C]//Proceedings of the 14th ACM international conference on web search and data mining. 2021: 833-841.\n\n[2] Park J, Song J, Yang E. Graphens: Neighbor-aware ego network synthesis for class-imbalanced node classification[C]//International conference on learning representations. 2021.\n\n[3] Song J, Park J, Yang E. TAM: topology-aware margin loss for class-imbalanced node classification[C]//International Conference on Machine Learning. PMLR, 2022: 20369-20383.\n\n[4] McPherson M, Smith-Lovin L, Cook J M. Birds of a feather: Homophily in social networks[J]. Annual review of sociology, 2001, 27(1): 415-44\n\n[5] Li W Z, Wang C D, Xiong H, et al. Graphsha: Synthesizing harder samples for class-imbalanced node classification[C]//Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2023: 1328-1340.\n\n[6] Zhou M, Gong Z. GraphSR: a data augmentation algorithm for imbalanced node classification[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2023, 37(4): 4954-4962." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "* In the experiments, what does the imbalance ratio represent specifically set within the range of 0.1 to 0.6?\n* Different dataset seems to show different performance. For example, in Table 10 when imbalance ratio = 0.1/0.2, while the LI of LIMO is the largest, the performances of RW and GS are better. Why?\n* The experiment results can be further explained. For instance, in Figures 5 and 10, no significant correlation is observed once the fraction exceeds 0.25. Additionally, in Table 8, when the imbalance ratio is set to 0.1, ACC and F1-Score remain the same across four baseline methods.\n* It is unclear whether the effectiveness stems more from well-synthesized node features or from the generated edges. Which one is more important? What are the appropriate $\\delta$ values for interpolating node features?" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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 experiments, what does the imbalance ratio represent specifically set within the range of 0.1 to 0.6?\n* Different dataset seems to show different performance. For example, in Table 10 when imbalance ratio = 0.1/0.2, while the LI of LIMO is the largest, the performances of RW and GS are better. Why?\n* The experiment results can be further explained. For instance, in Figures 5 and 10, no significant correlation is observed once the fraction exceeds 0.25. Additionally, in Table 8, when the imbalance ratio is set to 0.1, ACC and F1-Score remain the same across four baseline methods.\n* It is unclear whether the effectiveness stems more from well-synthesized node features or from the generated edges. Which one is more important? What are the appropriate $\\delta$ values for interpolating node features?" }, "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 easy to understand.\n* The paper focuses on the problem of class imbalance, a critical issue in many real-world applications.\n* Rich experiments have been conducted on a variety of datasets, examining different aspects." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents LIMO, an algorithm designed to tackle class imbalance in graph data. It works by adding synthetic edges between minority class nodes to increase Label Informativeness. It aims to improve the performance of GNNs on imbalanced datasets without significantly increasing the volume of data. The authors provide theoretical analysis and experimental results, showing LIMO outperforms existing methods, especially on heterophilous graphs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The paper lacks a discussion of several recent and relevant baselines for addressing class imbalance in graphs. Notably, recent methods such as GraphENS [1], TAM [2], and GraphSHA [3] are not included in the experimental comparisons.\n* The theoretical proof with t=1.31167628 is derived from simulation, but the paper does not provide a detailed explanation or justification of the simulation settings used. This lack of transparency limits the reliability of this theoretical result.\n* While the paper theoretically claims that LI is directly proportional to GNN accuracy, experimental results do not consistently support this. For instance, in Figure 9, the correlation between LI and accuracy weakens when added_fraction > 0.5, especially at fraction = 0.75, where no strong positive correlation is observed.\n\n[1] Park, J., Song, J., & Yang, E. (2021). Graphens: Neighbor-aware ego network synthesis for class-imbalanced node classification. In International conference on learning representations.\n[2] Song, J., Park, J., & Yang, E. (2022, June). TAM: topology-aware margin loss for class-imbalanced node classification. In International Conference on Machine Learning (pp. 20369-20383). PMLR.\n[3] Li, W. Z., Wang, C. D., Xiong, H., & Lai, J. H. (2023, August). Graphsha: Synthesizing harder samples for class-imbalanced node classification. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (pp. 1328-1340)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1.\tThe calculation of $t$ value given in Theorem 1 is not rigorous, since the authors explore it in a certain range using predefined steps, which might not be the optimal value. The proof of Theorem 2 is also not convincing.\n\n2.\tThe chosen baselines are too old, and some key baselines are not discussed and compared, e.g., GraphENS [1], and GraphSHA [2], etc.\n\n3.\tNo time and space complexity analyses are given to show the efficiency of the proposed model. \n\n4.\tThe authors claim that the proposed model does not significantly increase the data volume. However, the authors synthesize nodes and edges in the graph, which indeed increases the data volume.\n\n5.\tThe caption in Figure 1 does not provide any information, and the authors did not cite references appropriately. For instance, SMOTE was not cited upon its first mention. \n\n6.\tWhat labels are used in the Amazon datasets? The original datasets do not include labels for the user nodes.\n\n\nReferences:\n\n[1] Park J, Song J, Yang E. Graphens: Neighbor-aware ego network synthesis for class-imbalanced node classification[C]//International conference on learning representations. 2021.\n\n[2] Li W Z, Wang C D, Xiong H, et al. Graphsha: Synthesizing harder samples for class-imbalanced node classification[C]//Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. 2023: 1328-1340." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tThe paper is generally well-written and easy to understand, meanwhile, the results appear to be promising.\n\n2.\tExperiments are conducted on both homophilous and heterophilous datasets, which broadens its application. \n\n3.\tVarious ablation studies are given to show the impacts of the synthesized nodes and generated edges." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates class imbalanced graph representation learning, where certain classes have significantly fewer nodes. Ignoring such kind of information will result in biased learning or overfitting. The authors claim that existing oversampling techniques overlook the label informativeness of the graphs, which measures the amount of information regarding to its neighbor’s labels. To this end, the authors propose LIMO to oversample minority class nodes by maximizing LI. Synthetic node features are generated with SMOTE, then edges are added by analyzing the derivative of its LI. Comprehensive experiments on both homophilous and heterophilous datasets demonstrate that the proposed model can outperform recent baselines under different imbalance ratios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe calculation of $t$ value given in Theorem 1 is not rigorous, since the authors explore it in a certain range using predefined steps, which might not be the optimal value. The proof of Theorem 2 is also not convincing.\n\n2.\tThe chosen baselines are too old, and some key baselines are not discussed and compared, e.g., GraphENS, and GraphSHA, etc.\n\n3.\tNo time and space complexity analyses are given to show the efficiency of the proposed model." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Label Informativeness-based Minority Oversampling in Graphs" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024label,\ntitle={Label Informativeness-based Minority Oversampling in Graphs ({LIMO})},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xEDB5sSIK0},\nnote={under review}\n}" }, "abstract": { "value": "Class imbalance is a pervasive issue in many real-world datasets, particularly in graph-structured data, where certain classes are significantly underrepresented. This imbalance can severely impact the performance of Graph Neural Networks (GNNs), leading to biased learning or over-fitting. The existing oversampling techniques often overlook the intrinsic properties of graphs, such as Label Informativeness (LI), which measures the amount of information a neighbor's label provides about a node's label. To address this, we propose Label Informativeness-based Minority Oversampling (LIMO), a novel algorithm that strategically oversamples minority class nodes by augmenting edges to maximize LI. This technique generates a balanced, synthetic graph that enhances GNN performance without significantly increasing data volume. Our theoretical analysis shows that the effectiveness of GNNs is directly proportional to label informativeness, with mutual information as a mediator. Additionally, we provide insights into how variations in the number of inter-class edges influence the LI by analyzing its derivative. Experimental results on various homophilous and heterophilous benchmark datasets demonstrate the effectiveness of LIMO in improving the performance on node classification for different imbalance ratios, with particularly significant improvements observed in heterophilous graph datasets. Our code is available at \\url{https://anonymous.4open.science/r/limo-1A36/}" }, "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": [ "class imbalance", "graph neural networks", "mutual information", "label informativeness" ] }, "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/4d537bb414fef21c92bed3a168d411bcebb78fa6.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": "Label Informativeness-based Minority Oversampling in Graphs (LIMO)" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- What is the computational overhead of the proposed method compared to standard sampling?\n- Have you explored using other vision-language models besides CLIP for external conditioning?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper presents a interesting approach to address the issue of NSFW content generation in SBMs by utilizing external conditioning signals.\n- The SFW sampler allows for user-defined NSFW classes, making it adaptable to different settings and applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a safe-for-work (SFW) sampling method for diffusion models to prevent the generation of not-safe-for-work (NSFW) content. The key innovation is using external multimodal models (specifically CLIP) as a source of conditioning to guide samples away from undesired regions in the ambient space." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- No ablation studies on different choices of external multimodal models besides CLIP.\n- The proposed method is a direct application of manifold-preserving sampler.\n- The quantitative results in experiments are not convincing. The proposed method does not show 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": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "NA" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- **Complementing Direct Classification:** Could the authors discuss how their method complements or improves on using a direct classifier to detect NSFW content? Specifically, are there notable trade-offs in generation fidelity when comparing this approach to post-generation filtering?\n\n- **Stable Diffusion Configurations:** Stable Diffusion has various versions and parameter settings that might impact results. Could the authors clarify which versions (and parameter settings) were used in their experiments? A sensitivity analysis would improve replicability." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- **Significance:** The studied problem is very important in practice.\n\n- **Readability:** This paper avoids overly complex writing, which enhances comprehension for a broad audience.\n\n- **Minimal Impact on Image Quality:** The proposed method has a minor impact on the quality of benign samples, maintaining the aesthetic appeal of images that do not require correction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a method to make diffusion models safer by reducing the likelihood of generating NSFW content. The authors leverage an external signal from CLIP to guide the generation process, steering it away from harmful or explicit images. With the introduction of a Conditional Trajectory Correction step, the model subtly aligns generated content to safe categories without compromising image quality.\n\nThis flexible approach allows users to define what qualifies as harmful content, with customizable categories suited to different contexts. Experiments on Stable Diffusion demonstrate that the method reduces NSFW content with minimal impact on image quality, while a prompt-alignment metric assesses faithfulness to the user’s intent. Overall, this method provides a practical step toward safer generative models aligned with user-defined standards." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Moderate Impact:** The experimental results show a limited reduction in NSFW content. The evaluation metrics are somewhat narrow; incorporating widely adopted metrics like FID or Inception Score could provide a more comprehensive view of how the method impacts visual quality and generation fidelity.\n\n- **Computational Overhead:** The Conditional Trajectory Correction step likely increases computational and time costs due to added gradient estimations at each inference step. The authors do not discuss this in detail, and it would strengthen the paper to include a comparison of inference times with and without this step. Additionally, while the Conditional Diffusion Trajectory Correction adapts the manifold-preserving guidance from He et al., 2024, this adaptation may not represent a substantial advancement over existing methods.\n\n- **Parameter Sensitivity:** The model’s effectiveness depends on tuning parameters like the harmfulness threshold and guidance strength, which may vary by application. This sensitivity could hinder usability, as careful parameter adjustments might be needed for different scenarios. Including guidance or analysis on parameter selection would enhance practicality." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "In general, I'm willing to raise my score, if an area can be explored where the proposed method yields a significant performance improvement. Regarding the aforementioned weaknesses, I pose the following questions:\n\n1. $p_h$ is defined implicitly using trained classifiers. If the probability density $p_h$ of harmful samples is reliant on classifiers, what are the benefits of using your approach, rather than removing harmful descriptions from the prompt? \n\n2. Have you considered using a more elaborate set of concepts $\\mathcal{C}$ for harmful content detection? How would expanding $\\mathcal{C}$ impact the robustness and generalizability of the CTC approach?\n\n3. In addition to question 2: How does the cardinality $|\\mathcal{C}|$ affect the computational cost of the proposed CTC method and how does it affect performance? Can you provide an upper bound on computational cost or some time measurements for inference, especially for increasing $|\\mathcal{C}|$?\n\n4. The results indicate limited improvement over Erasing Stable Diffusion (ESD). Could you elaborate on any advantages of CTC in scenarios or datasets where it may outperform ESD? Were there specific cases in your experiments where CTC provided a distinct advantage?\n\n5. Given that the effectiveness of CTC is influenced by the threshold $\\eta$. The hyper-parameter selection of $\\eta=0.23$ was hand-selected based on a grid search. How sensitive is the performance of your approach to different values $\\eta \\in (0,0.23)$?\n\n6. Could you clarify any potential limitations in terms of scalability or generalization of the proposed CTC method? Specifically, are there specific concepts $\\mathcal{C}$ where the approach struggles, and what could be done to address these limitations?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors propose an interesting approach to generating SFW samples with SBMs, leveraging the solver's working principles for the reverse process. Herein lies the potential to extend the method to a broader range of applications beyond content moderation. By directly influencing the sampling trajectory based on explicit probability thresholds, this method could adapt to domain-specific needs where content control is essential, such as ensuring ethical content in SBM-driven artwork or meeting regulatory standards in automated media production. Furthermore, the Conditional Trajectory Correction (CTC) is theoretically grounded." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the challenge of minimizing the generation of undesired, harmful content in Score-Based Models. The distribution of such harmful samples is represented by a probability density $p_h$. \n\nThe authors aim to estimate the conditional probability $p_h(x_0 | x_t, t)$ to identify how likely a sample $x_t$ will produce harmful content at state $x_0$ in the generative process. They propose searching for candidate samples $x_{t-1}$ in the vicinity of $x_t$ that satisfy two criteria: i) they are valid samples, and ii) they yield low density $p_h(x_0 | x_t, t)$.\n\nTo implement this, the authors introduce a Conditional Trajectory Correction (CTC) mechanism. The proposed CTC evaluates the NSFW probability of each clean point prediction and applies a corrective adjustment if this probability exceeds a predetermined threshold $ \\eta > 0 $. This approach aims to effectively reduce the generation of harmful content while maintaining the integrity of valid sample generation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One of the primary limitations of this paper lies in its reliance on an implicitly defined $p_h$ using trained classifiers. This approach introduces a dependence on the quality and generalizability of these classifiers, which by the authors' admission may not consistently or accurately capture all harmful content, especially if the classifiers have limited scope or biased training data.\n\nAdditionally, the method requires a predefined set of harmful concepts, $\\mathcal{C}$, to guide content filtering. In this study, the largest considered set was $\\mathcal{C}$ = {violence, nudity, NSFW, harmful}. However, this may not be comprehensive enough to capture the full range of potentially harmful content across various contexts, particularly in domains where nuanced or emerging types of harmful content need to be addressed. This limitation in scope could reduce the method’s effectiveness in broader applications.\n\nFinally, the reported results show little to no improvement over Erasing Stable Diffusion (ESD), a competing approach. This raises questions about the practical advantages of the proposed Conditional Trajectory Correction (CTC) method. Given its performance, it remains unclear if the added complexity of CTC justifies its use over established methods like ESD, especially in settings where computational efficiency and ease of implementation are critical considerations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Please see above weaknesses" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The paper is easy to follow and self-contained.\n- The experimental design and the proposed method are sound and make sense to me. \n- Sections 4 and 2.2 are quite useful in understanding the generalizability of the method.\n- Aesthetic degradation experiment is important for the work and I commend the authors for doing that." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies an important problem of safe-for-work sampling from diffusion models to avoid generating explicit content. To this end, they propose a novel approach to leverage CLIP-based embeddings of harmful concepts and guide the sampling of diffusion models away from this embedding space. They specifically use manifold preserving guidance and only apply it when the classifier is confident enough (i.e., its probability is above a threshold). The proposed method is then compared against baselines to find competitive performance with baselines in explicit content detection and quality and prompt-concordance of the generated image." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The empirical comparison with traditional guidance-based methods is missing. It has been discussed that this is different from classifier-based guidance but no empirical evidence is provided to show that the proposed method is better. \n- A related technique of negative prompts is not compared or discussed:\n - https://minimaxir.com/2022/11/stable-diffusion-negative-prompt/\n- In almost all cases and especially in Table 1, ESD outperforms the proposed method and this negative result is not adequately discussed. This is strange since it seems that the proposed method is not working. \n- The only positive result is when the prompts are already safe (Table 3), which seems to indicate that the proposed method is beneficial due to its soft ways of conditioning and steering away from certain regions in an embedding space.\n- Novelty of the method is also limited since it uses off-the-shelf classifiers and manifold-preserving guidance approach along with a simple thresholding mechanism on the confidence. \n- In the conclusion, the authors note that explicit conditioning is important to uphold the bias of existing diffusion models. This is interesting but due to the lack of any empirical evidence to support this claim, it is hard to establish that explicit conditioning is indeed the way forward. Thus, if this is a claim that the authors want to establish, some results would be really useful and relevant. \n- Other important related work are missing from experiments and/or discussion.\n - Hong, Seunghoo, Juhun Lee, and Simon S. Woo. \"All but One: Surgical Concept Erasing with Model Preservation in Text-to-Image Diffusion Models.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 38. No. 19. 2024.\n - Lyu, Mengyao, et al. \"One-dimensional Adapter to Rule Them All: Concepts Diffusion Models and Erasing Applications.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.\n - Pham, Minh, et al. \"Circumventing concept erasure methods for text-to-image generative models.\" The Twelfth International Conference on Learning Representations. 2023." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We avoid the generation of NSFW images in diffusion models by modifying the sampling process with an external conditioning source." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024sfw,\ntitle={{SFW} sampling for diffusion models via external conditioning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=xETLME9sNq},\nnote={under review}\n}" }, "abstract": { "value": "Score-based generative models (SBM), also known as diffusion models, are the de facto state of the art for image synthesis. Despite their unparalleled performance, SBMs have recently been in the spotlight for being tricked into creating not-safe-for-work (NSFW) content, such as violent images and non-consensual nudity. This article proposes a safe-for-work (SFW) sampler for SBMs implementing a Conditional Trajectory Correction step that guides the samples away from undesired regions in the ambient space using external multimodal models as the source of conditioning. Furthermore, using Contrastive Language Image Pre-training (CLIP), our method admits user-defined NSFW classes, which can vary in different settings. Our experiments on the text-to-image SBM Stable Diffusion validate that the proposed SFW sampler effectively reduces the generation of explicit content, as assessed via independent NSFW detectors. Furthermore, the proposed correction comes at a minor cost in image quality and has an almost null effect on samples that do not need correction. Our study confirms the suitability of the SFW sampler towards aligned SBM models." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "diffusion", "score-based", "safeness", "alignment", "guidance" ] }, "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/d2481dd5d5bb5ca364b74728905636868fbe5367.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": "SFW sampling for diffusion models via external conditioning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]