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": 4 }, "primary_area": null, "questions": { "value": "The NeurIPS 2023 oral by Kingma and Gao unifies under a common framework the diffusion variants of, e.g., flow matching and optimal transport. This could be useful, given that both are discussed in your work. You might also take a look at their Appendix J, which has some discussion of the frequency analysis of diffusion noise that may have some overlap with your proposed frequency-dependent noise.\n\nCan we hear some audio examples? It is difficult to judge the results otherwise, and the MUSHRA test (unlike an AB or ABX test) doesn't provide sufficient granularity to discern statistical significance between, e.g., FlowDec and DAC." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Figures 2 and 3 are great visualizations for the diffusion dynamics and proposed improvements.\n\nGood selection of baselines, including retraining baseline models to account for differences in parameter count.\n\nThe authors mention that they will be open-sourcing their code." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes FlowDec, a 48 kHz general audio codec with a flow-matching diffusion post-filter. FlowDec modifies the DAC audio codec with different loss functions, the stochastic post-filter, and frequency-dependent noise. Evaluations demonstrate strong improvements over prior post-filter and diffusion-based methods, but only small improvements over DAC." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "You should consolidate the six main contributions. I recommend at most three or four. Some of the contributions currently listed seem minor, or are a side effect of another contribution.\n\nA primary issue with the proposed system is RTF. The logical alternative (DAC) is still orders of magnitude faster. I do still think this is an important paper in helping close the gap.\n\nA minor note, but a 44.1 kHz sample rate is sufficient to prevent audible aliasing above 20 kHz. The other rationales given for using 48 kHz over 44.1 kHz are convincing enough for me." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. Could the authors clarify the main purpose of FlowDec? Why is it presented as a standalone codec rather than positioning the postfilter as an enhancement tool?\n2. What is the rationale for choosing to enhance only outputs from a non-adversarial codec instead of using GAN output as the initial estimate for FlowDec's postfilter? Could the authors specify what benefits NDAC offers over DAC in this proposed two-stage setup?\n3. The STFT settings seem unconventional. A 1534-point FFT would definitely be slower than a 1536-point FFT due to FFT efficiency with highly composite sizes. I assume the goal is to get exactly 768 frequency bins, but why not just use a 1536-point FFT with 769 bins? Overall, this may be negligible given the slower inference, but I'm curious about the tradeoff here." }, "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, with clear explanations and informative illustrations that help clarify key points.\n- It provides useful context relative to previous score-based models. It also includes valuable theoretical insights and formalism to support its approach.\n- It performs on par with state-of-the-art GAN codecs and shows slight improvements for speech signals. FlowDec may handle high-frequency harmonics better, where GAN-based codecs often introduce periodicity and harmonic artifacts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces FlowDec, a neural audio codec with a two-stage approach: (1) an autoencoder with residual vector quantization, trained without adversarial loss, and (2) a postfilter that reduces coding artifacts and enhances perceptual quality. FlowDec adapts conditional flow matching for signal enhancement, achieving improvements over previous score- and flow-based models. Both listening tests and objective metrics show that FlowDec provides perceptual quality competitive to state-of-the-art GAN-based codecs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper introduces FlowDec as a codec, but its primary focus seems to be a flow-based postfilter that enhances outputs from an *underlying* codec. The main weakness of this paper is the decision not to explore a GAN-based codec combined with the proposed postfilter. The paper argues against adversarial training in Section 3.3, stating that the “generated phase may be very different from the original phase.” However, FlowDec itself doesn't seem to preserve phase. Results presented in Section 5.1 show that FlowDec performs worse than GAN-based codecs in terms of SI-SDR and fwSSNR, which are sensitive to phase shifts. This might still be acceptable for a low-bitrate codec, particularly if perceptual quality is prioritized. However, the paper should more clearly justify why a GAN isn't used as the underlying codec. A GAN-based codec with a flow-based postfilter might not only reduce coding artifacts but also achieve higher reconstruction metrics. This would call into question the rationale behind the proposed non-adversarial DAC (NDAC) model. Overall, the distinction between FlowDec as a novel audio codec and its role as an enhancement tool for *any* audio codec could be clarified." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "There have been several concurrent works adopting flow matching with similar task (i.e. mel spectrogram vocoder and discrete codec decoder), such as RFWave [1] and PeriodWave [2,3]. These work directly utilize flow matching as the decoder rather than postfilter, with several configs being order of magnitude faster than this work.\n\nWhile the direct comparison may not be strictly necessary, can the authors provide conceptual comparisons between the aforementioned methods along with possible advantages of FlowDec?\n\n[1] Liu, Peng, and Dongyang Dai. \"RFWave: Multi-band Rectified Flow for Audio Waveform Reconstruction.\" arXiv preprint arXiv:2403.05010 (2024).\n\n[2] Lee, Sang-Hoon, Ha-Yeong Choi, and Seong-Whan Lee. \"PeriodWave: Multi-Period Flow Matching for High-Fidelity Waveform Generation.\" arXiv preprint arXiv:2408.07547 (2024).\n\n[3] Lee, Sang-Hoon, Ha-Yeong Choi, and Seong-Whan Lee. \"Accelerating High-Fidelity Waveform Generation via Adversarial Flow Matching Optimization.\" arXiv preprint arXiv:2408.08019 (2024)." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Contrary to adversarial training (a dominant approach in vocoder/codec) which requires domain specific expertise to stabilize training and tune the hyperparameters of multiple losses, FlowDec simplifies the trianing pipeline by elimiating the adversarial losses. In my opinion, this can be considered as one of the first work that achieves competitive quality to GAN-based models with RTF < 1. \n\n* The proposed design choice is well justified overall, both in the theoretical and empericial perspective. It provides enough rigor in experimental detail, including toy experiments and detailed ablation study of each component." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "FlowDec is an improved version of ScoreDec (uses a score-based generative model as a postfilter), by switching the objective to flow matching. It further proposes a joint flow matching objective tailored for the postfiltering task (e.g. mean-shifted noise with frequency-dependent diagonal covariance). This makes it faster than real time (unlike ScoreDec) as a practically viable, full-band, and universal audio codec without adversarial training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* While the evaluation results from the paper look convincing overall, no demo page has been provided at submission. It's hard to form the reviewer's opinion on the subjective quality without access to the demo. I hope the authors can provide the samples for the reader to evaluate the subjective quality themselves. \n\n* Although it improved the RTF by a large margin, it still requires multiple NFE from the postfilter resulting in RTF of 0.22-0.23 which is considerably slower than recent, feed-forward audio codecs." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- The question arises as to why the waveform is not reconstructed directly from discrete tokens using flow matching. To provide a more comprehensive evaluation, could you elaborate on the potential computational or quality trade-offs between direct reconstruction via flow matching and the method proposed in the paper?\n- Regarding line 168-171, the presence of $p_{data}(\\cdot|y)$ is not clear as it seemingly should just be $p_{data}(x)$. This needs further clarification and justification.\n- From line 220 to 223 and in Figure 2, it may be inappropriate as we cannot obtain $x_1$ for real postfiltering problems. This aspect requires reconsideration or better explanation.\n- In Figure 3, the use of a large $\\sigma_t$ of 0.1 o illustrate that FlowAVSE is non-contractive is not convincing, especially since the original FlowAVSE employs a value of 0.04. It would be beneficial for the authors to provide a clearer explanation for this choice or consider revising the value used. \n- As $y$ is a time-domain signal, it is unclear how to apply frequency-dependent $\\sigma_y$ since it cannot be directly applied. The authors need to provide more details on this aspect." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Possibly the first application of flow matching in postfiltering.\n- Beautiful experiment result plots and strong objective and subjective results.\n- Enhance DAC by incorporating a multi-scale Constant-Q Transform (CQT) loss." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors introduce FlowDEC, a flow matching-based post-filtering method designed to enhance the audio quality decoded from discrete audio tokens. This method has demonstrated strong objective and subjective results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The latency of the proposed method is at least an order of magnitude higher than that of DAC. To facilitate a comprehensive evaluation, it would be beneficial to receive specific latency measurements for both the proposed method and the DAC. Furthermore, identifying the applications or use cases that could be critically impacted by this higher latency would provide valuable context. Moreover, to support the validation of the method’s performance, it would be advantageous if the authors could provide an online demo that allows for direct comparison of the latency between the two methods.\n- The baselines are limited. There are other diffusion-based works that reconstruct audio waveforms from discrete tokens, e.g. Multi-band diffusion [1]. \n- The concept of selecting a data-dependent prior, as proposed, is not entirely novel, with works such as Priorgrad [2] having implemented a similar strategy. To better understand the relationship between the proposed method and existing approaches, it would be helpful to have a clearer comparison of their similarities and differences. \n- Multi-band diffusion [1] uses frequency-dependent noise levels, which is a notable feature that could be further discussed in comparison to the proposed method.\n\n[1] \"From Discrete Tokens to High-Fidelity Audio Using Multi-Band Diffusion\" by Robin San Roman, Yossi Adi, Antoine Deleforge, Romain Serizel, Gabriel Synnaeve, and Alexandre Défossez.\n\n[2] \"PriorGrad: Improving Conditional Denoising Diffusion Models with Data-Dependent Adaptive Prior\" by Sang-gil Lee, Heeseung Kim, Chaehun Shin, Xu Tan, Chang Liu, Qi Meng, Tao Qin, Wei Chen, Sungroh Yoon, Tie-Yan Liu" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "FlowDec is a flow-based postfilter codec for general audio without adversarial training, and a competitive alternative to current GAN-based SOTA codecs." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024flowdec,\ntitle={FlowDec: A flow-based full-band general audio codec with high perceptual quality},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uxDFlPGRLX},\nnote={under review}\n}" }, "abstract": { "value": "We propose FlowDec, a neural full-band audio codec for general audio sampled at 48 kHz that combines non-adversarial codec training with a stochastic postfilter based on a novel conditional flow matching method. Compared to the prior work ScoreDec which is based on score matching, we generalize from speech to general audio and move from 24 kbit/s to as low as 4 kbit/s, while improving output quality and reducing the required postfilter DNN evaluations from 60 to 6 without any fine-tuning or distillation techniques. We provide theoretical insights and geometric intuitions for our approach in comparison to ScoreDec as well as another recent work that uses flow matching, and conduct ablation studies on our proposed components. We show that FlowDec is a competitive alternative to the recent GAN-dominated stream of neural codecs, achieving FAD scores better than those of the established GAN-based codec DAC and listening test scores that are on par, and producing qualitatively more natural reconstructions for speech and harmonic structures in music." }, "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": [ "audio", "audio codec", "generative models", "flow matching", "postfilter", "signal enhancement" ] }, "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/c10512ecac00cfcd8f490464b3ebab006bfce3db.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": "FlowDec: A flow-based full-band general audio codec with high perceptual quality" }, "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": "•\tThe problem only considers the univariate case. Can the model extend to the multivariate time series problem?\n\n•\tI want some more explanation about the effectiveness of informed prior distributions. Why does closedness of the prior and data distribution imply easy learning. Do you have any experiments about train efficiency or path efficiency? \n\n•\tCan the given prior (Gaussian process) extend to the arbitrary prior ? for example, refer to [1].\n\n•\tDo you have any theoretical evidence about how the selection of kernel functions effect to the model performance? (ex. OU kernel is better when the data follows OU process)\n\n[1] Leveraging Priors via Diffusion Bridge for Time Series Generation, Arxiv24" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "• By utilizing Gaussian process to the conditional flow matching, the model reflects the temporal dependencies of the given time series data better.\n\n• The model enables both unconditional and conditional generations.\n\n• By conditional prior sampling, the unconditionally trained model could follow the given guidance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The existing diffusion models have problems in the time series generation since the data and prior distributions differ. The authors handle this problem by utilizing conditional flow matching framework. They propose TSFlow, which sets the prior distribution as Gaussian process to make the prior distribution close to the data distribution. Also, they propose conditional prior sampling which makes an unconditionally trained model possible for probabilistic forecasting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Please refer 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": 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": "- The exact problem statement was a bit unclear to me. In lines 154-156 the authors describe a time series as a vector in $\\mathbb{R}^L$. Does this mean that the authors only work on time series having a fixed length $L$, or does the method allow for variable-length time series? Similarly, are the authors assuming that the time series all share a fixed discretization (i.e., there are some fixed times $t_1, \\dots, t_L$ corresponding to $y_1, \\dots, y_L$), or can the discretization vary across time series? Is this discretization assumed to be uniform, or can it be irregular as well?\n- It seems to me that the setup is not limited just to forecasting, but could be applied to general conditional generation tasks, e.g., imputation. Have the authors tried anything beyond forecasting?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The investigation of techniques for conditional sampling is pretty thorough, and I think the proposed methods are quite interesting\n- The proposed method obtains fairly strong empirical results, and the empirical evaluation is convincing \n- Throughout the paper is very clear and well-written" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the use of flow matching for time series forecasting. The authors first propose the use of flow matching techniques for unconditional generation using Gaussian process priors (section 3.1.1), followed by a technique for conditioning an unconditional model by sampling a relevant prior $x_0$ (section 3.1.2) or guidance (section 3.1.3), and finally the authors discuss a technique which uses a data-dependent Gaussian process prior for conditional sampling (section 3.2). The proposed methodology is empirically validated on several univariate time series datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There is some highly relevant related work that the authors do not discuss. [Functional Flow Matching, AISTATS 2024](https://arxiv.org/abs/2305.17209) proposes the use of GP priors in conjunction with flow matching and studies techniques for forecasting with these models. Similarly, [Conditional Flow Matching for Time Series Modeling, SPIGM@ICML 2024](https://openreview.net/forum?id=Hqn4Aj7xrQ) uses GPs with flow matching for time series. The authors should cite these works and discuss the differences with their proposed method.\n- There are some (relatively minor) clarity issues throughout\n - The use of equation 9 was a bit unclear to me. Why is this specific form of $q_1$ chosen? Some justification for this modeling choice would be good.\n - In Section 3.1.2, I am guessing that once we sample $x_0 \\sim q_0(x_0 \\mid y^p)$, then we use $x_0$ as an initial condition for the flow model to generate new samples $y \\mid x_0$. Is this the case? If so, it would help to state this explicitly somewhere in the paper.\n - In Line 303, the authors write \"approximating q_0(x_0 \\mid y^p)$ with $q_0(x_0 \\mid y^p)$. I think I understand what is meant here, but this seems to be a typo." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "See above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "**Incorporating Gaussian Process Priors:** The main contribution—replacing the typical isotropic Gaussian prior $q(x_0)$ with a data-dependent conditional prior $q(x_0∣y^p)$ is well-motivated. GP priors are naturally suited for time series due to their ability to model temporal dependencies, and this idea is a clear innovation over existing flow matching methods.\n\n**Empirical Performance:** The empirical results show that TSFlow performs well compared to state-of-the-art models across various benchmark tasks.\n\n**Flexibility in Conditional and Unconditional Modeling:** The approach supports both unconditional and conditional generation. While unconditional generation has fewer use cases compared to conditional generation, it is a feature often overlooked in time-series analysis. The ability to use the same model for both tasks—by applying conditioning only during inference—adds versatility to the paper." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces TSFlow, a model for probabilistic time series forecasting that enhances generative modeling by incorporating Gaussian Process (GP) priors within the Conditional Flow Matching (CFM) framework. The use of more informative GP priors helps align the prior distribution with the temporal structure of the data, potentially improving both performance and leading to runtime efficiency by simplifying the probability paths. TSFlow is flexible and can be used for both, conditional and unconditional, time-series generation.\n\nThe model demonstrates strong performance across several benchmark datasets, but there are aspects of the paper that could benefit from further clarification and improvement. Depending on the author's response, I am willing to increase my score from \"weak reject\" to \"weak accept\"." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Majors**\n\n**Difficulty in Parsing for Non-Experts:** The paper assumes substantial familiarity with flow matching and related generative methods, making it challenging for readers without a deep background in these specific techniques. It took for me considerable time to fully grasp the concepts, which suggests that the paper might also be difficult to read for a broader audience.\n\n**Lack of Runtime Analysis:** The authors propose replacing the isotropic Gaussian prior with a more complex GP prior. Howeer, they do not address the increased computational cost that comes with using GP priors. Although the GP prior is more suited to time-series tasks, this advantage must be weighed against the computational overhead. The paper should clearly state the theoretical runtime complexity of using GP priors and provide empirical runtime comparisons in the experiments. The reduction in NFEs is a positive, but its trade-off with the computational cost of the GP prior must be considered.\n\n**Baseline Comparison:** A simple baseline for the forecasting tasks could involve using Eq. 6 but with an isotropic Gaussian prior. Please include this method as a comparison partner in Table 3. If it is not a valid approach, it should be explained why such a baseline is excluded.\n\n**Inconsistent Findings on Kernel Choice:** The periodic kernel minimizes the Wasserstein distance in Figure 2, suggesting it aligns well with the data distribution. However, in Table 1, the periodic kernel does not significantly outperform other kernels, and it is not even considered in Table 4.2. This inconsistency is counterintuitive and warrants further discussion. Moreover, the necessity for different hyperparameter choices across tasks (e.g., generative modeling vs. forecasting) weakens the \"one model for all\" argument, suggesting more task-specific tuning may be required.\n\n**Minors:**\n\n**Missing Experiments on Guided Generation (3.1.3):** I could not find experiments on Guided Generation (3.1.3) in the paper. Where can I find them?\n\n**LPS Score Clarification:** The Linear Predictive Score (LPS) is not sufficiently explained. Please provide more details on how the LPS is calculated, specifically on what the linear model is regressed against and what the output represents.\n\n**GP Hyperparameter Optimization:** You state in your paper that you did not fit the GP hyperparameters. This is somewhat unexpected, I would have had a strong guess that learning the hyperparameters of the GP brings a large benefit to the model. Can you perform a small experiment to evaluate the difference in performance with/without GP hyperparameter optimization?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. If my understanding is correct, the overall idea of the model is to push an initial GP to the target time-series. Therefore, the sequence length $L$ in the problem formulation section is equivalent to the vector field dimension $d$ in the background section. Can the author(s) kindly confirm if this is true? It would be nice to clarify the setting a little bit. I think the main source of confusion is that there are two time indices in this paper: the time in flow-matching and the time in the time-series, and these are orthogonal to each other. I think this should be made clear somewhere in the paper.\n\n2. Following the first question, it would be nice to indicate in Figure 1 the time-series index and the flow-matching index. Moreover, different notations should be used. (For instance, on line 202-203, $t$ should not be used for the kernel because it has already been used in Eq. (1).)\n\n3. How would the model respond to a growing $L$? That is, there are two subquestions in this query:\n\ta. What is the time complexity of the model and how does it compare to other models that you benchmarked against?\n\tb. When $L$ is large, Eq. (6) seems to suffer from the curse of dimensionality and the integral would be impossible to discretize. How does the model see this issue, or is it not relevant?\n\n4. The section \"Effect on the Optimal Transport Problem\" only considers the distance between sequences but not anything about the training. Can you show some experiments where models that use the periodic kernel are indeed easier to train on tasks that involve periodicity?\n\n5. On line 302-303, you wrote \"we additionally condition the prior distribution on the observed past data by approximating $q_0(\\mathbf{x}_0 | \\mathbf{y}^p)$ with $q_0(\\mathbf{x}_0 | \\mathbf{y}^p)$.\" I assume there is a typo. What would be the intended sentence?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The model is flexible with many possible options; the informed prior sampling is intuitive.\n\n2. The model shows promising results on univariate benchmark tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a model that uses flow-matching algorithms to generate and forecast univariate time-series. When trained unconditionally, it replaces the isotropic Gaussian prior distribution with an informed one and uses conditional sampling and inferencing. The model itself can also be made conditional. Experiments show that the model surpasses existing ones on benchmark tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The model right now is only discussed and validated on univariate time-series.\n\n2. While there are many proposed options in the model, there lacks an ablation or a comprehensive comparison of different choices. Neither is there any theoretical insights in the proposed model.\n\n3. The presentation of the model needs a bit more clarification. The problem formulation can potentially be expanded a bit more (see questions below). The discussion of different training/inference choices is a bit dense. Maybe some concise pseudocode or flowchart in the main text could be helpful." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose TSFlow, a conditional flow matching model for time series forecasting that leverage domain-specific priors." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024flow,\ntitle={Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uxVBbSlKQ4},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in generative modeling, particularly diffusion models, have opened new directions for time series modeling, achieving state-of-the-art performance in forecasting and synthesis. However, the reliance of diffusion-based models on a simple, fixed prior complicates the generative process since the data and prior distributions differ significantly. We introduce TSFlow, a conditional flow matching (CFM) model for time series that simplifies the generative problem by combining Gaussian processes, optimal transport paths, and data-dependent prior distributions. By incorporating (conditional) Gaussian processes, TSFlow aligns the prior distribution more closely with the temporal structure of the data, enhancing both unconditional and conditional generation. Furthermore, we propose conditional prior sampling to enable probabilistic forecasting with an unconditionally trained model. In our experimental evaluation on eight real-world datasets, we demonstrate the generative capabilities of TSFlow, producing high-quality unconditional samples. Finally, we show that both conditionally and unconditionally trained models achieve competitive results in forecasting benchmarks, surpassing other methods on 6 out of 8 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": [ "flow matching", "time series forecasting", "generative modeling", "deep learning" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/0f3327414b38b730ae6923baecff14dfd1d5c15d.pdf" }, "presentation": null, "primary_area": { "value": "learning on time series and dynamical systems" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Flow Matching with Gaussian Process Priors for Probabilistic Time Series Forecasting" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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 the weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The work propose design a Topic Control Mechanism (TCM) to emphasize the target domain-specific knowledge. Combined with Token Topic Modeling (TTM), Masked Region Modeling (MRM), and Text Image Matching (TIM), KLAFT highlights the target related\nknowledge (i.e., the knowledge lift).\n\n- The propsoed KLAFT improves expressive captioning tasks by aligning and amplifying target knowledge, with the potential for\nParameter-Efficient fine tuning (PEFT) at low computational cost." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work present a visual tuning framework, Knowledge Lift Alignment Fine Tuning (KLAFT), which enhances the expressive image captioning capabilities of Pretrained Language Models (PLMs). The innovation of KLAFT lies in its approach to addressing the disparities in knowledge - visual versus textual via MAM and source versus target domain via TCM. These hidden spaces are conceptualized as distinct sub-networks, each possessing specific knowledge, KLAFT adjusts the weights of these sub-networks in a fine-grained manner. The empirical studies demonstrate that KLAFT improves expressive captioning tasks by aligning and amplifying target knowledge." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The writting of this work is quite poor. The inroduction part does not clearly introduce the background and motivation of the problem to be solved. The approach part is also not written in a concise and well-organized manner.\n\n- The approaches compared in this work are not proposed recently. As a result, the validity and innovation of the proposed modules is not convincing.\n\n- The authors did not conduct sufficient ablation experiments for different loss functions and did not give more qualitative 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": 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": "* In Figure 1 (left), a plot of token distributions is shown, but there is no information about the dataset. What are the source and target domains? How were these lines plotted? Given the significance of comparing the two domains, this should be one of the most important analyses in the paper.\n* I get the main idea of TCM. Yet, there are some phrases in the paper making the details of TCM unclear to me:\n * In line 234, `\"distribution over tokes, $\\mathbf{V}$. What is $\\mathbf{V}$? It only appears once in the paper.\n * In the paragraph following Eq (5), the term $b_z$ is mentioned twice, though it doesn’t appear in Eq (5). This inconsistency makes it challenging to understand Equation (5) fully. Additionally, the variable $w$ is not explained in the paragraph.\n* In lines 338-339, it's stated that three hyperparameters are all set to 0.1, which seems to be less commonly done. What is the rationale for this choice? Is there any ablation study to support it?\n* In Sec. 5.2, the authors compare S4 to S3 to showcase the effectiveness of TCM. Yet, these two settings differ a lot:\n * S3: \"seq2seq under fine-tuning GPT-2 over 100% training data (COCO+Conceptual Captions)\"\n * S4: \"prefix 100% training data with frozen setting (only COCO)\"\n * Given the differences in both tuning paradigm and datasets, how can TCM’s impact be fairly assessed in this context?\n* Could you provide more details on human evaluation?\n\n*I'm open to adjusting my rating after the authors' response.*" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* Increasing the expressiveness of the generation process in VLMs is a promising research direction.\n* The proposed TCM/TTM approach is intriguing and shows potential.\n* The performance gains over baseline models are satisfactory." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Knowledge Lift Alignment Fine-Tuning (KLAFT) to enhance the expressiveness of image captioning capabilities in pretrained Vision-Language Models (VLMs). KLAFT primarily introduces a Topic Control Mechanism (TCM) combined with Token Topic Modeling (TTM) to enable topic-guided tuning and decoding for VLMs. The proposed method is evaluated across various settings." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The presentation of the paper lacks clarity. I’ve outlined some key issues below:\n * The term \"KEIC\" appears three times—once in Figure 1, once in the title of Section 4, and again in the conclusion. These are critical sections. I assume \"KEIC\" should actually be \"KLAFT.\"\n * Some typos or unclear sentences can a problem, e.g. ``tokes`` appears multiple times.\n * The image in Table 1 is not clear, and Figure 1 also lacks effective delivery and clarity.\n * I assume all the citations use ``\\cite{}``. I think most citations should likely use ``\\citet{}`` for smoother integration.\n * And some other issues, see question section as well.\n* Overall, the claims around the design of the Mapping Layer (MaL) and the Modified Attention Mechanism (MAM) may be overstated. These components are fairly common in VLMs, and MAM appears to be a direct adaptation from one of the primary baselines, VisualGPT.\n* The choice of main baselines seems outdated, although results from some recent VLMs (e.g., LLaVA, BLIP2) are also included." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Please refer to the Weaknesses. The following is a minor question.\n\n1. Considering the fast-evolving nature of the VLM development, are some recently proposed and commonly used multimodal LLMs, such as VILA or InternVL-2, also applicable to this proposed fine-tuning framework?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Improving the quality of captions generated from VLMs has great potential for several real-world applications, such as manufacturing and healthcare.\n2. The proposed framework is reasonable to address the considered task." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a Knowledge Lift Alignment Fine Tuning (KLAFT) framework to improve the image captioning capability of multimodal LLMs. More specifically, this paper aims to encourage the generated captions to be detailed and comprehensive. To achieve that, this paper explores fine-grained alignment and designs MAM and TCM attention mechanisms. Experiments and quantitative comparisons are conducted on commonly used benchmark datasets, including MS COCO and Conceptual Captions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. As described in the Abstract, the goal of this paper is to generate detailed and comprehensive captions. However, the benchmark datasets adopted by this paper (e.g., MS COCO) are typically coarse-grained with relatively short captions. As we know, current VLMs, like LLaVA, are good at generating detailed and fine-grained image captions. Is using COCO to evaluate the detailed caption capability of VLMs suitable? My concern lies in adopting such traditional image captioning benchmarks (e.g., COCO), which cannot properly measure the detailed caption capability of VLMs and reflect the actual improvement of the proposed method.\n2. What are the differences between the source and target domains claimed in this paper? Does the difference lie in the visual domain shift? Does the difference lie in the caption style? Does the difference lie in the amount of (labeled) data? More clarification and explanation are encouraged to make this paper more easy to read.\n3. The proposed method is trained by multiple training objectives. It is complicated to model training and balance the weights of each loss function (in Eq.(9)). How does this paper design experiments to determine the weight of each loss term? Is the model training sensitive to different weights of loss terms?\n4. The visual presentation of this paper has some room to improve. For example, the image in Table 1 is out of the table. In addition, for Table 3, it is better to directly indicate the meaning of each row in the table instead of just mentioning it in the captions." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "KLAFT aims to generate detailed and comprehensive captions by leveraging fine-grained alignment between PLMs and target domain datasets." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024knowledge,\ntitle={Knowledge Lift Alignment Fine Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uxYbEAEWm4},\nnote={under review}\n}" }, "abstract": { "value": "We present a visual tuning framework, \\textbf{K}nowledge \\textbf{L}ift \\textbf{A}lignment \\textbf{F}ine \\textbf{T}uning (KLAFT), \nwhich enhances the expressive image captioning capabilities of Pre-trained Language Models (PLMs), including LLMs and VLMs.\nAs this task involves generating more detailed and comprehensive captions than basic image descriptions,\nthe core idea behind KLAFT is that fine-grained alignment could exploit the capabilities of PLMs and a given target domain dataset.\nThis idea motivates and challenges us to explore the framework that deeply understands both given images and text for this alignment and tuning PLMs towards expressive image captioning.\nThis direction modifies the attention mechanism (Modified Attention Mechanism, MAM) and develops both a Topic Control Mechanism (TCM) and their training objectives.\nThe innovation of KLAFT lies in its approach to addressing the disparities in knowledge - visual versus textual via MAM\nand source versus target domain via TCM.\nAs these hidden spaces are conceptualized as distinct sub-networks within the PLM, each possessing specific knowledge,\nKLAFT's unique contribution is in aligning and adjusting the weights of these sub-networks in a fine-grained manner,\nand fine-tuning this PLM.\nOur empirical studies demonstrate that KLAFT significantly improves expressive captioning tasks by aligning and amplifying target knowledge, with the potential for Parameter-Efficient Fine-Tuning (PEFT) at low computational cost." }, "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": [ "PEFT", "PLM", "LLM", "VLM", "Multi-modal", "Image captioning" ] }, "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/c1aef593ef14be15d6dce332c782c53eb47a4c65.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": "Knowledge Lift Alignment Fine Tuning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How do the authors position their work in the field of technical game design? \n2. What do longer gameplay examples look like? \n3. How do the authors respond to my concerns around their \"Evaluation of LLM Generations\"?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The main strengths of the paper are (1) the originality of entirely basing a virtual pet game on LLMs, (2) the significance of the two primary technical contributions to figure coherent image generation and real time LLM application uses, and (3) the quality of the image portion of the evaluation/experiments. The paper is also overall well-written, especially in terms of the technical aspects. These primarily justify the positive aspects of my above scores." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors present \"UNBOUNDED\" a virtual pet game based on LLMs. The authors present two novel technical elements to support this game: a regional IP-Adapter for character-environment consistency and a domain-specific distillation approach for finetuning a smaller LLM to allow for real time gameplay. The authors then present experiments comparing their approach to baselines in terms of visual quality and consistency, and overall game quality as scored by GPT-4o." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper has two major weaknesses in its present draft, the way it positions itself in terms of prior work and the evaluation. \n\n### Prior Work\n\nThe authors do not appear to have engaged with the field of technical games research at all. This is a shame, as their game can be understood as an AI-based game [1] or specifically a NN game [2]. More broadly, the authors' work in terms of generation would fit into the Procedural Content Generation paradigm [3]. There is a 3+ decades long history of work in this space that is relevant to the authors' work [4], with significant recent work at the intersection of LLMs and game generation [5], and it is crucial that the authors update the introduction and related work to situate their work within this field.\n\n### Evaluation\n\nThe authors primarily evaluate their work in terms of image quality and I have no major concern with these results (though the DreamSim results seem to incorrectly attribute the best score to their system). However, I have major concerns with the \"Evaluation of LLM Generations\", which is functionally the closest thing presented to a holistic evaluation of UNBOUNDED. Firstly, the norm within the area would be to make use of a user study for evaluation purposes [6,7]. Particularly since the authors make such an effort to allow for real time control, this would seem to follow naturally. Secondly, the evaluation is somewhat concerning due to the training process. Specifically, since the distillation approach makes use of GPT-4 to simulate user interaction data and the authors sample for uniqueness/variety, its fairly likely that the newly sampled user interaction data approximations will be very similar to these. Further, the largest model the authors compare to is again a GPT-4 variant. Given that GPT-4 is being used for both generation of data, as the system, and as the evaluation, there's a clear risk for bias. This could be addressed by any number of ways, including automated metrics for measuring game quality developed in prior game generation-like work [8,9]. But at present, it is difficult as an outside reader to get a sense of the actual quality of UNBOUNDED as a game. Failing this, it might be beneficial to include longer examples of gameplay in the supplementary materials at least.\n\n### Figures\n\nWhile these are not of major concern, I wanted to note two aspects of figures that I felt could be improved. First, the authors make use of a large number of fonts in their figures, and their primary font is a bit difficult to read. I'd suggest updating the figures to use a single, more legible font. Second, Figure 4 is largely superfluous. If the authors wish to retain it, it might be a better fit for an appendix.\n\n1. Treanor, M., Zook, A., Eladhari, M. P., Togelius, J., Smith, G., Cook, M., ... & Smith, A. (2015). AI-based game design patterns.\n2. Zhu, J., Villareale, J., Javvaji, N., Risi, S., Löwe, M., Weigelt, R., & Harteveld, C. (2021, May). Player-AI interaction: What neural network games reveal about AI as play. In Proceedings of the 2021 CHI Conference on Human Factors in Computing Systems (pp. 1-17).\n3. Shaker, N., Togelius, J., & Nelson, M. J. (2016). Procedural content generation in games.\n4. Pell, B. (1992). METAGAME: A new challenge for games and learning.\n5. Gallotta, R., Todd, G., Zammit, M., Earle, S., Liapis, A., Togelius, J., & Yannakakis, G. N. (2024). Large language models and games: A survey and roadmap. arXiv preprint arXiv:2402.18659.\n6. Anjum, A., Li, Y., Law, N., Charity, M., & Togelius, J. (2024, May). The Ink Splotch Effect: A case study on ChatGPT as a co-creative game designer. In Proceedings of the 19th International Conference on the Foundations of Digital Games (pp. 1-15).\n7. Guzdial, M., & Riedl, M. O. (2021). Conceptual game expansion. IEEE Transactions on Games, 14(1), 93-106.\n8. Khalifa, A., Green, M. C., Perez-Liebana, D., & Togelius, J. (2017, August). General video game rule generation. In 2017 IEEE Conference on Computational Intelligence and Games (CIG) (pp. 170-177). IEEE.\n9. Guzdial, M., & Riedl, M. (2018, September). Automated game design via conceptual expansion. In Proceedings of the AAAI Conference on Artificial Intelligence and Interactive Digital Entertainment (Vol. 14, No. 1, pp. 31-37)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "To which extent are results from Figure 6 consistent with results from Table I?\nHow have you ensured the quality and fairness (for evaluation) of the text prompts sample?\nHow can the current approach maintain narrative consistency over a number of interactions?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This is an ambitious piece of work, tackling difficult technical issues in balancing generation and interaction, as well as various training and fine-tuning protocols. It appears to have innovated on a number of aspects, for instance the dynamic mask to balance character and environment generation. It gives some interesting, detailed, insights such as those of Figure 4. The improvement over IP-Adapter thus appears more than incremental. Another strong points is the Small LLM distillation framework.\nThe comparative LLM evaluation has staged an appropriate number of LLM, capturing size and diversity. \nFinally, the paper demonstrates a good command of recent relevant work and compares its approach to really recent alternative methods (2023, 2024), which is welcome." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work combines various aspects of Generative AI to develop a LLM-based game engine. The design principles refer to open-world, continuous interaction games with limited simultaneous characters but significant opportunities for situation generation and interaction. Capitalizing upon recent research in the field, the system introduces a number of technical innovations both in terms of architecture (in the form of a dynamic regional prompt adapter) and in term of implementation (distillation of Small LLM). With an existing implementation, the paper includes elements of comparative evaluation, both quantitative and qualitative." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "From a fundamental perspective, there is a lack of awareness of some game design issues. The introduction is relatively naïve considering how these issues are addressed in game/play theory (Caillois, Huizinga). With the fine line between simulation and entertainment, and notwithstanding the reference to The Sims and Tamagotchi, contemporary audiences tend to have gameplay expectations that can only be met with sustained and meaningful (in the narrative sense) interactions. This feeds into the overall perspective of the evaluation methods which in the paper are primarily focused individual ‘event’ generation. While sustained UX evaluation may be at odds with the system’s current level of maturity, some of this could have been addressed through even slightly longer exchanges or chains of events construed as minimal narrative experience (which is claimed in the abstract but does not really materialize in the examples). The small examples given, including on Figure 1, do not convey a very strong sense of gameplay. \n\nThere is no evidence that the number of user-simulator examples is at all sufficient for the intended purpose of multi-topic data collection. Nor is there any discussion of some hierarchical organization of topics that would come close to scenarios, game genres or elementary forms of plot backbones.\n\nDespite the statement on replicability, the level of implementation details is unclear. Implementation details are rather cursory and uneven with some in-depth details but no systematic discussion, and minor duplications from other sections.\n\nEvaluation has both positive (see above) and negative aspects.\nUnless I am mistaken, Table I shows a couple of inconsistencies for results.\nOn Environment Consistency, how can 0.322 be the best score when it is in-between 0.381 and 0.257? How can 0.675 be the best score in-between 0.595 and 0.832?\nIt is difficult to find the Qualitative component of evaluation very convincing in the absence of a more systematic approach for generating test situations. From a similar perspective, the example Prompts do not contain advanced mechanisms for controlling response, or specific examples. One would thus wonder how LLM response can be kept consistent across runs, and what type of statistical sampling has been associated to LLM use?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024unbounded,\ntitle={Unbounded: A Generative Infinite Game of Character Life Simulation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uy31tqVuNo},\nnote={under review}\n}" }, "abstract": { "value": "We introduce the concept of a generative infinite game, a video game that transcends the traditional boundaries of finite, hard-coded systems by using generative models. Inspired by James P. Carse's distinction between finite and infinite games, we leverage recent advances in generative AI to create Unbounded: a game of character life simulation that is fully encapsulated in generative models. Specifically, Unbounded draws inspiration from sandbox life simulations and allows you to interact with your autonomous virtual character in a virtual world by feeding, playing with and guiding it - with open-ended mechanics generated by an LLM, some of which can be emergent. In order to develop Unbounded, we propose technical innovations in both the LLM and visual generation domains. Specifically, we present: (1) a specialized, distilled large language model (LLM) that dynamically generates game mechanics, narratives, and character interactions in real-time, and (2) a new dynamic regional image prompt Adapter (IP-Adapter) for vision models that ensures consistent yet flexible visual generation of a character across multiple environments. We evaluate our system through both qualitative and quantitative analysis, showing significant improvements in character life simulation, user instruction following, narrative coherence, and visual consistency for both characters and the environments compared to traditional related approaches." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Text-to-Image Generation", "Interactive Image Generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/9802d28b8f281f44f05ccd196684783a80a6e51a.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": "Unbounded: A Generative Infinite Game of Character Life Simulation" }, "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": "- W.r.t. the finite sample analysis, is the proposed algorithm robust towards any noise in the samples?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- This paper studied an important and practical problem of model multiplicity. The paper overall is well-organized and presented with a good clarity. \n- It is in particular helpful to have the illustrative example in Figure 1, which directly shows that it is insufficient to only update two predictive models so that they have improved squared loss and nearly agree on their individual predictions almost everywhere.\n- Theoretical guarantee shows that the new algorithm ReDCal provides an improved accuracy and approximately agrees on the best-response action almost everywhere compared to the prior work." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studied the problem of model multiplicity in downstream decision-making. In this setting, two predictive models of equivalent accuracy do not agree on their predictions for the downstream decision-making problem.\nThe paper proposed a new calibration framework which calibrates the predictive models with respect to both a finite set of downstream decision-making problems and the individual probability prediction. Further, the paper proposed an algorithm that first reconciles the differences in individual probability prediction then calibrates the updated models so that they are indistinguishable from the true probability distribution to the decision-makers together with its finite-sample analysis. Numerical experiments on two datasets demonstrate the effectiveness of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The experimental results with the HAM10000 dataset show substantially larger error bars, and much less smooth convergence. It is helpful to provide more details on this differences between the two sets of results.\n- The experiments only compared to one other baseline proposed in (Roth et al 2023). How does the proposed algorithm compared to other related works in the model multiplicity?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. One thing I didn't understand was if you are given two models at the end of ReDCal, how do you choose which model to use? Why is it important that the two models agree in the decisions they make if you already control the decision-loss gap?\n2. If you don't want to affect the decision-loss, it makes sense to me that you just wouldn't change the two predictors. Is that the case and is it reflected in your algorithm if you choose the correct parameters?\n3. Why does the algorithm converge so fast in the numerics? The upper bound on the time steps seems to suggest the number of time steps that should be run before convergence should be large given the choices of $\\beta$ and $\\alpha$. Does this imply the bound is is potentially very loose and is it possible for it to be made tighter?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper highlights an important problem, that improvements to prediction models can hurt downstream decision-making since downstream decision-makers may have loss functions that do not necessarily align with prediction accuracy. The paper combines existing work in multi-calibration with work in model multiplicity to solve this problem. The algorithm proposed by the paper seems novel and provides what seems to be sensible theoretical guarantees that trade-off between improvements to prediction accuracy and preserving decision-loss. This seems to be a relatively significant improvement to the Reconcile algorithm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the Reconcile algorithm proposed in [1] which takes two prediction models and outputs two models that have better prediction accuracy which is quantified by the Brier score. They expand on the existing method by proposing a new reconcile algorithm that looks to improve prediction accuracy while also preserving the downstream decision-loss. They provide theoretical bounds that show their method trades off improvement to prediction accuracy at the cost of downstream decision-loss. They provide numerical results to show their algorithm helps reduce the loss gap after running their version of the Reconcile algorithm. \n\n\n\n-----\n[1] Aaron Roth, Alexander Tolbert, and Scott Weinstein. Reconciling individual probability fore- casts. In Proceedings of the 2023 ACM Conference on Fairness, Accountability, and Trans- parency, FAccT ’23, pp. 101–110, New York, NY, USA, 2023. Association for Computing Ma- chinery. ISBN 9798400701924. doi: 10.1145/3593013.3593980. URL https://doi.org/ 10.1145/3593013.3593980." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper has weaknesses in its presentation as well as results that seem somewhat suspicious/hard to interpret precisely.\n\nThe following items could be addressed and improved for presentation:\n- In the paper's introduction, the authors mention calibration several times, but for someone not immediately familiar with the literature it's hard to understand what it is formally. It becomes a little better defined at Lemma 2.6, but having extra background or explanation in the introduction would be helpful in understanding the high-level concepts used to construction the paper's proposed algorithm. \n- The paper could provide better background on the Reconcile algorithm which seems to be a key motivation for pursuing the work. There is no high level description of how the algorithm works, so Theorem 2.7 for example feels impossible to verify. Additionally, the Reconcile algorithm reproduced in the appendix seems to be missing key elements, like defining $h$ and $\\mu$. The former seems to be an important update step.\n- The paper does not define notation well. For example $\\mu$ and $A$ are not defined or are defined offhandedly, making it hard to decipher when it shows up later in the text.\n- In Theorem 2.7 the distribution $\\mathcal{D}$ is not defined which makes the last expectation in the proof hard to verify. I could not immediately come up with a distribution where the expectation is 1/2. \n- Some notation seems to be incorrect or have a typo. For example the Definition 2.4 of $E_{\\ell,a}$ is an indicator outside a set. I don't think this is standard notation and is not precise. My guess is it's a typo.\n- Is the notation $\\ell_a$ necessary? If so how do you obtain $\\ell_a$ given $\\ell$?\n- The descriptions of Decision-Calibration, ReDCal, and Decision-Calibration + ReDCal are confusing to me. What exactly is the algorithm for the first and last algorithms. When you count the number of time steps for ReDCal, which loop are you counting? Does it include the loops in decision-calibration? \n\nThe following items can be addressed to help improve the results of the paper:\n\n**Questions related to Theorem 3.4**\n- In Theorem 3.4, does item 1. have a typo? \n- Define $A$ the number of actions in the theorem statement for part 3. since it is off-handedly defined somewhere else in the paper. \n- How do I choose $\\beta$, $\\alpha$, and $\\eta$ if I wanted to guarantee a certain decision-making loss level while optimizing the Brier scores? Managing the trade-off between prediction accuracy and decision-loss would help improve the impact of this theoretical result. Right now it seems hard since the term $T_i \\beta$ seems hard to control.\n- The decision gap in the numerical results don't seem to correspond to the bound in 3. Shouldn't it be close to 0 if you choose $\\beta = 0.00001$?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The paper currently appears mostly in good shape; my main small improvement suggestions are about the experimental section (at least by lightly tweaking existing experiments/adding another simple but conceptual one). \n\nIn addition, even though the writing of the paper is mostly quite good, it could still be improved in a few places. For instance, the subsection about handling more than 2 models is somewhat confusingly written: e.g. consider the sentence about exponentially many output predictors in the number of models k, followed by the sentence that union and Chernoff-Hoeffding bounds \"suggest\" that the sample complexity scales linearly in k --- a reader who may not have checked the details in the appendix may easily confuse this for saying that the method may not necessarily be scalable in k, even though it actually is. Also, the event collection setup in that case, which involves reconciling a single base model with every other model, is not described that well in prose --- so consider e.g. including a quick diagram which could e.g. have the base model with \"reconciliation arrows\" pointing to the other models, or something similar." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper is overall a good and novel contribution to the predictive multiplicity literature. Specifically:\n\n+ The contribution of this paper appears new in the model multiplicity literature: it gives a rigorous method for reconciliation of any two models with provable guarantees in terms of the resulting model loss (for which only one algorithm exists in the literature), while at the same time ensuring in a rigorous way that downstream decision making is not affected negatively (which is new); and it moreover extends this method to more than 2 models in a natural way. \n\n+ In fact, from what I can tell, it has an even broader scope than typically considered in the model multiplicity context, as it does not require the to-be-reconciled models to start off having similar or equal accuracy. When the to-be-reconciled models do have similar accuracy to begin with, then the paper's contribution intuitively appears to offer the interesting and thought-provoking semantics of: \"given two different models with similar performance, you can bypass the multiplicity issue by finding an even higher-performing model (which ensembles/combines the two initial models), with similar or better downstream decision making guarantees\"." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a new algorithm for reconciling two or more models in a supervised setting; the goal is to address the model multiplicity problem (which refers to the phenomenon that there may be > 1 model with similar accuracy but substantial disagreements among predictions) in a decision-aware way. Specifically, the paper shows how to perform model reconciliation in a way that (1) makes the models agree in the prediction space, (2) does not hurt (and possibly even improves) the squared loss of either model, and (3) does not hurt (and possibly even improves) the decision loss of a given downstream decision-maker. (The decision loss is an arbitrary linear loss that the decision maker uses to map predictions to actions.) In addition, the paper performs preliminary experiments showing that the method outperforms (with respect to decision loss) a prior reconciliation algorithm on some vision datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While overall I believe this to be a good-quality paper, there is the following (relatively non-major) consideration that I would call a weakness:\n\n- The paper currently appears written with a primarily theoretical audience in mind, but I think it could still do a better/more thorough job coming up with/describing experiments. It currently gives two semi-synthetic ones. In the first one, linear decision losses are generated in a Gaussian manner --- so that the two vision models are essentially being calibrated to realizations of random noise (granted, the experiment does illustrate the point that the new method achieves better decision-making properties than the old one, but the setup still sounds strange). In the second one, the decision loss is a \"loss function motivated by medical domain knowledge in Zhao et al (2021) and additional random noise\" --- and in this case, the paper should at the very least clearly state what the loss function in Zhao et al (2021) is, and perhaps provide more principled perturbations of it than Gaussian noise ones.\n\nNeither of these experiments, however, connects in any way to the examples given earlier in the paper on how disregarding decision loss can lead to its deterioration during the reconciliation procedure. These examples are in some sense prototypical of how reconciliation could hurt downstream decision making; and given how easy it is to come up with a synthetic task/models + synthetic loss clearly showcasing them (i.e. making it to where some significant mass of items will flip across the decision boundary as a result of the reconciliation), I would like to urge the authors to do just that as it will make the paper more logically coherent." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Please see my questions above. Additionally: \n\n1) How does the framework perform across different domains or datasets that exhibit varying levels of complexity and noise?\n2) As mentioned, determining optimal ranges and conditions for key hyperparameters parameters across varying contexts can be challenging. \n3) Although the paper addresses computational efficiency, the scalability of the algorithm for very large datasets remains an open question. How will the computational complexity grow with increased data size?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper presents a novel framework to address the issue of model multiplicity in predictive modeling for decision-making, using multi-calibration techniques. It identifies specific hyperparameters, such as loss margin and decision-calibration tolerance as the driving parameters of key results. In practice, these parameters will need to be adjusted to mediate trade-offs between model fairness and computational efficiency. The authors provide both theoretical insights and empirical validations of their approach. The paper makes a reasonable contribution by integrating theoretical multi-calibration concepts into empirical settings and testing them across multiple datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper provides a framework using multi-calibration to reconcile model multiplicity in downstream decisions. It seeks to develop a framework to address the inherent discrepancies between predictive models, which often result in varying decision recommendations despite equivalent accuracy levels. The authors propose an algorithm that aligns predictive models with decision-making, improving agreement on best-response actions and losses. Empirical validation shows enhancements in downstream decision performance and addressing prediction disagreements, relevant to multi-calibration advancements. Proposed approach improves the consistency and utility of models in decision-based applications, including cases where only empirical data is available." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Comments/Questions to Authors: \n•\tThe initial concept of multi-calibration was introduced by Hebert-Johnson et al. in 2017 which focused on ensuring fairness across overlapping subpopulations. Recent advancement includes extension of multi-calibration to game dynamics and multi-objective learning – for example, work by Nika Haghtalab and Eric Zhao, which utilize game dynamics to connect fairness and optimization. Another line of work by Roth et al deals with a wide array of issues from online multicalibrated learning to omnipredictors. While the article does a good job at comparing and showing differences with respect to work of Roth, the conceptual difference with respect to work of Haghtalab et al. would be much appreciated. \n\n•\tThe algorithm’s performance depends on several hyperparameters (such as loss margin and decision-calibration tolerance). It seems techniques to fine-tune parameters such as loss margin and decision-calibration tolerance are crucial for achieving the results and trade-offs that is important in practice (e.g., between prediction fairness and accuracy). These are critical for adapting multi-calibration models to robustly function across heterogeneous data distributions and complex decision environments, the paper should clearly discuss how the choice of these hyperparameters impact the performance of proposed algorithm. I am also curious about the third result of Theorem 3.2 (remark 3.3). How does the impact of this result show up in practice/experiments? I believe a major limitation of the paper is lack of explicit focus on loss margin and decision-calibration tolerance hyperparameters.\n\n•\tThe paper could better motivate the empirical tests conducted and discuss how they confirm the theoretical claims and show improvements in decision-making outcomes and model agreement. Avenues for future work on validating the proposed algorithm's effectiveness in real-world data scenarios should be discussed. \n\n•\tAdditional literature – please clarify the relationships / novelty of this paper with respect to work of these authors. I would be especially interested in knowing the relevance of this paper with issues of fairness-accuracy tradeoff. \n\n“Inference for an Algorithmic Fairness-Accuracy Frontier,” Yiqi Liu and Francesca Molinari (2024). This work provides a consistent estimator for a fairness-accuracy frontier. Method for testing fairness-related hypotheses in algorithmic decision-making seems relevant. \n\n“Fair Representation: Guaranteeing Approximate Multiple Group Fairness for Unknown Tasks,” Xudong Shen and Yongkang Wong and Mohan S. Kankanhalli, IEEE Transactions on Pattern Analysis and Machine Intelligence (2021). Explores approximate fairness using fair representation across multiple fairness notions. \n\n“Multigroup Robustness,” Lunjia Hu and Charlotte Peale and Judy Hanwen Shen (2024). This work establishes a connection between multigroup fairness and robustness and discusses robustness algorithms tailored for subpopulations under data corruption.\n\n“Inherent Trade-Offs in the Fair Determination of Risk Scores,” J. Kleinberg and S. Mullainathan and Manish Raghavan (2016). Shows inherent trade-offs in algorithmic fairness in risk scores and formalizes three fairness conditions and proves their incompatibility in most cases.\n\n“Multi-group Agnostic PAC Learnability,” G. Rothblum and G. Yona, International Conference on Machine Learning (2021). Provides a framework for multi-group agnostic PAC learning.\n\nAlso important is work by Michael P. Kim et al. on Multiaccuracy: Black-Box Post-Processing for Fairness in Classification (2018): It introduces a foundational framework for ensuring fairness in classification, with a focus on multi-group fairness using multiaccuracy post-processing techniques. This paper sets a critical baseline for future research on fairness in predictions." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024reconciling,\ntitle={Reconciling Model Multiplicity for Downstream Decision Making},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uy4EavBEwl},\nnote={under review}\n}" }, "abstract": { "value": "We consider the problem of \\emph{model multiplicity} in downstream decision-making, a setting where two predictive models of equivalent accuracy cannot agree on what action to take for a downstream decision-making problem. Prior work attempts to address model multiplicity by resolving prediction disagreement between models. However, we show that even when the two predictive models approximately agree on their individual predictions almost everywhere, these models can lead the downstream decision-maker to take actions with substantially higher losses. We address this issue by proposing a framework that \\emph{calibrates} the predictive models with respect to both a finite set of downstream decision-making problems and the individual probability prediction. Specifically, leveraging tools from multi-calibration, we provide an algorithm that, at each time-step, first reconciles the differences in individual probability prediction, then calibrates the updated models such that they are indistinguishable from the true probability distribution to the decision-makers. We extend our results to the setting where one does not have direct access to the true probability distribution and instead relies on a set of i.i.d data to be the empirical distribution. Furthermore, we generalize our results to the settings where one has more than two predictive models and an infinitely large downstream action set. Finally, we provide a set of experiments to evaluate our methods empirically. Compared to existing work, our proposed algorithm creates a pair of predictive models with improved downstream decision-making losses and agrees on their best-response actions almost everywhere." }, "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 multiplicity", "multi-calibration", "decision-making", "uncertainty quantification" ] }, "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/aac0d0609216b2203552241e99ed33485a586a95.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": "Reconciling Model Multiplicity for Downstream Decision Making" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Please see the weakness listed above." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Authors provide an analysis of explainable and interpretable methods for audio deepfake detection methods. The authors compare these methods on two different benchmark datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper analyses the existing explainable methods, such as Occlusion and Attention visualization, for deepfake audio detection tasks. The authors show results using three baseline models such as AST, GBDT, Wav2Vec. The authors evaluate these methods on two existing datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper is merely an analysis paper of existing explainable methods. There are no significant novel contributions to this paper. The authors mention in line 052 that the contributions are \"Empirical evaluations of novel explainability methods for audio transformers\". However, I cannot judge what novelty is in attention visualization for transformers. It has happened a lot in literature. \n- When authors compare explainable methods for audio deepfake detection, they must show the results on existing approaches such as Shapley[1]. \n- The authors should include more explainable mechanisms as baselines, such as [2], [3].\n- The authors should show results on multilingual deepfake audio datasets such as MLAAD, DECRO, and WaveFake since AST and Wav2vec are pre-trained on the AudioSet dataset, primarily English. Evaluating a multilingual dataset would help strengthen the analysis.\n- Authors show results on the ASVspoof5 dataset. The authors mention that the dataset was released in June 2024 (line 750). However, The dataset is still not available for public use and review. Authors should show the results on the publicly available datasets.\n- FakeAVCeleb dataset is primarily an audio-video deepfake dataset, not only for audio deepfake detection. The FakeAVceleb dataset contains 500 real videos, which means there should be 500 real audio. However, authors in line 7716 mention 9712 real audio samples. Why is there a difference in the numbers?\n- Even the baseline models such as GBDT, Wav2Vec and AST are general audio architectures, not specifically designed for audio deepfake detection. Authors must show results on models such as ASSIST, RawGAT-ST and some state space models such as RawBMamaba.\n- More then 50% of the paper is just explaining trivial and non paper contributions only.\n\n\n\n[1] Explaining deep learning models for spoofing and deepfake detection with SHapley Additive exPlanations\n[2] Listen to Interpret: Post-hoc Interpretability for Audio Networks with NMF\n[3] Focal Modulation Networks for Interpretable Sound Classification" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. It is recommended to incorporate a wider variety of audio forgery datasets to validate the model's generalization capability across different forgery techniques, aligning more closely with the diversity encountered in real-world scenarios.\n\n2. While interpretability is essential, it would be beneficial to analyze how interpretability can contribute to designing better models or evaluating existing ones. Adding such analysis could enhance the paper's contribution by illustrating the practical impact of interpretability on model improvement and assessment." }, "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 interpretability framework for the audio domain, incorporating interpretability techniques from visual and natural language processing into audio Deepfake detection, providing a clearer interpretative path for the model's black-box decision-making process.\n\n2. Two interpretability methods (attention visualization and occlusion techniques) are systematically explored to assess the interpretability of Transformer-based audio detection models, with a comparison of each method's strengths and weaknesses.\n\n3. Cross-validation using the ASVspoof and FakeAVCeleb datasets demonstrates the model's generalization capability across varying data distributions, simulating data shift scenarios common in real-world applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the limitations of existing audio Deepfake detection models in terms of generalization ability and interpretability in real-world scenarios. It proposes a novel interpretability approach and establishes a new benchmark to assess model generalization performance. The study trains models on the ASVspoof dataset and evaluates them on the FakeAVCeleb dataset, demonstrating the superior performance of Transformer-based audio detection models on unseen data. Additionally, the paper introduces attention visualization and occlusion techniques to enhance model interpretability, aiming to bridge the gap between model performance and interpretability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The author mentions three limitations in the Limitation section, of which the latter two could serve as directions for future work. However, the first limitation needs to be addressed in the current phase of the task: relying solely on the ASVspoof and FakeAVCeleb datasets may not cover the full range of audio deepfake techniques encountered in real-world scenarios, presenting a dataset limitation. A wider variety of datasets and scenarios is needed to demonstrate the robustness and completeness of the current approach.\n\n2. The introduced attention visualization and occlusion techniques are computationally intensive, potentially affecting the efficiency of practical deployment. Further analysis and comparison of computational costs would be beneficial.\n\n3. The paper’s contribution is somewhat limited, as occlusion and attention visualization are commonly used techniques in computer vision and natural language processing. While adapting these methods for audio Deepfake detection is interesting, the lack of specific modifications tailored to the characteristics of audio forgery detection tasks reduces the overall impact of the proposed approach. Simple method transfer limits the originality of the 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": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How are these attention roll-out and image occlusion-based analyses aiding explainability specific to the audio deepfake analysis, and how does this contribution differ from the contributions of attention roll-out and image occlusion-based analysis methods in image feature explainability?\n\n2. Are the normalized token attention plots taken as an average of multiple audio samples or the entire dataset? How do these token attention plots vary with different samples? Further information is needed on the normalized token attention plots. \n\n3. It is stated that “..we can pinpoint specific frames that were instrumental in the classification and inspect them more closely…. and we observe that influential tokens typically appear in groups.” This is one of the primary focus of the paper, and further analysis should be done to elaborate these findings.\n\n4. What is the significance of training ASVspoof5 and inferring on FakeAVCeleb over existing benchmarks, beyond the test of generalizability? What specific challenges might inferring on FakeAVCeleb bring that training on ASVspoof5 would not cover?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors have analyzed the information available from Attention Rollouts and Image Occlusion methods. They also noted that some very short frames from audio signal representation are influential to transformers in classifying and these frames typically appear in groups, which, if further explored, may potentially lead to better interpretability of audio deepfake classifications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel explainability framework for audio transformers in audio deepfake classification. It proposes utilizing image occlusion to detect feature importance and attention roll-out to understand features better. It also open-sources a novel benchmark for detecting audio deepfakes in real-world cases, which consists of training on ASVspoof5 dataset and testing on FakeAVCeleb dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "As the authors themselves have pointed out, attention roll-out and image occlusion-based analysis have been in existence for quite some time, but the novelty of the proposed work lies in applying them in spectrograms to aid in the explainability of audio deepfake analysis. However, how these attention roll-out and image occlusion-based analyses are aiding explainability specific to the audio deepfake analysis is not adequately explained, and how their contribution differs from already existing contributions of attention roll-out and image occlusion-based analysis methods in image feature explainability remains unclear. \n\nThey have utilized the occlusion method in an attempt to explain how the model is reaching these decisions, but as they themselves pointed out, it was not helpful in explaining the model’s decision-making. They have also used an attention visualization method and stated that they can attribute specific frames that were instrumental to classification. However, using attention visualization to attribute where a transformer model is putting importance is not novel, and their analysis does not show enough contribution specific to explaining the decision process in audio deepfake classification in transformers. \n\nThey have proposed a new benchmark, which consisted of training on one existing dataset and testing on another. They have not provided adequate explanations as to how their novel benchmark would be more helpful in audio deepfake classification, and the idea of testing on a new dataset itself is not particularly novel." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. What is the technical novelty contribution of the paper's usage of existing explainability methods? Were any changes needed to adapt these existing methods for audio classification? Why not introduce modifications that can take advantage of features specific to audio modality?\n2. The result of the explainability methods didn't seem helpful for explaining model behavior. Can you provide more insights regarding this? For example, a concrete deepfake audio sample and its corresponding outputs for explainability and your analysis that will help reason about model behavior? Also, it would be helpful if a scenario is provided where the techniques are useful in practice and have a big impact on our understanding of a model's capabilities.\n3. Why is the benchmark considered a novel contribution? This kind of evaluation is very common in deep learning models." }, "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 very well written, the concepts are easily understood and the limitations of the work are discussed as well.\n2. The subject of explainability in audio deepfake detection is a significant and timely problem, especially as deepfake generation technology is readily accessible to the public and has a high capacity of being misused. Research into this subject is very important due to the impact on society this technology can have.\n3. The paper makes a difference between interpretability and explainability and proposes that explainable methods should provide interpretable explanations that are sample-specific, time-specific, and feature-specific. This robust definition of explainability would ensure greater applicability of these methods to interpret black-box models." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on the limitations of explainability in current audio deepfake detection methods and provides three-fold contributions to the subject. Firstly, the paper proposes a conceptual explainability framework emphasizing sample-specific, time-specific, and feature-specific explanations interpretable by humans. Secondly, the paper provides empirical evaluations of novel explainability methods for transformer-based detection models, using occlusion and attention visualization techniques. The occlusion method masks out sections of a mel-spectrogram to reveal which parts are most important for final classification. The attention visualization technique and roll-out method are utilized to get a distribution of attention across all layers and can be used to point out parts of the input sequence most relevant for the final classification. Finally, the paper also introduces a novel framework to evaluate the generalizability of deepfake classifiers, where models are trained on one dataset but evaluated on another dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Both the proposed methods for audio explainability, occlusion and attention visualization, are already existing methods that are very common in literature, especially for vision and language tasks. This is also recognized by the paper, which mentioned that their contribution is porting the methods to the audio task. But to do that, the paper didn't make any modality-specific changes to the methods or edit the methods in any way. For the occlusion method, it's well-known that mel-spectrograms can be treated as images and the paper directly used them on the traditional method. For the attention-visualization method, this is modality agnostic and is based on transformer architecture which takes tokens as input. The roll-out attention method was introduced for natural language tokens and the paper claims to adapt the method for audio tokens, but it's unclear what kind of modifications they did except just replacing the tokens. So the novelty introduced by the paper in these methods is questionable.\n2. The paper provided the results of their methods in Figure 4 and Figure 5. First of all, the paper understands that the result in Figure 4 doesn't help explain the model's decision-making (line 420) and instead is suggestive of transformer model behavior. Second of all, the result in Figure 5 is also not very helpful in human interpretation. When these methods are applied to a visual or textual domain, a human can more easily interpret the segment and decision-making rationale. This suggests that more work needs to be done in the audio domain to make the results more human-interpretable. These observations are also recognized by the paper in their limitation section. So the usability of these methods is questionable.\n3. The paper claims to introduce a novel benchmark to evaluate the generalization capabilities of deepfake audio classifiers, where they train the model in the ASVspoof5 dataset and evaluate it on the FakeAVCeleb dataset. Here the only contribution of the benchmark is training on one dataset and evaluating on another dataset. This mechanism is already well-known and well-practiced to show the generalization capability of a model to out-of-domain datasets. Here, there is no contribution to the dataset, no new evaluation metric is introduced or any other changes are proposed. Thus, it's questionable to consider this benchmark as a contribution. The abstract also claims to \"open-source a novel benchmark for real-world generalizability\". The question is what is there to \"open-source\" here, as the datasets are already available." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Novel explainability methods and a generalizability benchmark for deepfake audio detection." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024toward,\ntitle={Toward Robust Real-World Audio Deepfake Detection: Closing the Explainability Gap},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uy9oR0nYCW},\nnote={under review}\n}" }, "abstract": { "value": "The rapid proliferation of AI-manipulated or generated audio deepfakes poses serious challenges to media integrity and election security. Current AI-driven detection solutions lack explainability and underperform in real-world settings. In this paper, we introduce novel explainability methods for state-of-the-art transformer-based audio deepfake detectors and open-source a novel benchmark for real-world generalizability. By narrowing the explainability gap between transformer-based audio deepfake detectors and traditional methods, our results not only build trust with human experts, but also pave the way for unlocking the potential of citizen intelligence to overcome the scalability issue in audio deepfake detection." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "self-supervised learning", "explainability", "deepfake audio", "generalizability" ] }, "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/e7538d6cb10fefbb335617742b271ff0af53c441.pdf" }, "presentation": null, "primary_area": { "value": "interpretability and explainable AI" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/6f3b568cf9e13f170898a344e43cdcfff98b28b1.zip" }, "title": { "value": "Toward Robust Real-World Audio Deepfake Detection: Closing the Explainability Gap" }, "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": "Adding an algorithm highlighting the differences from the original SDS would help readers better understand the optimization process.\n\nSince the paper claims to introduce minimal additional computational cost compared to SDS, more details regarding this cost should be included.\n\nIn addition to SDS-like optimization methods, there is another line of text-to-3D generation approaches that directly train on large-scale 3D datasets to enable feedforward generation, such as [1, 2]. Although these approaches are not directly comparable, clarifying this distinction in the related work section would be helpful.\n\nConsidering similar efforts to restrict the noise sampling space, a noise recalibration scheme was introduced in [3]. A discussion on the similarities and differences is recommended.\n\nref:\n\n [1] LGM: Large Multi-View Gaussian Model for High-Resolution 3D Content Creation.\n [2] 3DTopia: Large Text-to-3D Generation Model with Hybrid Diffusion Priors.\n [3] Diverse and Stable 2D Diffusion Guided Text to 3D Generation with Noise Recalibration." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper addresses an important issue - multi-view inconsistency in SDS. \nThe method is well-motivated, and the design is intuitively sound.\nThe proposed method is shown to work well with various SDS variants, including 3DFuse, GaussianDreamer and ProlificDreamer.\n\nThe presentation is clear.\nGood visualizations in Fig. 3, Fig. 4 and Fig. 5." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on optimization based text-to-3D generation. The authors propose a geometry-aware score distillation method to address the multi-view consistency issue in SDS. \nSpecifically, the authors propose to use 3D consistent noising by warping noises of nearby views with the current 3D modelling parameters. Based on this, a correspondence-aware gradient consistency loss is introduced to encourage the multiview consistency of SDS gradients between nearby viewpoints." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "No video demos were shown to validate the generated 3D results.\nSome of the generated results still appeal oversmoothness and unrealistic (Fig. 6, Fig. 11)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 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. Efficient Incorporation of 3D Consistency: The paper directly adjusts the gradients of Score Distillation Sampling (SDS) without requiring 3D data to fine-tune a 2D diffusion model. This allows for the incorporation of 3D consistency with minimal computational overhead.\n\n2. Significant Improvement over Baselines: The proposed method shows noticeable enhancements in both 3D consistency and appearance details when compared to baseline models such as GaussianDreamer, ProlificDreamer, and MVDream." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a method called Geometry-aware Score Distillation. The experimental results demonstrate that the proposed approach can improve performance and address geometric inconsistencies in SDS-based text-to-3D generation tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Dependency on Initial Point Cloud: The method requires establishing correspondences between different views, necessitating an initial point cloud. The paper primarily compares against Gaussian splatting-based methods like GaussianDreamer. According to previous literature [1], when shape initialization is used as a constraint, the Janus problem is substantially mitigated. Therefore, the effectiveness of the proposed method requires more rigorous justification.\n\n2. Comparative Evaluation Concerns: In comparing their method with ProlificDreamer, the authors use 3D-Fuse—originally designed to address the Janus problem—as a baseline. This choice potentially diminishes the perceived effectiveness of the proposed method. Additionally, as observed in Figure 7, the method does not significantly enhance 3D consistency compared to ProlificDreamer; it primarily reduces floating artifacts in the background. Similar issues have been discussed in prior work [2] and can be addressed by adjusting the Classifier-Free Guidance (CFG) in diffusion models. Thus, further experimental evidence is needed to substantiate the authors' claim of enhanced 3D consistency.\n\n[1] Chen, Cheng, et al. \"Sculpt3d: Multi-view consistent text-to-3d generation with sparse 3d prior.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024.\n\n[2] Yang, Xiaofeng, et al. \"Learn to Optimize Denoising Scores: A Unified and Improved Diffusion Prior for 3D Generation.\" European Conference on Computer Vision. Springer, Cham, 2025." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "My major concern is the applicability of the proposed method in latent diffusion models. Please refer to the weaknesses outlined above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well-organized, easy to follow, and clearly presented.\n2. The proposed method offers a fresh perspective on the multi-face Janus problem of the SDS loss, particularly from the standpoint of random noise sampling.\n3. Although the solution is an extension of existing work, it is still novel in the 3D generation field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to address the well-known Janus problem of the SDS loss. Specifically, it proposes modifying the noise sampling process of the SDS loss by replacing random noise with 3D-consistent noise. The 3D-consistent noise is generated by extending the integral noise method of Chang et al. (2024) using an intermediate point cloud representation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My primary concern lies in the use of latent diffusion models, as all experiments in the paper are conducted on latent diffusion models. While I do not doubt that the proposed method could be effective on pixel-space diffusion models, the entire analysis may not be applicable to latent diffusion models.\n\n- **First,** 3D-consistent noise in latent space does NOT equate to consistent gradients in pixel space. Imagine a simple case where the latent map is shifted to the right or left by a few pixels—the VAE would not decode the same image with the same shift in pixel space. In the case of SDS combined with latent diffusion, the gradient at the latent space needs to pass through a VAE encoder composed of multiple convolutional layers and activation functions. As a result, simply keeping the gradient of a portion of the latent map unchanged does not ensure that the corresponding image regions will also receive an unchanged gradient. \n\n- **Second,** this limitation has already been highlighted in the original work of Chang et al. (2024) (see their Appendix E). The authors of the integral noise method in Chang et al. (2024) noted that their method does NOT perform well in latent diffusion models and provided three reasons for this. Could the authors clarify why integral noise fails in the temporal domain but succeeds in the 3D case using LDM?\n\n2. The paper does not address the limitations of the proposed method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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. I recommend including a pseudo-algorithm to enhance comprehension. Is the noise regenerated after each batch of gradient updates? Does the computation of correlation occur solely within the same batch?\n\n2. Are noise particles not visible considered in the computation? If so, why not focus on visible particles, aligning more closely with the computation in $\\int$-noise [1]?\n\n3. There are missing experimental details as noted in weakness 3. How many prompts do you experimented with?\n\n[1] How I Warped Your Noise: a Temporally-Correlated Noise Prior for Diffusion Models (https://warpyournoise.github.io/)" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The proposed noising method establishes a correlation between camera views. The noising methods can generate i.i.d. Gassian noise correctly while maintaining correlation properities.\n\n2. The introduction of the *correspondence-aware gradient consistency loss* serves as a regularization technique to mitigate geometry artifacts." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a method to enhance SDS-based 3D generation through 3D consistent noising and geometry-aware gradient warping. The authors presented an algorithm that could warp the gaussian noise at different camera view while maintaining Gaussian properity for one camera view. The 3D consistent noising aims to improve gradient consistency, while the geometry-aware gradient warping seeks to reduce geometric artifacts." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The motivation behind *3D consistent noising* lacks clarity. Although the authors reference its similarity to $\\int$-noise [1] for improving texture consistency in video generation, the discussion on its application in 3D generation is insufficient. This omission raises questions about how 3D consistent noising can enhance 3D generation specifically. The introduction of 3D consistent noising appears unrelated to the 3D score interpretation of SJC, raising concerns about its relevance and application within this context. Thus, I suggest the authors to provide a more detailed explanation of how 3D consistent noising specifically addresses challenges in 3D generation.\n\n2. The noising method is confined to point cloud (3D GS) representation, limiting its applicability to other forms such as NeRF or Mesh. Additionally, the algorithm appears to overlook occlusion; for instance, a Gaussian particle not visible from the current camera view may still influence noise computation, resulting in questionable correlations compared to surface warp (optical flow) as in $\\int$-noise [1].\n\n3. The experimental evaluation is weak, with qualititive results from both the baselines and proposed methods showing low quality (Fig. 6, 7, 11). The results strongly underperform SoTA implementations like ProlificDreamer (VSD) and LucidDreamer (ISM), undermining the validity of the user study. Furthermore, this paper is lack of quantitative metrics. For the CLIP score in Fig. 13, details on the number of prompts and seeds used in the experiments are missing, and a more thorough experimentation (with at least 25 prompts) is needed.\n\n4. The *correspondence-aware gradient consistency loss* remains difficult to grasp despite the explanations in Sec. 4.4. The rationale behind why sharp changes are likely artifacts needs clarification. I suggest the authors to provide a more intuitive explanation of why sharp changes are likely to be artifacts. Additionally, the ablation study in Fig. 8 fails to effectively illustrate its impact; more results with varied seeds or prompts would strengthen the argument." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024geometryaware,\ntitle={Geometry-aware Score Distillation via 3D Consistent Noising and Gradients},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uyzkKPvVyS},\nnote={under review}\n}" }, "abstract": { "value": "Score distillation sampling (SDS), the methodology in which the score from pretrained 2D diffusion models is distilled into 3D representation, has recently brought significant advancements in text-to-3D generation. However, this approach is still confronted with critical geometric inconsistency problems such as the ``Janus problem''. We provide a novel insight into this problem, hypothesizing that the incorporation of 3D awareness into the 3D noising process and gradient distillation process may bring about enhanced consistency between gradients, leading to improved fidelity and geometric consistency. To achieve this, we propose a simple yet effective approach to achieve a 3D consistent, geometry-aware noising process, leveraging the advantages that 3D Gaussian Splatting possesses as an explicit 3D representation. Combined with our geometry-based gradient warping and our novel gradient dissimilarity loss, we demonstrate that our method significantly improves performance by addressing geometric inconsistency problems in text-to-3D generation with minimal computation cost and being compatible with existing score distillation-based 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", "Score Distillation Sampling", "Text-to-3D Generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/d2e3237d70da157f5de95b5e86a8e41cc390ad5a.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": "Geometry-aware Score Distillation via 3D Consistent Noising and Gradients" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you for your detailed review and for pointing out critical areas of improvement. Your feedback is highly valuable in refining our manuscript. Below, we respond to your comments and outline the revisions we will implement:\n\n1. Details on the IVC Dataset: We acknowledge that the current manuscript lacks sufficient detail about the IVC dataset. In response, we will provide a comprehensive description of the dataset's source, content, and annotation process. This will include details about how the dataset was curated from Parpola’s CISI volumes and Mahadevan’s seminal work, and how manual annotation was performed to ensure data quality.\n\n2. Functionality and Illustration of the ProtoSegment Model: We understand that the description of the ProtoSegment model may not have been sufficiently detailed. To address this, we will include a more thorough explanation of the model’s functionality, emphasizing how the integration of a segmentation encoder contributes to enhanced feature extraction. \n\n3. Sloppy Text and Language Quality: We appreciate your feedback on the language and writing quality. We will conduct a comprehensive revision of the manuscript to improve the clarity, coherence, and academic quality of the text. Specific instances, such as the noted section on page 3 (lines 147-148), will be rephrased for better readability​.\n\n4. Relevance of References: We will carefully review all references to ensure they contribute meaningfully to the content.\n\n5. Ethical Concerns Regarding Acknowledgment Section: Thank you for highlighting the potential ethical issue regarding the acknowledgment section. We will modify this section to comply with ICLR’s double-blind submission policy by removing or anonymizing any grant details to avoid revealing author identity.\n\nResponses to Specific Questions:\n\nReal Contribution of the Model: The key contribution of ProtoSegment lies in its unique integration of a segmentation encoder within the Prototypical Networks framework. This allows for more refined feature extraction and segmentation of input images, particularly benefiting tasks with limited data. We will expand the discussion on this novelty to clearly differentiate it from existing methods​.\n\nProcurement of the IVC Dataset: The IVC dataset was curated from sources such as Parpola’s CISI volumes and Mahadevan’s \"The Indus Script: Texts, Concordance and Tables.\"" }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you for your detailed review and constructive feedback on our paper. We greatly appreciate your thoughtful comments, which provide valuable insights for enhancing the quality and impact of our work. Below, we address your points and outline our planned revisions:\n\n1. Scope and Relevance of the Topic: We understand that grapheme recognition in ancient scripts may appear niche, potentially limiting its perceived influence in broader OCR fields. We aim to show how techniques designed for complex, data-limited environments can contribute to advancements in general OCR and machine learning methods.\n\n2. Contribution and Novelty of the Model: The ProtoSegment model does build upon existing frameworks, but its key contribution lies in integrating a segmentation encoder that focuses on enhancing feature extraction for highly detailed and complex input images, such as those from ancient scripts. This innovation is tailored to handle the challenges posed by limited data, making it more than a simple application of existing methods. We will revise the methodology section to clarify and highlight this unique aspect​.\n\n3. Justification for CNN Architecture: The use of a basic CNN architecture was a deliberate choice to balance computational efficiency with performance, particularly in the context of few-shot learning where simplicity can lead to better generalization on small datasets. We will provide further justification for this choice and discuss its comparative advantages in low-data scenarios​.\n\n4. Clarification of Figure 1: We acknowledge that Figure 1 may be unclear. We will enhance this figure by improving its visual quality and adding a more detailed caption that explains each component and its role in the model architecture.\n\n5. Public Release of the Dataset: We understand the importance of dataset accessibility for reproducibility and wider academic impact. We plan to release the annotated dataset upon acceptance of the paper for transparency and facilitate further research in this field. \n\n6. Experimental Validation and Ablation Studies: Your point regarding the lack of ablation studies is well-taken. We will include an ablation study that dissects the contributions of individual components, such as the segmentation encoder and the CNN backbone, to provide empirical validation of their impact on performance." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you for your detailed review and for providing valuable feedback on our paper. We appreciate your time and insights, which will help us enhance the quality of our manuscript. Below, we address your comments and outline how we will revise the paper:\n\n1. Clarifying Dataset Details: We acknowledge your concern about the lack of clarity regarding the dataset. To address this, we will add more comprehensive information about the creation, annotation, and characteristics of our custom Indus Valley Civilization (IVC) dataset. This will include details on the sources used, annotation tools and processes, and class distribution to provide transparency and better understanding.\n\n2. Novelty of Network Architecture: We understand that the novelty of the proposed ProtoSegment model may not have been sufficiently highlighted. While ProtoSegment builds on the existing Prototypical Networks framework, its key innovation lies in the integration of a segmentation encoder to enhance feature extraction in cases of limited data. This segmentation step allows the model to isolate and focus on meaningful visual features within each region, facilitating better discrimination between graphemes with subtle differences. Unlike traditional Prototypical Networks, which process entire images or use simpler feature extraction methods, ProtoSegment leverages this tailored segmentation to enhance its feature maps before prototype computation. We will revise the method section to more clearly articulate this novelty and its impact on the grapheme recognition task​.\n\n3. Paper Blindness Compliance: We appreciate your observation regarding the paper's anonymity. We will ensure that any potentially identifying information is removed or anonymized to comply with double-blind review standards.\n\n4. Writing Edits and Flow Improvements: To improve readability, we will:\n\n- Edit the paper to ensure a smoother flow of content, with better transitions between sections.\n- Enhance the explanation linking Figure 4 to the relevant text, making the connection clearer and easier to follow​.\n- Revisit the entire paper to standardize abbreviation definitions, ensuring terms like \"CNN\" are defined only once and referenced consistently thereafter​.\n\n5. Reference Formatting and Citation Issues: We will review and correct the formatting of references throughout the manuscript to ensure they are distinct from the main text and properly cited. \n\n6. Response to the Question on Dataset Annotation: To answer your question on dataset annotation, specifically, the dataset was annotated manually. The authors examined each image and with feedback from experts marked individual graphemes according to predefined classification criteria. The annotations followed Mahadevan's taxonomy of labels for the IVC dataset." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": { "value": "Thank you for your thorough review and valuable feedback on our submission. We appreciate your time and constructive comments, which will guide us in improving our manuscript. Below, we address each of your points in detail:\n\n1. Writing and Presentation Issues: We acknowledge that the manuscript contains issues related to incorrect symbols and formulas, confusing citation formats, and the use of lower-resolution figures. To address this:\n\n- We will revise all mathematical notations to ensure they are correct and properly defined within the text.\n- Citation formats will be standardized and checked to prevent blending with the main content, enhancing readability.\n- We will replace all blurry figures with higher-resolution images saved in vector format (PDF), ensuring clarity.\n\n2. Clarification of the Segmentation Encoder's Role: You raised an important question regarding the role of the segmentation encoder. The encoder is indeed a convolution-based encoder-decoder designed to isolate individual graphemes within complex input images. This isolation helps the model focus on relevant features and contributes to improved recognition performance. We will expand the explanation in Section 3.2 to highlight this functionality and its significance in ProtoSegment​.\n\n3. Explanation of MobileNet's Role: We appreciate your observation regarding the lack of clarity about MobileNet’s role in Figure 1. MobileNet is integrated as a component of the initial digitization step to pre-process and refine input character images before they are processed by ProtoSegment. This connection will be elaborated on to provide better context​.\n\n4. Explanation of K-way and N-shot Terms: We will add an explanation in the tables and main text to define K-way (the number of classes) and N-shot (the number of examples per class) for readers who may not be familiar with few-shot learning terminology​.\n\n5. Model Training and Experimental Clarification: To address the confusion regarding Figure 4, where the backward pass is shown pointing to the support and query samples:\n\n- We will revise the figure and description to accurately reflect the training flow and clarify that the support and query samples serve as inputs during the training episodes but are not updated as part of the backward pass.\n\n6. Analysis of Experimental Results: We understand your request for an explanation of why the 20-way results were worse than the 5-way. This discrepancy is primarily due to the increased difficulty of classifying among more classes, which impacts accuracy. We will provide a more detailed analysis in the experimental results section to discuss this phenomenon and its implications​." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "The authors have provided details on the grant which funded their research in an acknowledgment section, this could reveal their identity. Hence I believe this is violating ICLR submission policy." }, "flag_for_ethics_review": { "value": [ "Yes, Other reasons (please specify below)" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "Seems the method described in this paper is just an off-the shelf algorithm - Could you specify what was the real contribution?\n\nWhere from did this IVC dataset was procured? \n\nWhat was the reason for putting the details of the grant in the acknowledgement section?? This is completely against ICLR submission policy as this might reveal the authors identity." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The only positive aspect of this article is the topic." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents \"FLAIR\", a foundational model designed for the grapheme recognition of the Indus Valley script, an ancient un-deciphered writing system. Recognizing the limited availability of labeled data, the authors leverage few-shot learning (FSL) through prototypical networks enhanced with a custom segmentation encoder called ProtoSegment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(i) No details on IVC dataset. \n\n(ii) What is the functionality of the protosegment model is also not properly illustrated and hence the key contribution (if any at all ) also cannot be perceived. This part should have been aided with more illustrative diagrams. \n\n(iii) Sloppy text - for example in page 3 line 147-148. \n\n(iv) Extremely poor language and sentence formation. \n\n(v) irrelevant references - just for the sake of filling up the paper , for example Line 44 in page 1." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "As shown in Weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tFLAIR fills a critical gap in ancient script recognition, providing a versatile model not previously available in OCR or grapheme recognition.\n2.\tProtoSegment outperforms existing few-shot and deep learning methods, achieving higher accuracy in grapheme classification tasks across both datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces FLAIR, a few-shot learning model for recognizing graphemes from the undeciphered script of the Indus Valley Civilization (IVC). Utilizing prototypical networks and a specialized encoder, FLAIR excels at digitizing IVC seal graphemes, outperforming traditional method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe paper focuses on grapheme recognition in ancient scripts, which is a niche topic and represents a small subfield of OCR. This has weak influence on our scholar field, which makes this paper is not suitable for a top-tier conference like ICLR. Additionally, general OCR methods might also perform well on this dataset.\n2.\tThe proposed method largely relies on existing approaches (CNN backbone + Classifier head), merely applying the framework on your dataset. This raises concerns about the contribution and innovation of this work.\n3.\tThe method employs a very basic CNN architecture for the classification task, which seems outdated in the current era of large models. Moreover, referring to it as a \"foundational model\" appears somewhat exaggerated.\n4.\tFigure 1 is also quite unclear.\n5.\tFurthermore, will this paper release the dataset publicly? If not, the lack of innovation in your method significantly diminishes the paper's contribution to the academic community.\n6.\tThe experimental section lacks ablation studies to validate the components of your proposed method.\n7.\tAs shown in Table 1, the accuracy of your method and other state-of-the-art approaches has reached over 98%, even approaching 99%. In such cases of minimal improvement, it is difficult to determine whether the results stem from experimental variability or the enhancements offered by your 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": 2 }, "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": "* How was the dataset annotated?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "* Applying deep learning techniques to scripts could be very beneficial to archaeologist and linguists to help study many undeciphered scripts. This could also be applied to other ancient scripts outside of scripts used in indux valley civilizations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper describes a method, FLAIR, to classify graphemes in ancient Indus valley civilization scripts. FLAIR adopts a few-shot learning approach to circumvent small datasets by introducing the protosegment model designed for images of graphemes. The authors evaluated their technique on an existing dataset OmniGlot as well as their custom Index valley civilization scripts dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* Details about the dataset is not clear.\n* There is limited novelty in the proposed network architecture\n\n* The paper is not blinded\n* The writing of the paper needs edits. e.g.,\n * The flow of the paper is hard to follow. \n * It wasn’t easy to link Figure 4 to the text describing it\n * References are not correctly added throughout the paper\n * Abbreviations are not defined carefully (e.g., convolutional neural network -> CNN was defined 3 times)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "This paper does not involve any ethics issues." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In the paper, all figure uses jpg or png format. The drawn image should be saved in pdf format before being inserted into the paper.\n2. In section 3.2 PROTOSEGMENT MODEL, there are a lot of errors in mathematical symbols and formulas. Some symbols appear out of thin air without explanation, which is not conducive to readers' understanding.\n3. What role does Deep Learning: MobileNet mentioned in Figure 1 play in the entire task? The article does not explain it clearly.\n4. In Tables 1 and 2, it should be explained clearly what K-way and N-shot refer to, as this may be confusing to readers who are not familiar with Prototypical Learning.\n5. There are errors in the reference citation format in the paper, and the content of the references is mixed with the main text, which is not conducive to the reader's reading experience.\n6. Why does the Backward Pass in Figure 4 point to the Support Sample and Query Sample from the network? Aren't these samples extracted from the dataset and cannot be updated?\n7. In Table 1, why is the result of 20-way worse than that of 5-way? I hope to see the explanation of this experimental phenomenon." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper uses meta-learning and few-shot learning to perform classification and recognition tasks on a small sample of Indus Valley Civilization. The article improves Prototypical Networks and proposes ProtoSegment, which achieves state-of-the-art performance on the IVC Dataset. The findings of this paper may contribute to new discoveries and interpretations of ancient texts of the Indus Valley Civilization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The article proposes FLAIR, which uses a few-shot learning method to recognize individual characters from a limited set of Indus scripts. FLAIR uses prototypical networks and ProtoSegment to extract complex features in grapheme images to achieve recognition of Indus script. FLAIR was pre-trained on the Omniglot dataset and then migrated to the recognition and classification tasks of the IVD dataset, achieving state-of-the-art results. FLAIR's ability to perform efficient feature extraction from small samples and its potential adaptability to unseen symbols make it a powerful tool not only to digitize and analyze ancient scripts but also potentially aid in their decipherment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "This paper has major writing problems. For example, there are a large number of incorrect symbols and formulas in the text, the citation format of references in the text is incorrect and confusing, the pictures in the text are blurry, and the model training process is not clearly explained, which will cause great confusion for readers who are not familiar with Prototypical Networks. In terms of innovation, although the paper proposes a relatively novel task, there are few improvements to the methods used. The paper only adds a segmentation encoder to the original Prototypical Networks. If I understand correctly, the segmentation encoder should be a convolution-based encoder-decoder, but I don’t understand what specific role this network can play and why it can segment images into individual graphemes. In the experimental part, the article lacks qualitative analysis of the experimental results. Why is there such a result? What caused the difference in the experimental results? What conclusions can we draw from the experimental results? I think these should be added to the paper." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Updating prototypical networks and applying few-shot learning for Indus Valley script recognition with Omniglot dataset testing and validation" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024flair,\ntitle={{FLAIR}: A Foundation Model for Grapheme Recognition in Ancient Scripts with Few-Shot Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uz4QiNHB16},\nnote={under review}\n}" }, "abstract": { "value": "The Indus Valley Civilization (IVC) left behind an undeciphered script, posing a significant challenge to archaeologists and linguists. This paper introduces FLAIR, a few-shot learning approach that aims to establish a foundational model for recognizing and identifying individual graphemes from the limited available Indus script. As a foundational model, FLAIR is designed to be versatile, supporting multiple potential applications in script recognition and beyond. It leverages prototypical networks combined with a modified proposed encoder network for segmentation, ProtoSegment to extract intricate features from the grapheme images. We evaluate FLAIR’s ability to generalize from minimal data using IVC grapheme classification tasks and further experiment with pre-trained Omniglot models for fine-tuning. Additionally, we simulate real-world data scarcity by intentionally restricting training data on the Omniglot dataset. Our experiments demonstrate FLAIR’s accuracy in digitizing and recognizing Indus Valley seal graphemes, outperforming traditional machine learning classification approaches. These results underscore FLAIR's potential not only for the digitization of ancient scripts with limited labeled datasets but also for broader applications where data is scarce. FLAIR’s success in grapheme recognition highlights its promise as a foundational model capable of extending to other undeciphered writing systems, thereby contributing to the integration of classic scientific tools and data-driven approaches." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Foundation Model", "Few-Shot Learning", "Prototypical Networks", "Encoder Network", "Indus Valley Civilization Script", "Omniglot Dataset" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/5c9a2ba9199b47060126dbfc66e13b3aa3d9d84a.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": "FLAIR: A Foundation Model for Grapheme Recognition in Ancient Scripts with Few-Shot 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": null, "code_of_ethics": null, "comment": { "value": "Thanks for your comments, we decide to make a thorough revision and resubmit this manuscript." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Please refer to the weaknesses part." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed method offers a practical advantage by boosting T2V model performance without retraining, making it suitable for integrating with existing models.\n\n2. The TSG and FME components are well-designed to tackle specific weaknesses in T2V generation, providing both structural coherence and dynamic motion in synthesized videos.\n\n3. The authors validate their methods on multiple backbones like AnimateDiff and VideoCrafter2 illustrating its generalizability across various models, including potential applications in image-to-video tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a training-free, plug-and-play approach called BroadWay to enhance text-to-video (T2V) generation models. The proposed method is designed to address common challenges including structural artifacts, temporal inconsistencies, and limited motion dynamics in T2V models, by manipulating temporal attention maps. The two main components are Temporal Self-Guidance (TSG), which improves structural plausibility and consistency by reducing disparities between attention maps across decoder blocks, and Fourier-based Motion Enhancement (FME), which amplifies motion by increasing the energy in the attention maps through frequency manipulation. Experimental results demonstrate BroadWay's effectiveness across multiple T2V backbones, showing substantial improvements in video quality, structural consistency, and motion richness without additional training or memory cost." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The hyper-parameters, especially \\alpha (for TSG) and \\beta (for FME), may need to be manually tuned for different backbones, reducing the plug-and-play convenience for some users. For example, the \\beta parameter varies greatly for different T2V base models (1.5 for AnimateDiff and 10.0 for VideoCrafter2). Is there any guidance for this? The author should provide a sensitivity analysis or guidelines for selecting these parameters across different backbones. This would help users more easily apply the method to various models.\n\n2. All the T2V backbones are based on the U-Net structure with an interleaved spatial-temporal attention module. How about the model with DiT, as well as full 3D attention? It is encouraged to discuss the potential of the proposed method when applying to models with different architectures like DiT or full 3D attention.\n\n3. It is highly encouraged to use VBench to evaluate the proposed method since this metric could evaluate the jittery and motion magnitude of the videos.\n\n4. UNet should be U-Net." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "How is the hyper-parameter \\tau decided? How does it affect the performance of the method?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The analysis of the temporal attention map is interesting and provides a lot of insights.\n- The paper reads well and is easy to follow.\n- The ideas of both components are interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a training-free method for improving both temporal consistency and motion intensity of video diffusion models. The first component is temporal self-guidance. The analysis shows that disparities between temporal attention maps across different blocks are related to the structure coherence of generated videos. To create videos with better structure coherence and temporal consistency, the temporal attention map of the first upsampling block is added to the attention maps of subsequent blocks. The second component is frequency spectrum re-weighting. The attention maps are decomposed into a high-frequency part and a low-frequency part. To increase the motion intensity, the high-frequency part will be multiplied by a scalar greater than 1." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- There are only ~10 example video results for each model in the supplemental material.\n- For different models (AnimateDiff, VideoCrafter2), the default hyper-parameters are quite different. Users might need heavy manual tuning for these hyper-parameters.\n- Ablation experiments of different values of alpha and beta are missing.\n- VBench is a standard video generation benchmark, but the paper doesn't use the metrics from VBench." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- Will the method be extended to DiT-based video generation.\n- Several relevant training-free enhancement methods, such as Freeu and VideoElevator, also adopted Fourier-based Enhancement. More discussion and comparison is suggested to clarify their difference.\n- Albeit training-free is convenient for use, training usually is beneficial to performance. Some discussion on this issue is preferred." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "+ A training-free method BroadWay to improve the quality of text-to-video generation. \n+ Temporal Self-Guidance and Fourier-based Motion Enhancement. \n+ Effectiveness on several existing video generation models such as AnimateDiff and VideoCrafter2." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presented a training-free method BroadWay to improve the quality of text-to-video generation. BroadWay involves two major components, Temporal Self-Guidance and Fourier-based Motion Enhancement. Experiments show its effectiveness on several existing video generation models such as AnimateDiff and VideoCrafter2." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Will the method be extended to DiT-based video generation.\n- Several relevant training-free enhancement methods, such as Freeu and VideoElevator, also adopted Fourier-based Enhancement. More discussion and comparison is suggested to clarify their difference.\n- Albeit training-free is convenient for use, training usually is beneficial to performance. Some discussion on this issue is preferred." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "In conclusion, considering the weaknesses mentioned above, this paper cannot be considered a well-prepared version for ICLR. Therefore, I lean towards rejecting this manuscript." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The authors demonstrate the effectiveness of BroadWay on various T2V backbones and I2V tasks, showing its versatility and potential for widespread adoption.\n2. This paper is well-organized, and the authors provide a clear explanation of the proposed method, making it easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a training-free method called BroadWay to improve the quality of text-to-video (T2V) generation models without introducing additional parameters, memory, or sampling time. BroadWay consists of two components: Temporal Self-Guidance and Fourier-based Motion Enhancement. The former improves structural plausibility and temporal consistency by reducing the disparity between temporal attention maps across decoder blocks. The latter enhances motion magnitude and richness by scaling the high-frequency components of the temporal attention maps. The authors demonstrate the effectiveness of BroadWay on various T2V backbones, including AnimateDiff and VideoCrafter2, and show that it can also be applied to image-to-video (I2V) tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tFirstly, this paper focuses on methods that do not require training to improve the performance of video diffusion models, which has been explored by existing works, including FreeInit [1], UniCtrl [2], and I4VGen [3]. However, these works are not adequately discussed in this paper. In fact, [1] also explores the motion degree of generated videos from the perspective of high and low frequency decoupling, and [2] studies the attention layer of video diffusion models. However, the omission of these works means that this paper cannot be considered a well-prepared version for ICLR. This also affects the evaluation of the novelty of this paper.\n2.\tExperiments. In fact, there are many benchmarks focused on video generation, such as T2V-CompBench [4] and VBench [5], which provide reliable evaluation metrics. However, this paper hardly provides quantitative evaluation results. For the evaluation of the motion degree of generated videos, the \"Dynamic Degree\" in VBench provides a reference.\n3.\tAblation experiments. Providing only visualization results in Figure 10 is not convincing, and more results, including quantitative results, are necessary. Moreover, in line 215, the authors should provide more discussion on the selection of up_blocks.1.\n4.\tIn the limitations section, the authors also mention that the proposed method is parameter-sensitive, and it would be better to provide more experimental results on this issue.\n5.\tI have doubts about the results in Table 2(b). For AnimateDiff, when LoRA is not introduced, the generated videos are temporally chaotic, corresponding to a higher motion degree. Therefore, FreeInit and others introduce Realistic Vision V5.1 LoRA as the video baseline, which is not mentioned in this paper. Therefore, the authors need to provide more explanations for Table 2.\n\n[1] FreeInit: Bridging Initialization Gap in Video Diffusion Models, Wu et al., ECCV 2024\n[2] UniCtrl: Improving the Spatiotemporal Consistency of Text-to-Video Diffusion Models via Training-Free Unified Attention Control, Chen et al., arXiv 2024\n[3] I4VGen: Image as Free Stepping Stone for Text-to-Video Generation, Guo et al., arXiv 2024\n[4] T2V-CompBench: A Comprehensive Benchmark for Compositional Text-to-video Generationrk, Sun et al., arXiv 2024\n[5] VBench: Comprehensive Benchmark Suite for Video Generative Models, Huang et al., CVPR 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": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No ethics review needed." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. I am concerned about the reasoning and justification for the proposed method.\n a. In L83, the authors claim temporal inconsistency and motion artifacts are related to temporal attention map disparity. However, why such disparity lead to artifacts? In Sec 4.2, the authors also modify and amplify several frequency components, why such operation not lead to disparity and degrades results?\n b. In L87~89, the meaning of `energy` is not defined and unclear. In the second row of Fig. 2, the background region slightly translates, but shows a large response similar to foreground cat, why is that? Is this `rich motion`?\n c. In L210, why to choose up blocks.1 as anchor? Does temporal attention maps in different blocks share similar meaning and can be computed as in Equ (3)? \n d. In Sec 4.2, why the authors choose to modify in frequency space? The artifacts or motion inconsitency do not necessary happen in frequency domain.\n e. In Equ (5), A is a 3D tensor, it is not clear how to do the fourier transform. Does that mean a 1D FFT along the last axis?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper tries to solve an important question of motion quality in video generation. The authors propose a new method to modify temporal attention maps using self-guidance and frequency spectrum re-weighting. The method itself seems reasonable and interesting. \n2. The results shown in the paper seem promising and superior to baseline model.\n3. The paper writing is clear and very easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces BroadWay, a training-free method to improve T2V video quality, especially structural plausibility, temporal and consistency. BroadWay consists of two components: Temporal Self-Guidance, which enhances structural plausibility and temporal consistency by reducing disparities in temporal attention maps, and Fourier-based Motion Enhancement, which amplifies motion by increasing the energy in the maps. Experiments and results seem promising." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. My primary concern lies with the reasoning and justification for the proposed method. Many of the descriptions and claims in the paper appear to be rather ad hoc, lacking detailed analysis or theoretical proof to substantiate their validity.\n\n2. The authors clearly described their method. However, it is not very clear for the reason and hyper-parameter choice for each step.\n I think there must be some explanation under these steps, and it would be much better if more discussion and explanations are provided.\n\n3. The results in the main paper seem promising. However, the videos in the supplementary material seem to be slightly better, especially in consitency and motion magnitude, but still contain artifacts. I wonder maybe the proposed temporal attention map modifications are not essential to this problem?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "BroadWay provides a training-free and plug-and-play option to enhance the overall quality of current T2V backbones." }, "_bibtex": { "value": "@misc{\nbu2024broadway,\ntitle={BroadWay: Boost Your Text-to-Video Generation Model in a Training-free Way},\nauthor={Jiazi Bu and Pengyang Ling and Pan Zhang and Tong Wu and Xiaoyi Dong and Yuhang Zang and Yuhang Cao and Dahua Lin and Jiaqi Wang},\nyear={2024},\nurl={https://openreview.net/forum?id=uzKG83YJ3t}\n}" }, "abstract": { "value": "The text-to-video (T2V) generation models, offering convenient visual creation, have recently garnered increasing attention. Despite their substantial potential, the generated videos may present artifacts, including structural implausibility, temporal inconsistency, and a lack of motion, often resulting in near-static video. In this work, we have identified a correlation between the disparity of temporal attention maps across different blocks and the occurrence of temporal inconsistencies. Additionally, we have observed that the energy contained within the temporal attention maps is directly related to the magnitude of motion amplitude in the generated videos. Based on these observations, we present BroadWay, a training-free method to improve the quality of text-to-video generation without introducing additional parameters, augmenting memory or sampling time. Specifically, BroadWay is composed of two principal components: 1) Temporal Self-Guidance improves the structural plausibility and temporal consistency of generated videos by reducing the disparity between the temporal attention maps across various decoder blocks. 2) Fourier-based Motion Enhancement enhances the magnitude and richness of motion by amplifying the energy of the map. Extensive experiments demonstrate that BroadWay significantly improves the quality of text-to-video generation with negligible additional cost." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Jiazi_Bu1", "~Pengyang_Ling1", "~Pan_Zhang1", "~Tong_Wu2", "~Xiaoyi_Dong1", "~Yuhang_Zang1", "~Yuhang_Cao3", "~Dahua_Lin1", "~Jiaqi_Wang1" ] }, "authors": { "value": [ "Jiazi Bu", "Pengyang Ling", "Pan Zhang", "Tong Wu", "Xiaoyi Dong", "Yuhang Zang", "Yuhang Cao", "Dahua Lin", "Jiaqi Wang" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "generative models", "text-to-video generation", "video quality enhancement" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "bu|broadway_boost_your_texttovideo_generation_model_in_a_trainingfree_way" }, "pdf": { "value": "/pdf/5c556dd169d39408a8439dd58e08a5983703161d.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/c66068bfdbd36567c6250eae8c42f33c35fccb3b.zip" }, "title": { "value": "BroadWay: Boost Your Text-to-Video Generation Model in a Training-free Way" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1) Results from Table 1 shows that the proposed method is only marginally better than RivaGAN w/o real images conditioning. Did you check RivaGAN with the same setup as the last row? \n2) Why do you compare only to SVD in Table 2? \n3) The authors appeal to decreased visual quality of videos generated with the existing watermarking methods, but did not organise subjective study to show that their method outperforms the others. Video quality metrics are known to have limited capabilities to estimate AIGC. Subjective comparison is required in this work, can you provide one?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) The proposed method does not require additional training, which could simplify integration with existing diffusion models.\n2) The framework’s ability to detect tampering both within frames and across the sequence" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper addresses the need for the control of integrity and misuse of AI Generated Content. The usual approach is to use watermarks for video domain, but they are underdeveloped and have a post-processing manner, which results in video quality degradation. The authors propose a novel way of watermarking images while generating it (VideoShield). Their method is training-free and works with diffusion-based video models. Authors extract the watermark from the images using DDIM Inversion." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) Notation T_temp for a threshold in tampering is a bit confusing, considering T_p is a tensor.\n2) No introduction for chacha20\n3) No formula for calculating video quality is given \n4) Figure 2 provides only abbreviations in the legend\n5) Over all paper, formulas and tables seem to have decreased font. Moreover, some tables overlap the text (e.g. Table1). This may be a reason for a possible desk rejection." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Can the author provide more comparative details on how VideoShield compares to other watermarking frameworks or post-processing methods in terms of tamper localization accuracy and robustness?\n2. Can the author share the computational performance of VideoShield, especially regarding the running time of different video resolutions or models?\n3. Will the author consider testing the robustness of VideoShield under other types of distortion, such as extreme video compression, frame rate variations, or color adjustments? These additional distortions are common in real-world applications and can enhance confidence in VideoShield's resilience.\n4. The results indicate that VideoShield has a certain dependence on video quality. Does the performance of VideoShield still vary due to resolution or model complexity? If so, can the author provide more insights or data on these factors?\n5. Can VideoShield adapt to image tampering localization?\n6. Is the watermark capacity of 6 512 bits suitable for all video resolutions, or is the number of bits that can be embedded flexible based on the characteristics of the video?\n7. Placing visual content in the main text seems better." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The paper introduces an innovative approach to watermarking in diffusion-based video generation models by embedding watermarks during the generation process, which diverges from traditional post-processing methods. Compared to image watermarking, this is a less explored field. The originality of embedding watermarks during the generation process, as well as the novel dual tampering localization, is a meaningful supplement to this field.\n2. The technical methods are rigorous and fully described. This paper uses DDIM inversion to provide a solid foundation for watermark embedding and extraction without affecting video quality. The extensive experimental evaluation of multiple T2V and I2V models strongly demonstrates this method's robustness, flexibility, and effectiveness. The author's detailed analysis of different watermark extraction and localization scenarios further strengthens the contribution of this article to this field.\n3. This paper is well-organized and provides a clear understanding of motivation, methods, and experimental setup.\n4. VideoShield's training-free and high-fidelity watermarking method provides a reliable and efficient solution for generating watermarks in video models. This paper addresses a highly relevant issue - ensuring the integrity of content in videos generated by artificial intelligence." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper \"VideoShield: Regulating Diffusion-Based Video Generation Models via Watermarking\" presents VideoShield, a novel framework for embedding watermarks in diffusion-based video generation models, including both text-to-video (T2V) and image-to-video (I2V) models. The paper addresses the need for content control and authenticity verification in AI-generated video content, focusing on integrating watermarks during the video generation process to prevent quality degradation. Key contributions include In-Generation Watermark Embedding, Tamper Localization for Videos, and Watermark Extraction and Robustness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although this article introduces relevant watermarking and tampering localization methods, there is no comparative analysis with other video generation or post-processing watermarking methods. For example, including baselines for time tampering localization or using alternative methods to evaluate the robustness of specific types of distortions.\n2. VideoShield is positioned as a non-training and efficient framework, but this paper lacks specific comparisons with baselines regarding time/computational complexity, such as runtime.\n3. Although these experiments covered various distortions, they did not explore broader real-world attack scenarios, including testing with denser video compression techniques, color distortion, or frame rate changes, which could further validate the robustness of VideoShield in various real-world environments. In addition, analyzing the performance of VideoShield under adversarial conditions where attackers actively attempt to bypass watermarks could be an interesting solution.\n4. The author can further discuss the limitations by listing the performance of VideoShield under different video quality and generation settings. This method seems to perform better on higher-quality video output, but further discussion on factors that may affect watermark integrity, such as video resolution or content complexity, will further demonstrate the effectiveness boundary of VideoShield.\n5. This article briefly introduces the adaptability of image watermarking, but can VideoShield adapt to image tampering localization scenarios?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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": "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 is well-written.\n- This paper proposes a novel scheme for proactive video forensics." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces VIDEOSHIELD, a novel watermarking framework for diffusion-based video generation models. By embedding the watermark directly during the generation process, VIDEOSHIELD eliminates the need for additional training, providing a cost-effective solution for safeguarding generated videos. Additionally, the model includes a proactive tamper detection mechanism. Using template bits derived from the watermark, the authors enable both temporal and spatial localization of tampering within the video." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tThe proposed model may not entirely align with watermarking application scenarios. In its context, verification requires the recipient to compare the watermarked video against the original video's bit template. This requires the verifier to either have access to the original video to generate the bit template using the same encryption method, or to obtain the bit template directly from the video creator. Such requirements differ from standard watermark extraction practices, where verifiers typically can extract watermark information without needing the original video.\n\n2.\tThe model uses the CHACHA20 stream cipher to generate template bits, requiring the algorithm to use the same key with different nonces for each encryption. Does this mean that a unique random number needs to be set for each video? If so, how does the method handle the storage of a large number of video-to-key mappings? Additionally, it’s unclear whether the primary goal of this approach is to protect the model itself or the generated videos.\n\n3.\tThe proposed method uses watermarking for proactive tamper detection; however, the baseline experiments in the paper compare it primarily with passive detection methods (e.g., MVSS-Net, HiFi-Net), which may not be entirely appropriate. It would be more suitable to compare the approach with other active tamper detection methods. Below are references to such methods for consideration:\n[1] Zhang X, Li R, Yu J, et al. Editguard: Versatile image watermarking for tamper localization and copyright protection[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 11964-11974.\n[2] Zhou Y, Ying Q, Wang Y, et al. Robust watermarking for video forgery detection with improved imperceptibility and robustness[C]//2022 IEEE 24th International Workshop on Multimedia Signal Processing (MMSP). IEEE, 2022: 1-6.\n\n4.\tThere are formatting issues: the bottom line of Table 1 overlaps with the text below, and the font size in Table 3 is too small, making it difficult for readers to follow." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 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 framework looks heavy, how about the computational overload? \nThe relationship with \"Gaussian Shading\" needs to be further clarified." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The topic is interesting. Using DDIM Inversion, the video can be reversed to its original watermarked noise. The paper is written and organized well. Experiments are plenty and convincing. The ablation study verified the contribution of each design towards the whole framework." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a VideoShield, which embeds watermarks directly during video generation. It maps watermark bits to template bits, which are then used to generate watermarked noise during the denoising process. Template bits allow precise detection for potential spatial and temporal modification. Extensive experiments demonstrate its effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "No obvious drawbacks are found. However, typesetting needs to be improved, such as keeping the same or similar text size of the table as the text, no overlapping between table and text." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Address the weakness, especially the novelty issue." }, "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 introduces a Gaussian Shading watermark embedding method that does not require training, thereby avoiding additional computational overhead.\n2. Unlike previous methods that focus solely on spatial tampering localization, VIDEOSHIELD takes into account both temporal and spatial tampering localization during the tampering detection process. The paper introduces Hierarchical Spatial-Temporal Refinement, which enables the generation of more detailed masks for tampering localization, significantly improving the accuracy of the detection." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The watermark can actively identify generated content and is currently the primary method used for detection and attribution in diffusion models. However, there has been no watermarking method specifically addressing video generation models based on diffusion models. This paper proposes VIDEOSHIELD to fill this research gap. VIDEOSHIELD introduces a Gaussian Shading watermark embedding method that does not require additional training. In addition to its basic watermarking functionality, VIDEOSHIELD also acts as a fragile watermark, enabling both temporal (across frames) and spatial (within individual frames) tampering localization. Furthermore, it introduces Hierarchical Spatial-Temporal Refinement to enhance accuracy. Experimental results demonstrate that VIDEOSHIELD performs effectively on both T2V and I2V diffusion models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The innovation presented in this paper is moderate. It extends Gaussian Shading from images to videos and further explores its potential as a fragile watermark. \n2. The presentation of the text has significant issues, particularly in Section 3.4, \"TAMPER LOCALIZATION.\" In this section, the extensive use of symbols due to the inclusion of formulas creates obstacles to understanding the paper's content. The frequent reuse of subscripts p and q makes reading quite difficult; the matrix subscript notation is inconsistent, such as in C_(p(q=M(p))), which is recommended to be changed to C_(p,M(p)). Additionally, the comparison bits matrix Cmp could be written as CMP to standardize with other matrix symbols and to avoid confusion with the subscript p. In line 334, the letter m is used both for the variable and for watermark information, which may lead to confusion. After reviewing the main text, I found that in Supplementary Material A3.1, the authors provide a simple example of the process. I strongly recommend including Figure 7 in the main text to aid reader comprehension. Furthermore, there is some overlap between Table 1 and the Datasets section.\n3. In my view, the authors have not addressed an important issue regarding the sequence of watermark attribution and tampering localization. The paper does not demonstrate the accuracy of the watermark after spatial tampering. If we assume that a video undergoes spatial tampering and the tampered areas are detected through localization, but the watermark attribution fails, we cannot conclude that a third party maliciously altered the generated video. This raises the question of whether the watermark has become ineffective in such a scenario." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- In Table 1 and other tables, the metric \"Video quality\" is mentioned. How is it specifically calculated? The authors mentioned it is the average of a series of metrics; if possible, could you provide the specific values for each metric?\n- Currently, the VideoShield method can hide 512 bits. What is the upper limit for watermark capacity?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- VideoShield leverages the unique properties of active watermarking, innovatively combining the model watermarking task with the tampering localization task, demonstrating significant performance advantages.\n- This method requires no training and follows a zero-shot paradigm, making it easy to reproduce and validate its performance.\n- This method is based on the properties of diffusion and Gaussian distribution, allowing it to be generalized to other diffusion-based models, such as text-to-image models, exhibiting good generalization capabilities." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This article proposes a zero-shot video generation model with active watermarking for video copyright protection and tampering localization. Based on the DDIM inversion process, the watermark information hidden in the initial Gaussian noise can be detected again from the generation results, and this active watermark is robust against video tampering, image degradation, and other attacks. Furthermore, for the task of video tampering localization, the authors have designed a detection method that localizes tampering in both temporal (frame position) and spatial (image information) aspects. Experiments have proven the performance advantages of VideoShield." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The threshold and hyperparameter settings in the method section are mostly the result of manual searches. The authors may consider exploring how these thresholds could be adaptively searched in the future." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024videoshield,\ntitle={VideoShield: Regulating Diffusion-based Video Generation Models via Watermarking},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=uzz3qAYy0D},\nnote={under review}\n}" }, "abstract": { "value": "Artificial Intelligence Generated Content (AIGC) has advanced significantly, particularly with the development of video generation models such as text-to-video (T2V) models and image-to-video (I2V) models. However, like other AIGC types, video generation requires robust content control. A common approach is to embed watermarks, but most research has focused on images, with limited attention given to videos. Traditional methods, which embed watermarks frame-by-frame in a post-processing manner, often degrade video quality. In this paper, we propose VideoShield, a novel watermarking framework specifically designed for popular diffusion-based video generation models. Unlike post-processing methods, VideoShield embeds watermarks directly during video generation, eliminating the need for additional training. To ensure video integrity, we introduce a tamper localization feature that can detect changes both temporally (across frames) and spatially (within individual frames). Our method maps watermark bits to template bits, which are then used to generate watermarked noise during the denoising process. Using DDIM Inversion, we can reverse the video to its original watermarked noise, enabling straightforward watermark extraction. Additionally, template bits allow precise detection for potential spatial and temporal modification. Extensive experiments across various video models (both T2V and I2V models) demonstrate that our method effectively extracts watermarks and detects tamper without compromising video quality. Furthermore, we show that this approach is applicable to image generation models, enabling tamper detection in generated images as well." }, "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", "watermarking", "tamper localization" ] }, "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/7a72970c782630a628e0cd92de170187e69e2e99.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": "VideoShield: Regulating Diffusion-based Video Generation Models via Watermarking" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Have you considered evaluating Selective Attention on other architectures, such as encoder or encoder-decoder transformers? If so, could you share any preliminary results or insights?\n\n2. How sensitive is the performance of Selective Attention to the memory budget per layer? Could you offer guidelines on tuning this budget for tasks beyond language modeling?\n\n3. Have you tested Selective Attention on tasks other than language modeling, such as question answering or text summarization? How might the approach handle tasks with varied context and attention needs?\n\n4. Could you provide further comparisons between Selective Attention and other efficient attention methods, such as sparse or linear attention, to clarify the unique advantages of Selective Attention?\n\n5. How does context pruning with Selective Attention affect tasks requiring long-range dependencies? Are certain types of tokens or information more prone to being pruned?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Efficient Memory Management**: The Selective Attention mechanism effectively prunes unneeded tokens, significantly reducing memory usage during inference without degrading model performance. This efficiency gain is particularly valuable for scaling transformers in resource-constrained environments.\n2. **No Additional Parameters**: Selective Attention operates without introducing new parameters or significantly increasing computational overhead, which preserves the simplicity of the transformer architecture.\n3. **Theoretical Motivation and Experimental Rigor**: The paper provides a thorough theoretical motivation for selective attention, supporting it with extensive experimental results on benchmarks like HellaSwag and C4, which demonstrate the effectiveness of the approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a parameter-free modification to the transformer attention mechanism called Selective Attention, which reduces the attention to irrelevant or unneeded tokens in a sequence. This approach enhances performance in language modeling tasks by reducing memory usage and computational costs without compromising model quality. Experiments show that transformers with Selective Attention achieve comparable performance to larger models with traditional attention mechanisms, providing improved efficiency and effectiveness across various model sizes and tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Limited Scope of Model Architectures**: The experiments are primarily conducted on decoder-only transformer models. Further analysis is needed to verify if Selective Attention can similarly benefit encoder models or encoder-decoder models used in other tasks, such as translation or summarization.\n2. **Potential Over-Reliance on Hyperparameter Tuning**: Selective Attention’s performance may depend on optimal memory budget settings per layer, which could complicate deployment in different tasks or models. Although a memory reduction process is described, further tuning could be required in practical applications.\n3. **Dataset and Task Diversity**: While Selective Attention shows improvements in language modeling tasks, testing it on a wider range of tasks (e.g., text generation or long document understanding) would strengthen the case for its generalizability and adaptability to diverse applications.\n4. **Comparison with Similar Methods**: The paper could benefit from a more detailed comparison with other recent efficient attention mechanisms, such as sparse attention or adaptive attention approaches, to highlight any relative advantages Selective Attention may have." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- In the selection attention computation, why use a single head’s logits to mask all heads? Did you try keeping each head separate (i.e. using the logits head i for the selection mask of head i)?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Given the simplicity of the proposed approach and the reported performance gains, it seems that selective attention could be a significant addition to the transformer architecture if it is further validated.\n- Even without the performance gains, the efficiency gains (particularly without having to modify the pretraining loss) are quite relevant and make Selective Attention look like a viable alternative to other efficient attention mechanisms." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces *Selective Attention*, a parameter-free modification to the standard attention mechanism in transformers which helps reduce the attention given to irrelevant elements in the context. To do so, they introduce a *selection mask* in the attention computation to mask “irrelevant” token, and propose using the previous tokens’ attention on other tokens (on a specific head) as masking for future attention computations. On language modelling on C4, results show that transformers equipped with Selective Attention can achieve comparable performance to standard transformers with twice the attention heads and parameters, and slightly better performance on the downstream task of HellaSwag.\n\nThe authors then propose a method that uses the selection mask to prune elements from the attention's context buffer, reducing memory and computational demands during inference, and show that it can lead to significant memory saving (up to 47x for large context sizes) while preserving performance of non-selective attention transformers." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The main problem of this paper is weak experimental validation. The authors only show gains in the language modelling task (using the relatively noisy and deprecated C4 dataset) and on a single downstream task (where they show smaller gains). They don't compare to existing pretrained models or other efficient attention mechanisms. While the (limited) experimental results are promising, they are not enough to validate the approach. I suggest replicating the recipe of an existing (state-of-the-art) pretrained LLM, replacing the attention mechanism with Selective Attention. Additionally, in the current LLM area, more downstream tasks need to be tested to validate the approach. For the efficiency section, the authors should compare with other efficient attention mechanisms.\n- The presentation paper could also be better: it uses a whole page for a single figure and further page for intuition, both of which could be much shorter. It then moves decently important plots to the appendix. I suggest restructure the paper to give more space to the experimental results and comparison to other 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": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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 motivation is clear, and the execution of the idea to not to tokens that are already attended to is clever and clean\n- The proposed method is easy to implement, adds no parameters, and saves memory overhead\n- Interesting results and analysis" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper argues that, for transformer-based language models, selectively attending to a subset of the tokens in the context can improve the model's quality. Motivated by this intuition, the paper proposes a new attention technique, where if a token has already received high attention weights by previous tokens, it means that its content has already been \"absorbed\" in the model's representations, and therefore it should receive less attention weights onward. The paper proposes an implementation for this idea by reusing the attention logits of one of the attention heads, and thereby adding no additional parameters and only a small amount of computational overhead. Further, the proposed method also allows for pruning some tokens from the context when they will never be attended to again, which leads to memory saving.\n\nExperiments with a 100M transformer based LM trained on the C4 dataset suggests that the proposed approach achieves promising results on language modeling perplexity and one downstream task (HellaSwag)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- My main concern is about the limited empirical evaluation where only one downstream task is considered. As the paper argues, \"different tasks have different requirements,\" it is crucial to explore whether selective attention is broadly applicable by evaluating it on a diverse set of tasks with different requirements. This can be complemented with synthetic evaluations that might require the model to store the entire sequence, e.g., counting the number of a certain token in a sequence. \n- Besides a more comprehensive evaluation, I also encourage the authors to strengthen their findings by trying selective attention on larger-scale models and 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": 4 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The authors mention that they compute the S matrix using \"one of the existing heads\" (#167). Their code indicates that they simple take the first one. Is this ablated in any way? Does taking a different head / an average of all heads make much difference?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The method is simple and intuitive\n- Results are very strong and consistent, showing both accuracy improvements, and very impressive memory reduction\n- Experiments are quite extensive\n- Given these results, overall I tend to agree with the authors' statement in Section 9: \" Given that it adds no new parameters, only a negligible amount of compute, and provides consistent improvements, *selective attention might be a good default for transformer decoders*.\"" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents selective attention, a method for adjusting the attention weight of tokens based on the amount of attention they received from previous tokens. The authors also propose to use this method to drop some of the history tokens with the lowest (adjusted) weight, also proposing a new loss term to transformer-based LLM training. Extensive experiments show that this method both consistently improves performance, and reduces memory requirements substantially." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I like this paper a lot, I have several reservations.\n\nFirst, the level of novelty is not huge. The idea of pruning the least important token in each iteration is similar in spirit to cache-pruning methods such as H20 (Zhang et al., 2023), TOVA (Oren et al., 2024) and others. The authors basically apply a similar approach at training time. While this method works great (as noted above), which is a sufficient contribution on its own, and also present some novel concepts (e.g., the new loss term), it would be better to tone down the novelty claims a bit.\n\nSecond, I am not sure I am fully convinced by some of the major claims in the paper.\n \n(a) the examples in Section 2 are very appealing, but they are missing a critical baseline: how much attention does the vanilla transformer assign the pruned tokens in such cases? Perhaps it already knows it needs to ignore them? Again, the method works great so obviously something good is happening, but it is not clear the intuition is capturing the essence of it.\n\n(b) similarly, the authors say (#113) \"if token b has determined that token a is irrelevant or even misleading to future tokens such as c, there is nothing it can do in the given layer to correct for this.\". But I am not sure how their method allows the transfer of such negative signal. If I understand correctly, the authors penalize tokens that received high attention in previous steps (I.e., their value in F is high). I am not sure how information about irrelevant or misleading tokens can be propagated in such cases. \n\nThird, while the writing is generally clear, I found section 3 a bit confusing, requiring several reads to understand a fairly simple method. For instance, it is never explicitly mentioned that each token is penalized by the amount of attention it got from all subsequent tokens. I would also advise the authors to formally define masking, as it is not entirely trivial in this context." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024selective,\ntitle={Selective Attention Improves Transformer},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v0FzmPCd1e},\nnote={under review}\n}" }, "abstract": { "value": "Unneeded elements in the attention's context degrade performance. We introduce Selective Attention, a simple parameter-free change to the standard attention mechanism which reduces attention to unneeded elements. Selective attention consistently improves language modeling performance across model sizes and context lengths. For example, a range of transformers trained with the language modeling objective on C4 with selective attention perform equivalently to transformers with standard attention modules with ~2X more parameters and heads. In addition, selective attention allows reducing the size of the attention's context buffer, leading to substantial reductions in the memory and compute requirements during inference. For example, transformers with 100M parameters and context sizes of 512, 1,024, and 2,048 need 16X, 25X, and 47X less memory for their attention module, respectively, when equipped with selective attention, as those without selective attention, with the same validation perplexity." }, "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": [ "selective attention", "attention", "transformer", "llm", "language model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/e7294ff543e60337e66031c2be9d624a96b4b63b.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": "Selective Attention Improves Transformer" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. How does the computational complexity of ASLDiff compare to other state-of-the-art methods, especially for large-scale networks?\n\n2. Can the authors provide more insights into the model's performance on large real-world networks (e.g., millions of nodes)?\n\n3. Why the performance of ASLDiff under the SIS diffusion pattern is inconsistent?\n\n4. Can the proposed model handle dynamic networks where the topology changes over time?" }, "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 combines diffusion models with principles of information propagation, offering a unique solution to the source localization problem.\n\n2. ASLDiff leverages pre-calculated source estimations as informative priors, potentially improving efficiency and effectiveness.\n\n3. The authors test their model on both synthetic and real-world datasets, comparing it against several state-of-the-art methods.\n\n4. The proposed model shows promise for some real-world scenarios, such as epidemiology, and cybersecurity." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces ASLDiff, a new method for finding information sources in complex networks. ASLDiff combines diffusion models with information propagation principles to accurately locate sources across different network types and patterns. It uses pre-estimated source locations as guides, a diffusion process led by these estimates, and a GCN (Graph Convolutional Network) to capture key propagation details. ASLDiff outperforms certain existing methods, achieving up to 7.5%-12.1% higher accuracy on some real-world datasets and adapting effectively to some networks and scenarios." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the authors use some real-world datasets, more extensive testing on diverse real-world datasets could further validate the model's effectiveness.\n\n2. The paper doesn't discuss the computational complexity or resource requirements of ASLDiff compared to other methods.\n\n3. The paper does not sufficiently address the scalability of the proposed method for large networks (millions of nodes), which is essential for real-world applications. The largest tested graph contains fewer than 15K nodes.\n\n4. ASLDiff’s performance under the SIS diffusion pattern is inconsistent. Only three datasets are used, with performance varying across datasets and metrics. For instance, there is no improvement in AC on the Net and Jazz networks, and only a marginal increase from 0.984 (GCNSI) to 0.985 on the Power network." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 authors’ research scope is so wide that it may be incredible. They claim that they can solve problems in various fields, such as disease outbreaks, network security, etc. This is a great plan, however, maybe it is an exaggeration. The spread of infectious diseases and the spread of misinformation vary greatly. Maybe it is too ideal to solve these problems with only one model? This is my main concern for this paper, as complex networks are a very vast and intricate field.\n\nThe author's review is complete, but there are relatively few comparative methods published in the year 2024. Compared with the newest algorithms in the year 2024, can ASLDiff still maintain its advantage?\n\nAfter setting up the required environment for the author's code, I encountered an error stating \"models.guide: No such file or directory.\" I can't verify the results of the model, even though it is impressive." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "ASLDiff can train based on the few-shot learning.\n\nThe appendix contains a sufficient review of related work, indicating the adequate preparations for the proposed ASLDiff.\n\nASLDiff using a single snapshot is superior to multiple-snapshot-based works. I think it is awesome!" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In the social network background, the author employs the diffusion model to implement source detection tasks across different localization scenarios. The proposed ASLDiff can be trained in a few-shot manner." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "One of the contributions of this paper is the development of few-shot learning due to the limited data available in real-world scenarios. This should be a key focus for the author; however, it seems that the main body of the paper provides a very limited description of few-shot learning. Moreover, despite the thorough survey and preparations of the work, there are many details that need attention. For example, 'Diffusion' instead of 'Diffsion'; Equation (1), Equation (7) instead of Equation 1, equation 7. Also, providing code is intended to enhance confidence for your work, but the code is not runnable." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "see Weaknesses" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. This paper provides comprehensive experimental validation of ASLDiff.\n2. The code of ASLDiff is given.\n3. ASLDiff shows significant improvement on some datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes an Adaptive Source Localization Diffusion Model to face the challenge of data scarcity without specific propagation models. And evaluations of various propagation patterns and real network datasets demonstrate ASLDiff’s effectiveness." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. As the authors mention, the diffusion model is highly complex. However, the extent of this complexity is not quantified. Additionally, considering the high complexity, the motivation for using the diffusion model should be fully explained.\n2. The authors discuss the challenge that \"real-world networks typically exhibit unknown propagation patterns\", but they do not explain or demonstrate how ASLDiff understands different propagation patterns in different scenarios.\n3. Sections 2.1 and 3.2 are confusing, as they both seem to explain the IC and LT models redundantly.\n4. The authors have demonstrated that the closeness centrality of infected nodes and sources is very consistent, but it appears that the advisor used in ASLDiff is not based on closeness centrality." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "If LPSI is used to generate enough training samples to train the DL-based method, is it possible to achieve an approximation with adaptability, and what are the advantages of using a diffusion model in comparison?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "S1. The problem of source localization is important.\n\nS2. The problem statement and the rationale behind the method are clear.\n\nS3. The paper is a reasonable attempt to apply diffusion modelling to traditional propagation problems.\n\nS4. The evaluation is very comprehensive. The authors show strong quantitative results. Tables 1-2, with figs 3-6 give strong qualitative results and ablation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the source localization problem: given a graph with infection states of all nodes, aim to find the original propagation source. The paper proposes ASLDiff which utilizes diffusion model for source localization in complex networks and applies GNNs to enhance the model’s adaptability to diverse network topologies. Besides, it incorporates soft labels and a restructured label propagation process to capture essential propagation characteristics across various network topologies. However, the technique novelty is limited and methods are capped which subject to labelling dissemination methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. The forward process (e.g. the diffusion process) should be detailed. What is the state of the initial $X$ and what is the $X$ after noise added? Besides, it seems that you used the conditional diffusion method, but your condition is an approximate solution obtained from LPSI. This usually does not conform to our intuition, and I hope you can elaborate on the rationality of using the approximate solution of LPSI as a condition. And will using approximate solutions as conditions greatly limit the performance of your proposed model? If this impact does not occur, you should provide some theoretical proof. This is evident in the accuracy results, where ASLDiff underperforms compared to LPSI on some metrics.\n \nW2. The description of the method in the paper is unclear, and the innovative aspects of the technical details are not sufficiently articulated. For example, the part of the LPSI advisor that employs the conditional diffusion model lacks an explanation of how the diffusion model introduces innovation for the specific task.\n\nW3. The framework diagrams in the paper are not sufficiently clear. For instance, in Figure 1, the function $f(\\theta)$, which corresponds to the denoising network, is not intuitive. It is suggested to use clearer module diagrams to explain this. Additionally, the input-output relationships in Figure 2 are unclear. For example, the concept of $Y_{label}$ is not explicitly defined in the text, even though the generation method is explained. Further clarification is recommended." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a diffusion-based source localization method that can directly applied to real-world data in zero-shot manner after pretraining on simulation data with known propagation patterns and simple network topology." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024adaptive,\ntitle={Adaptive Source Localization on Complex Networks via Conditional Diffusion Model},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v0O9FrVTt1},\nnote={under review}\n}" }, "abstract": { "value": "Network propagation issues like the spread of misinformation, cyber threats, or infrastructure breakdowns are prevalent and have significant societal impacts. Identifying the source of such propagation by analyzing snapshots of affected networks is crucial for managing crises like disease outbreaks and enhancing network security. Traditional methods rely on metrics derived from network topology and are limited to specific propagation models, while deep learning models face the challenge of data scarcity. We propose \\textbf{ASLDiff}~(\\textbf{A}daptive \\textbf{S}ource \\textbf{L}ocalization \\textbf{Diff}sion Model), a novel adaptive source localization diffusion model to achieve accurate and robust source localization across different network topologies and propagation modes by fusing the principles of information propagation and restructuring the label propagation process within the conditioning module. Our approach not only adapts to real-world patterns easily without abundant fine-tuning data but can also generalize to different network topologies easily. Evaluations of various datasets demonstrate ASLDiff's superior effectiveness, accuracy, and adaptability in real-world applications, showcasing its robust performance across different localization scenarios. The code can be found at https://anonymous.4open.science/r/ASLDiff-4FE0." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Diffusion Model", "Knowledge Informed Machine Learning", "Source Localization", "Complex Network" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/9980d331fd11893cc30889890efe7447dd775a86.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": "Adaptive Source Localization on Complex Networks via Conditional Diffusion 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": 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": "•\tAlthough the proof is straightforward, it would be useful for the reader and for completeness to explain how Eq. (3) is derived in proof.\n•\tFor clarify, the role/design rationale behind each term in Eq. (5) can be explained briefly, although it is an existing method.\n•\tIn line 364, it is mentioned that the MSE reconstruction loss is used. How is this implemented together with the GLasso loss in Eq. (5)?\n•\tHow does the proposed method perform in terms of training time/cost?\n•\tWhat do different colours mean in Figure 1?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "•\tTo my knowledge, the proposed method is original. I like the proposed SPODNET, building on classical theory, simple but elegant.\n•\tThe paper is well written and well structured.\n•\tExperiment design is appropriate for demonstrating the effectiveness of the proposed method. I particularly like the result of UBG in Figure 6, highlighting more distinct structure that is interesting for this real-world data." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a new method SPODNET for learning SPD matrices by elements, supported by the classical Shur’s condition, where the matrix elements (u and v) to update are learned using neural networks. The work demonstrates the use of SPODNET for the sparse precision matrix learning task, and proposes three new model architectures to perform the learning, including UBG, PNP and E2E. Two sets of experiments were conducted for evaluation, using a synthetic data and a real-world data. The results show the effectiveness of the proposed methods, and their advantages over the compared ones." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "•\tThe paper would be stronger if they could include another real-world learning problem over SPD manifold. But I don’t see this as a major issue.\n•\tLack of discussion on the limitation of the proposed work.\n•\tThere are a couple of things that could be explained better, see my questions.\n•\tFigures 4 and 5 are too small, hard to read." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. Training details: What are the exact steps of the unrolling algorithm? How many unrollings are needed? From line 363, does the training posit an MSE loss on the intermediate $\\Theta$s or only the last one?\n2. GLasso: In Fig. 5, the F-1 performance of GLasso decreases when $n$ gets larger. This is weird because I expect it to recover the graph perfectly when $n\\gg p$. What are the authors' explanations about this?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is well-written and easy to follow. The idea of unrolling the column-row BCD algorithm to ensure SPD seems novel." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a deep learning-based approach to solving SPD problems such as covariance selection. The authors start with the block coordinate descent (BCD) algorithm and then unroll the optimization using neural networks. This follows with three different unrollings, where each preserves different levels of the problem structures (or inductive bias in the deep learning words). The authors evaluate the proposed SpodNet on synthetic data and the animal dataset against GLAD, GLasso, and other traditional approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I think this is a borderline paper in its current form. I value the novelty of the paper, but its numerical performance is not the most convincing. I think solving the following points will make the paper stand more firmly at my rating.\n\n1. GLAD-Z: SpodNet's NMSE performance on synthetic data seems to be consistently worse than GLAD-Z's. I understand the authors' argument that GLAD-Z is not SPD, but what if Z is projected onto the SPD cone? Will the projected Zs remain the lower NMSE scores? Because GLAD uses an ADMM-like algorithm, the learned $\\Theta$ matrices are not projections if I understand correctly.\n2. Large sample regime: The NMSE performance of SpodNet is no better or only marginally better than the baselines when $n>p$.\n3. Real-world datasets: The results of GLAD on the animal dataset are missing. Also, the paper will benefit from adding at least another real-world dataset.\n4. Figures: Some figures can be hard to read, especially Fig. 4-5. I suggest the authors use thinner lines and/or redesign their layout to make line plots larger." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See above" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. SpodNet enables increased expressivity for estimating SPD matrices with a neural network based parametrization, at the same time maintaining constraints like sparsity unlike other model based approaches, as verified by experiments on synthetic data.\n2. In order to make the method tractable the authors leverage the block structure of SPD matrices to restrict the complexity of each update step to $O(p^2)$.\n3. Experiments on synthetic data and graph topology estimation using SpodNet are presented, highlighting the effectiveness of the method. On the synthetic data, the proposed method is similar to model based approaches while achieving both SPD and sparsity at the same time." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a method for jointly learning Symmetric Positive Definite (SPD) and sparse matrices using a learning based approach. The proposed method SpodNet is an iterative method which learns column-row pairs stepwise where each step is parametrized by a neural network. The SPD constraint is then ensured via Schur’s condition." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The proposed SpodNet provides a novel approach to leverage neural networks and increase expressivity on constrained manifolds. However, the design of the neural networks itself is not discussed in detail. It is not clear to me why the authors choose the input features for $g$ as $\\theta_{22}$, $s_{22}$ and $\\theta_{12} \\theta_{11}^{-1} \\theta_{12}$. Similar for each of the three described approaches the explanation for choosing the input features is missing. I assume the features are chosen to best suit the model based approach and that performs well with gradient descent but the paper would benefit from a detailed explanation of the same.\n2. The method still seems relatively expensive in spite of the improved update rule. The overall cost as the authors mentioned is of the order of $O(p^3)$, how does this compare with the other model based approaches?\n3. Can the SpodNet framework maintain other structure constraints for example structural sparsity. In general what conditions would the constraints need to satisfy in order to be optimized with a SpodNet layer.\n\nSince the general literature of SPD matrix estimation points towards applications in computer vision, it would be informative to see an experiment for a vision task with SpodNet to verify the comparison with baselines and its scalability given its computational requirements." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a novel and generic learning module with guaranteed SPD outputs that can jointly handle additional structural constraints such as sparsity." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024schurs,\ntitle={Schur's Positive-Definite Network: Deep Learning in the {SPD} cone with structure},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v1B4aet9ct},\nnote={under review}\n}" }, "abstract": { "value": "Estimating matrices in the symmetric positive-definite (SPD) cone is of interest for many applications ranging from computer vision to graph learning. While there exist various convex optimization-based estimators, they remain limited in expressivity due to their model-based approach. The success of deep learning motivates the use of learning-based approaches to estimate SPD matrices with neural networks in a data-driven fashion. However, designing effective neural architectures for SPD learning is challenging, particularly when the task requires\nadditional structural constraints, such as element-wise sparsity. Current approaches either do not ensure that the output meets all desired properties or lack expressivity. In this paper, we introduce SpodNet, a novel and generic learning module that guarantees SPD outputs and supports additional structural constraints. Notably, it solves the challenging task of learning jointly SPD and\nsparse matrices. Our experiments illustrate the versatility and relevance of SpodNet layers for such applications." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "sparsity", "graphical lasso", "lasso", "deep learning", "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/2dbc07f8f1e6619f2154206d2ab38196f6213603.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/197819d45d9049396b52ffbccdf0b5897c6ac90e.zip" }, "title": { "value": "Schur's Positive-Definite Network: Deep Learning in the SPD cone with structure" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. “Aligning Motion Generation with Human Perceptions” by Wang et al., 2024, utilizes their proposed motion perception model as an effective supervision signal for training a motion generator. This paper similarly uses the motion perception model from Wang et al. as a supervision signal to aid motion generation. What are the distinction and novelty of this paper's method compared to the approach of Wang et al.?\n2. The use of **skeleton** rendering in the user study section for evaluation could be more reasonable, as there is a gap between **skeleton** and **SMPL** rendering. Some unnatural details are more noticeable in SMPL, using SMPL rendering for the user study might be more convincible." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper introduces human preference into the text-to-motion generation task by applying a motion perception model from the motion generation field. Additionally, two other rewards (text adherence and motion quality) are used to prevent potential degradation caused by the human preference reward. Experimental results demonstrate the superiority of the proposed approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes incorporating human preference priors into the text-to-motion task to enhance the quality of generated motions. Through the use of reinforcement learning and a motion perception model, the paper construct a human preference reward, enabling the generation model to learn human perception. To prevent degradation in other performance metrics resulted by human preference reward, the paper introduces a motion quality reward and a text adherence reward, forming a proposed multi-reward system. To mitigate potential training instability caused by multiple rewards, Pareto optimality is employed to balance the different rewards. The experimental results demonstrates superior performance in FID, R-Precision, human perceptual model scores, and user studies." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The most concerning limitation of this paper lies in its novelty. Introducing human perception into the field of motion generation is not novel, as also mentioned by this paper in line#86-87 (Voas et al., 2023; Wang et al., 2024). Wang et al. utilizes their proposed motion perception model as an effective supervision signal to finetune the motion generator. This paper similarly uses the motion perception model from Wang et al. as a supervision signal to train model for motion generation. Given these, the technical novelty is limited.\n\nThere are also insufficiencies in the experimental section: **1)** The main insufficiency is that the ablation study does not adequately demonstrate the effectiveness of the proposed multiple rewards. Table 2 lacks a baseline ablation result with no rewards applied, making it difficult to confirm the method’s effectiveness. **2)** In text-to-motion generation tasks, methods are typically validated on both the HumanML3D and KIT-ML datasets, but this paper only provides results for HumanML3D. **3)** Additionally, on the HumanML3D dataset, the MModality metric is usually reported; however, this paper neither provides MModality results nor explains why it was omitted. **4)** For the Diversity metric, Diversity closer to real test data is preferable, yet this paper labels higher values as better.\n\nMinor comments:\n1. The paper does not explain the metrics reported in the tables, such as Diversity.\n2. The “Conclusion and Future Work” section does not include future work but instead links to the appendix.\n3. Descriptions for figures should not be placed in the appendix; for example, <mm> and <mt> in Fig. 5 lack explanation in the main text." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "How does MotionRL handle the trade-offs between text adherence, motion quality, and human preferences during the training process?\n\nWhat are the limitations of relying on pre-trained human perception models for aligning generated motions with human preferences, and how can these limitations be addressed in future work?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Human-Centric Optimization: MotionRL uniquely incorporates human preferences into the optimization process, ensuring that the generated motions align better with human perception. This focus on human feedback addresses the limitations of traditional metrics that may not fully capture the nuances of human motion quality.\nMulti-Objective Optimization: The use of a multi-reward reinforcement learning framework allows MotionRL to balance multiple objectives simultaneously, such as text adherence, motion quality, and human preferences. This approach ensures that the generated motions are not only accurate and high-quality but also meet the subjective preferences of users.\nPareto Optimality: MotionRL introduces a novel multi-objective optimization strategy to approximate Pareto optimality. By selecting non-dominated points within each batch, the model learns to balance different rewards effectively, leading to more stable and optimal training outcomes. This method enhances the overall performance across various metrics compared to other algorithms." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces a novel approach that leverages Multi-Reward Reinforcement Learning to optimize text-to-motion generation tasks. Unlike previous methods that primarily focus on numerical performance metrics, MotionRL incorporates human preferences to enhance the alignment of generated motions with human perception. The approach uses a multi-objective optimization strategy to balance text adherence, motion quality, and human preferences, aiming for Pareto optimality. Extensive experiments and user studies demonstrate that MotionRL significantly outperforms existing methods in generating high-quality, realistic motions that align well with textual descriptions and human feedback. This work represents a significant advancement in the field of text-driven human motion generation, offering a more nuanced and human-centric approach to evaluating and improving motion quality." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Dependence on Pre-trained Perception Models: MotionRL relies on pre-trained human perception models to capture complex human perceptual information. If these models are not of high quality or do not accurately reflect human preferences, the performance of MotionRL could be significantly impacted. This dependency limits the flexibility and robustness of the approach.\n\nLimited Dataset Size: The text-to-motion domain typically has smaller datasets compared to other fields like image or text generation. This limitation makes the model more sensitive to small changes in input text and can lead to overfitting. The smaller dataset size also poses challenges in effectively training the model to generalize well across diverse motion scenarios.\n\nComplexity of Multi-Objective Optimization: While the multi-reward optimization strategy aims to balance text adherence, motion quality, and human preferences, it introduces significant complexity into the training process. Managing and fine-tuning multiple rewards can lead to unstable training and requires careful calibration to ensure that the model does not prioritize one objective at the expense of others. This complexity can make the approach less accessible and harder to implement effectively." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Could the authors elaborate on the potential for dynamically adjusting the weighting of rewards based on real-time feedback?\n\nHow does the model handle scenarios where text descriptions are ambiguous or open to interpretation? Is there a mechanism to weigh certain rewards more heavily in such cases?\n\nFor practical applications, are there plans to develop more user-friendly interfaces for non-technical users to fine-tune MotionRL’s generated motions?\n\nPlease also provide some details about the dataset you created" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper introduces a novel approach by combining reinforcement learning with human preference alignment for text-to-motion tasks, which is an underexplored area in motion generation. The application of multi-reward RL to text-to-motion represents a good contribution, especially as it integrates human perceptual feedback." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper \"MotionRL: Align Text-to-Motion Generation to Human Preferences with Multi-Reward Reinforcement Learning\" introduces MotionRL, a framework that leverages Multi-Reward Reinforcement Learning to enhance text-to-motion generation by aligning outputs with human preferences. Unlike prior models, MotionRL incorporates human perceptual data in its reward structure to address limitations in previous approaches focused on numerical metrics alone. MotionRL employs a multi-objective optimization strategy to achieve a balance between text adherence, motion quality, and human preferences, approximating Pareto optimality for optimal trade-offs. The authors demonstrate the model's effectiveness through extensive experiments and user studies, showing MotionRL's superior performance in terms of perceptual quality and alignment with human expectations." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The model’s reliance on pre-trained perception models, as mentioned in the limitations, could restrict generalizability. Fine-tuning without additional human annotations might limit the model’s adaptability to new or unique datasets, where the pre-trained perception model may not fully capture the nuances of human preferences.\n\nWhile the multi-reward RL framework is effective, there is limited discussion on dynamically adjusting the weight of each reward in real-time based on specific user feedback. A more adaptive reward weighting mechanism could further enhance user-centered customization.\n\nAlthough the human preference model provides valuable perceptual data, additional insights from continuous human interaction could help refine the model further. The paper would benefit from exploring how human feedback could be iteratively incorporated to improve long-term model performance, potentially through ongoing human-in-the-loop adjustments." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 did you come up with the balance between the three different rewards?\n\n* How is the pareto-based policy gradient optimization different from simple PPO?\n\n* What is the demographic data for the user study, did you do any formal survey? \n\nMinor things:\n- You introduce the abbreviation \"RL\" multiple times in the text\n- Line 200: the Appdix -> the Appendix" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* Introducing RL with the set of reward functions to optimize for multiple objectivesin text-to-motion generation seems to be a novel approach to that specific domain.\n\n* The idea is simple and easy to understand which is also further improved by a good presentation and writing. The images helps me understand what is the problem you're trying to solve.\n\n* I find the ablation in table 2 important as it shows that human perception does not always align with the metricses typically used." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper offers MotionRL, an approach for optimizing text-to-motion generation tasks using Multi-Reward RL. This method aims to align the generated motions with human preferences more effectively than traditional models by incorporating human perceptual models into the RL training process. This is done by employing a multi-objective optimization strategy to balance text adherence, motion quality, and human preferences, approximating Pareto optimality. The paper provides experiments, demonstrating that the proposed model outperforms baselines on traditional metrics and in user studies assessing human perception." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The overall novelty of the method is modest at best, but the domain application seem to be novel.\n\n* There is a great lack of details about the user study and \"volunteers for evaluation\" could mean anyone. Given that the participants was called \"volunteers\" (and not paid?) it seems like it's not randomly recruited people which makes it hard to evaluate the soundness of the user study. \n\n* The paper uses up all 10 pages but given that it's a simple idea with limited technical contribution I think it's excessive, especially since the limitation was pushed to the appendix." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. In line 95, it seems that the purpose of introducing RL is merely to supervise human preference, and the rationale for using RL isn’t well explained.\n\n2. The concept of Pareto optimality lacks a citation—what specific issue does this?\n\n3. There are factual inaccuracies in the stated contributions, as ReinDiffuse appears to have introduced RL into motion before this paper.\n\n4. The supplementary materials don’t clearly explain their purpose, which only provides some generated samples. \n\n5. The paper contains several grammatical errors. The title should change \"Align\" to \"Aligning\". A layout issue in line 292." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The experimental metrics show good results, and a motion generation framework based on RL has been designed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces MotionRL, a reinforcement learning (RL)-based framework for multi-objective optimization in human motion generation. Recognizing that prior methods often overlook human perceptual nuances, MotionRL integrates human perceptual priors into the generation process through a reward structure. The framework balances three objectives: human perception, motion quality, and text adherence. To address the challenges of optimizing multiple objectives, MotionRL incorporates a multi-reward optimization strategy that approximates Pareto optimality by selecting non-dominated points in each batch, ensuring better trade-offs across objectives." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The reason for introducing RL is not well explained." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024motionrl,\ntitle={Motion{RL}: Align Text-to-Motion Generation to Human Preferences with Multi-Reward Reinforcement Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v1OQ0kNq0w},\nnote={under review}\n}" }, "abstract": { "value": "We introduce \\textbf{MotionRL}, the first approach to utilize Multi-Reward Reinforcement Learning (RL) for optimizing text-to-motion generation tasks and aligning them with human preferences. Previous works focused on improving numerical performance metrics on the given datasets, often neglecting the variability and subjectivity of human feedback. In contrast, our novel approach uses reinforcement learning to fine-tune the motion generator based on human preferences prior knowledge of the human perception model, allowing it to generate motions that better align human preferences. In addition, MotionRL introduces a novel multi-objective optimization strategy to approximate Pareto optimality between text adherence, motion quality, and human preferences. Extensive experiments and user studies demonstrate that MotionRL not only allows control over the generated results across different objectives but also significantly enhances performance across these metrics compared to other algorithms." }, "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": [ "Motion Generation; 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/aabb496632c83d612ab54e5bebaa9a510cb0f077.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/1deeb9f4eb3998be4f76115a8debdb13a56484d8.zip" }, "title": { "value": "MotionRL: Align Text-to-Motion Generation to Human Preferences with Multi-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": 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. Add real-world experiments and more benchmark datasets (with larger scene scale).\n2. Add visualizations to the decomposed appearance throughout the figures in the paper.\n3. Add missing details in the method section.\n4. Add some comparison on the train - test efficiency and memory footprint.\n4. Revised the discussions and ablation studies as suggested in the weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well written and easy to follow. The idea is clean and the pipeline do not introduce additional hyper-parameter tuning and selection compared to some other recent methods for neural surface reconstruction / rendering.\n2. The idea to fuse multi-resolution grids for detailed surface reconstruction is novel. AniSDF’s fused-granularity structure balances high- and low-resolution information to improve convergence and accuracy. This approach allows for a more adaptive reconstruction that captures both overall structure and fine details, which is validated by their good geometry quality (chamfer) in the experiments.\n3. The use of ASG encoding in appearance modeling seems to be effective and handles specular reflections very well." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a high-quality 3D surface reconstruction method- AniSDF, which learns fused-granularity neural surfaces with physics-based encoding. The authors propose fused multi-resolution grids for geometry modeling, and adopt Anisotropic Gaussians for appearance modeling. With these designs, AniSDF can reconstruct objects with complex structures and produce high-quality renderings on benchmarked datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Experiments for reflective surfaces mainly come from synthetic data, it would be helpful to understand the model’s ability if we could see results of more real-world reflective surface data, such as trucks from Tanks and Templates, sedans from Refnerf, The Glossy-Real dataset from Nero.\n2. The appearance modeling involves blending view-based and reflection-based radiance fields, however, the method's ability to decompose base color and reflection color is unknown. It would be better if the author could add a visualization of view-based color, reflection-based color, and the blending weight.\n3. Some details of the methods are missing. The derivation of normal and the method for mesh extraction are not discussed.\n4. The reason behind the choice of grid level `m` and `l` is not clear. It would be clearer if an ablation study about grid resolution were added." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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. For Eq. 10, why would both c_view and c_ref depend on view directions? Would it encourage better diffuse-specular separation if one makes c_view view-independent?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper is well-written and easy to follow. \n2. Qualitative and quantitative results of the proposed method seems very strong, beating prior baselines like NeuS, RefNeRF, RefNeuS. Reconstructed meshes look clean with high-quality surface normals. \n3. The authors validate the proposed method on both synthetic datasets (Nerf-synthetic and Shiny-blender), and real datasets (DTU), and it's nice to see improvements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper provides a fused-granularity neural surfaces with physics-based anisotropic\nspherical Gaussian encoding for high-fidelity 3D reconstruction. The authors show state-of-the-art novel-view rendering and geometry reconstruction results on several datasets, including NeRF-Synthetic, Shiny Blender, and DTU datasets. The proposed method shows very convincing reconstruction of challenging specular and furry objects." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. To further convince me about the method's performance on reconstructing specularity, I would need to see the view synthesis videos, as opposed to the static frames shown in the paper and project page. Unfortunately, I could not find such videos (except the relighting ones). \n\n2. For the proposed blended radiance field (Eq. 12), I think it would be great to provide some visualizations of the individual components: w, c_view, c_ref, to better understand what each learnt components look like. \n\n3. I'm unsure why the fused-granularity hash grids actually works better than a plain multi-resolution hash grids. It seems to me that the major difference of it from a plain one is the additional handcrafted equation (6) that says the final SDF is an addition of coarse and fine SDF. In the plain multi-resolution hash grid, the final SDF is predicted by a MLP from concatenated multi-resolution features. This could benefit some justification." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "The reviewer would like to raise awareness of possible breach of double blind review rules.\n\nThe reviewer found a twitter page:\n\nhttps://x.com/zhenjun_zhao/status/1842119223646302292\n\nThis page introduces their paper with authors names explicitely shown." }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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. Modify the typo in Line 093 (‘Oue’) to ‘Our.’\n\n2. The physical-based rendering method via ASG is similar to $\\cite{yang2024spec}$. Is this work also inspired by similar works using ASG to learn the specular representation in 3DGS?\n\n3. The blended radiance fields with ASG encoding are composed of $c_{view}$ and $c_{ref}$ though a learnable weight. According to Eq. 4, the light field is modeled by $c_d$ and $c_s$, diffuse color and specular color. So, can $c_{view}$ be regarded as purely diffuse and $c_{ref}$ as the pure composition of specular? If so, can the radiance fields be considered purely diffuse when $\\omega$ is 1? If not, can this work disentangle the light field to only diffuse or specular field?\n\n4. Refer to $\\cite{han2023multiscale}$, they control the final color by adding the scale to the color calculated from ASG when retaining the diffuse color term calculated from the first three orders of SH. So, what motivates adding weight to diffuse and specular in this work? What are the differences between your light field calculation in the blended radiance fields with $\\cite{han2023multiscale}$?\n\n@article{han2023multiscale,\n title = {Multiscale Tensor Decomposition and Rendering Equation Encoding for View Synthesis},\n author = {Kang Han and Weikang Xiang},\n journal = {Computer Vision and Pattern Recognition},\n year = {2023},\n doi = {10.1109/CVPR52729.2023.00412},\n bibSource = {Semantic Scholar https://www.semanticscholar.org/paper/aa41843888fffada6335b6c5cdbcd2d4bb5cf9da}\n}\n\n@article{yang2024spec,\n title={Spec-gaussian: Anisotropic view-dependent appearance for 3d gaussian splatting},\n author={Yang, Ziyi and Gao, Xinyu and Sun, Yangtian and Huang, Yihua and Lyu, Xiaoyang and Zhou, Wen and Jiao, Shaohui and Qi, Xiaojuan and Jin, Xiaogang},\n journal={arXiv preprint arXiv:2402.15870},\n year={2024}\n}" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Originality:\nThe paper explores the potential of parallel using coarse and fine hash-grid to replace the general sequential coarse-to-fine structure, demonstrating the effects of experiments. Besides, this paper combines SDF learning with blended radiance field learning with anisotropic spherical Gaussian encoding to distinguish material information.\n\nQuality:\nThe quality of the paper is good, evidenced by detailed experiments and comprehensive comparisons with state-of-the-art methods. \n\nClarity:\nThe paper is well-structured and organized. \n\nSignificance:\nThe good geometry that disambiguates the reflective appearance is helpful in 3D reconstruction. The possible relighting application makes this research meaningful." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces AniSDF, an approach to enhance the quality of SDF-based methods in geometry reconstruction and novel-view synthesis tasks, enabling the physically-based rendering ability. Firstly, AniSDF uses the parallel branch structure of coarse hash-grids and fine hash-grids, replacing the former sequential coarse-to-fine training strategy, to learn a fused-granularity neural surface to improve the quality of SDF. Secondly, AniSDF uses Anisotropic Spherical Gaussian Encoding to learn blended radiance fields with a physics-based rendering, disambiguating the reflective appearance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Reference is insufficient: From Lines 130 to 135, the Sec. 2.1 is related to the attempts to improve the reconstructed geometry of Gaussians. Since 2DGS is used to compare, it is no reason to cite some papers focused on doing similar jobs: improving the surface reconstruction of 3DGS, like $\\cite{guedon2023sugar, lyu20243dgsr, chen2023neusg}$.\n\n2. The ablation study of the fused-granularity neural surface is not enough. The ablation study shows the comparison with the sequential coarse-to-fine method, but the technique with only coarse hash-grid and only fine hash-grid should also be demonstrated to prove the observations shown at the beginning of Sec.3.2. It could be better if the training time comparison is also shown in this part.\n\n@article{guedon2023sugar,\n title={SuGaR: Surface-Aligned Gaussian Splatting for Efficient 3D Mesh Reconstruction and High-Quality Mesh Rendering},\n author={Gu{\\'e}don, Antoine and Lepetit, Vincent},\n journal={CVPR},\n year={2024}\n}\n@article{lyu20243dgsr,\n title = {3DGSR: Implicit Surface Reconstruction with 3D Gaussian Splatting},\n author = {Xiaoyang Lyu and Yang-Tian Sun and Yi-Hua Huang and Xiuzhe Wu and Ziyi Yang and Yilun Chen and Jiangmiao Pang and Xiaojuan Qi},\n year = {2024},\n journal = {arXiv preprint arXiv: 2404.00409}\n}\n@article{chen2023neusg,\n title = {NeuSG: Neural Implicit Surface Reconstruction with 3D Gaussian Splatting Guidance},\n author = {Hanlin Chen and Chen Li and Gim Hee Lee},\n year = {2023},\n journal = {arXiv preprint arXiv: 2312.00846}\n}" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Questions:\n1. I compared the chamfer distance metric of the Neuralangelo method reproduced on the DTU dataset in the paper, and there is a significant gap. The original paper reported an average of 0.61 (which surpasses your method), while the reproduced result in the paper is 1.07. Could the authors clarify the reasons for this discrepancy? Specifically, did you maintain the same hyperparameter settings as Neuralangelo? Please provide detailed information on your experimental setup.\n2. Section 3.2 of the paper lists some issues related to coarse grid and fine grid training, but there are no corresponding experimental supports for these claims. Regarding the use of the coarse to fine method, you pointed out that thin structures may be discarded in the early training stages. Could you provide visualizations of surface reconstruction at different training stages, along with corresponding quantitative metrics, particularly for Neuralangelo and Neus2? This would help us assess the advantages and disadvantages in the reconstruction of detailed structures. I noticed your experiments in the ablation study, but they do not specify the experimental settings and only show the final results." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tInnovative Approach: The use of fused-granularity neural surfaces combined with ASG encoding for 3D reconstruction is novel and effective in balancing both coarse and fine details, leading to improved geometry and appearance quality.\n2.\tHigh Performance: Experimental results show significant improvements in rendering quality and geometry reconstruction, with better handling of reflective, luminous, and fuzzy objects compared to existing methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents AniSDF, a novel SDF-based method for high-fidelity 3D reconstruction that incorporates fused-granularity neural surfaces and anisotropic spherical Gaussian (ASG) encoding. AniSDF aims to achieve accurate geometry reconstruction and photo-realistic rendering by addressing challenges in neural radiance fields, such as geometry-appearance trade-offs. The approach uses parallel fused-granularity neural surfaces to balance coarse and fine details, and blended radiance fields with ASG encoding for modeling both diffuse and specular appearances. Extensive experiments demonstrate AniSDF's superiority in both geometry reconstruction and novel-view synthesis over prior methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tLimited Real-Time Capability: AniSDF cannot perform real-time rendering, which limits its applicability in time-sensitive applications such as interactive graphics or augmented reality.\n2.\tComputation Cost: The use of multiple neural networks and high-resolution hash grids could be computationally expensive, which may hinder scalability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "It’s unclear whether the larger network or the fused-granularity neural surface structure is responsible. What would happen if we set both the fine and coarse grids to the same resolution, either that of the coarse grid or that of the fine grid?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Pros:\nThe paper is well-written and generally easy to follow.\nExperimental results demonstrate incremental improvements in PSNR, which support the proposed approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The AniSDF paper introduces an innovative approach to high-fidelity surface reconstruction and photo-realistic rendering from multi-view images. This is achieved through a synergistic geometry network and appearance network, which together enable high-quality 3D reconstruction. Additionally, the authors propose a fused-granularity neural surface that aims to balance overall structural integrity with fine detail preservation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Cons:\nThe fused-granularity neural surface structure may lack novelty, as it essentially uses two parallel structures with different resolutions. It seems likely that resolution choices could impact the final reconstruction quality. Including experiments that vary resolution settings would clarify their effect. \n\n\nDespite claims of high-quality mesh reconstruction, Chamfer Distance results reveal performance gaps on certain objects (e.g., \"Chair\" and \"Mic\" categories) compared to methods like Neus and NeRO. Explaining these discrepancies would help elucidate the limitations." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024anisdf,\ntitle={Ani{SDF}: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v1f6c7wVBm},\nnote={under review}\n}" }, "abstract": { "value": "Neural radiance fields have recently revolutionized novel-view synthesis and achieved high-fidelity renderings. \nHowever, these methods sacrifice the geometry for the rendering quality, limiting their further applications including relighting and deformation. \nHow to synthesize photo-realistic rendering while reconstructing accurate geometry remains an unsolved problem. In this work, we present AniSDF, a novel approach that learns fused-granularity neural surfaces with physics-based encoding for high-fidelity 3D reconstruction. Different from previous neural surfaces, our fused-granularity geometry structure balances the overall structures and fine geometric details, producing accurate geometry reconstruction. \nTo disambiguate geometry from reflective appearance, we introduce blended radiance fields to model diffuse and specularity following the anisotropic spherical Gaussian encoding, a physics-based rendering pipeline. With these designs, AniSDF can reconstruct objects with complex structures and produce high-quality renderings. \nFurthermore, our method is a unified model that does not require complex hyperparameter tuning for specific objects. \nExtensive experiments demonstrate that our method boosts the quality of SDF-based methods by a great scale in both geometry reconstruction and novel-view synthesis." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Surface Reconstruction", "Neural Radiance Field" ] }, "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/612a8b13a1a4f68011420360d14bb62dac043d0c.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/abeb06a1d21aeb62e7b7d8264da9df458206efe7.zip" }, "title": { "value": "AniSDF: Fused-Granularity Neural Surfaces with Anisotropic Encoding for High-Fidelity 3D Reconstruction" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Several parts of the CurveLID metric requires fitting on training data of adversarial v.s. benign prompts. I’d expect that the performance will depend heavily on the coverage of training data distribution, but judging from the experimental results it seems the generalization is pretty descent, have the authors had any insights on why this is the case?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Novelty: The author revitalized LID score in detecting adversarial prompts by estimating it on the learned aggregated output of word level embeddings. It further proposes TextCurve to capture the extra curvature information around the prompt. The final decision takes account for both metrics via a learned linear combination.\n2. Extensive experimental evaluations. The paper includes a detailed amount of empirical evaluations and analysis to support the effectiveness of the proposed method, as well as providing insights into how different components affect its performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies adversarial prompt detection. It revitalizes metrics of LID from anomaly detection to this task, by estimating the score on a learned prompt-level representation instead of a suboptimal word-level one. The paper further proposes TextCurve score, and merges these two metrics into a single score for the final decision. The final binary classifier needs to be trained on a collect. Extensive empirical results and ablation study demonstrate the generation and robustness of the proposed metrics despite the need for training." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. How is curveLID’s defenses against jailbreaking algorithms rather than static harmful prompt? It Whitebox methods are not applicable in this setting, but the author can potentially try some blackbox methods, such as DrAttack or Multijail? It seems that the authors include several defense methods in the Appendix, but the experiment seems not directly related to the proposed metrics (I could be wrong).\n2. Minor: the appendix is very rich in useful information so it cannot be ignored. Tut currently the organization is quite flattened so a bit difficult to track everything. Organizing it into sections/subsections (e.g. all ablations are grouped together, all analysis are grouped together) might improve its 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": 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": "I recommend the following:\n1. A complete theoretical overhaul establishing proper connections between differential geometry and their implementation\n2. Justification for architectural choices and comparison with alternatives\n3. Analysis of and solutions for the sequence dependency problem\n4. Comprehensive evaluation across a wider range of linguistic variations and attack types\n5. Complete technical details enabling reproducibility\n\nWhile the empirical results appear promising, they cannot overcome the fundamental theoretical and methodological issues present in the paper." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "1. The paper introduces an innovative geometric framework (CurvaLID) that combines two complementary measures - Local Intrinsic Dimensionality (LID) and curvature - to detect adversarial prompts. Unlike existing approaches that rely on token-level analysis, the paper presents PromptLID which analyzes geometric differences at the prompt level, and TextCurv which captures semantic shifts at the word level.\n\n2. The paper achieves exceptional performance metrics like (i) Over 0.99 detection accuracy for English prompts, \n(ii) 0.994 accuracy for non-English prompts and (iii) Successfully reduces attack success rates to near zero\nIt also Includes detailed ablation studies showing the importance of both PromptLID and TextCurv components.\n\n3. The solution is highly efficient, requiring only 15 minutes of training on a single Nvidia H100 GPU, compared to competing methods that need up to 16 hours on more extensive hardware setups. The approach is model-agnostic, meaning it doesn't require access to or modification of the underlying LLM. The method also scales well to different languages without requiring retraining." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Summary:\nThis paper presents CurvaLID, a framework that claims to detect adversarial prompts in LLMs using geometric measures combining Local Intrinsic Dimensionality (LID) and curvature. While achieving reported empirical results of 99% detection accuracy, the paper suffers from fundamental theoretical flaws, questionable design choices, and significant limitations that are either unexplored or unacknowledged.\n\nMajor issues identified:\n\n1. The paper presents three classical curvature definitions (osculating circle, Whewell equation, differential geometric) but implements a formula (dθ/(1/||u|| + 1/||v||)) that has no clear mathematical connection to any of them. The reference to Whewell's equation (dφ/ds) lacks proper theoretical grounding. The authors don't define arc length, don't justify how embedding angles relate to tangential angles, and provide no proof of geometric equivalence to Whewell’s equation. The denominator term using inverse vector norms appears ad hoc and unfortunately comes with no geometric justification.Claims about semantic shifts relating to curvature are made without mathematical rigor or proof. The relationship between their geometric measures and adversarial behavior is assumed (lines 265-266) rather than proven. Therefore, there is a massive disconnect between the theory shown in Section 3 and what is finally implemented. \n\n2. The choice of architecture to use a CNN for text classification is unjustified by the authors. Why choose a CNN when there are more appropriate architectures (like Transformers, RNNs etc) available for sequential data?\nThe missing aspects are that: (i) there is no justification as to why CNNs would better capture textual properties. \n(ii) how are variable length sequences handled by a CNN? Is it via padding? \n(iii) How would a CNN capture long-range dependencies in text?\n(iv) CNNs typically rely on “spatial order”, what is the equivalent in this setting?\n\n3. In the multi-class classification definition, there is no proper specification given of the label space Q and its relationship to their task. It looks like it’s just a single label per training example. \n\n4. This paper provides no theoretical time and storage complexity analysis of their defense, especially for their key algorithms. While TextCurv computation is O(nd) for sequence length n and embedding dimension d, they don't discuss this in detail. The fact is that if “true differential geometry” metrics were used then the computational complexity is high and mostly polynomial time. \n\nThe complexity analysis for critical components should be presented. Such as CNN feature extraction, k-NN computation for PormptLID, Overall pipeline complexity including the MLP classifier. How does complexity scale with longer sequences or larger embedding dimensions or increased batch sizes?\n\nThis lack of complexity analysis makes it difficult to assess: i). Practical deployability of their method, ii). scalability to longer prompts, iii). real-time defense capabilities and iv). resource requirements for implementation\n\n\n5. The method relies critically on exact word order, making it vulnerable to simple paraphrasing. If there are basic linguistic variations like active/passive voice conversion, which retains the semantic meaning, this method would end up with two different representations (geometric measures) completely. This limitation isn’t mentioned or addressed. A simple text restructuring can circumvent this defense. \n\n6. This work presents a limited evaluation in terms of variety of attack types, even when focused on just single-step attacks. There is no analysis of the dataset diversity or potential overlaps between the 8 datasets proposed. There are no evaluations on social engineering attacks (as claimed in introduction) where they stop persuasion attacks for example. There is no testing of non-English attacks despite claims of language-agnostic performance. \n\n7. The CNN feature extraction process is not explained properly and leaves a lot of questions. The hyperparameter choices are arbitrary and do not come with a good justification. It would be nice to see training stability, convergence and sensitivity to initialization studies in the experiments. There must be comparisons made to move naive baseline approaches. \n\n8. There are some reproducibility concerns that must be highlighted too. Critical hyperparameters are left unspecified. Preprocessing details are left incomplete. There is no outline of the training environment and no discussion of potential implementation challenges that are faced.\n\n\nThe paper suffers from multiple fundamental flaws that significantly undermine its contribution. Currently, the paper presents a major disconnect between the presented theoretical framework (which is sophisticated and borrowed from differential geometry) and actual implementation (which is a heuristic). The sequence dependency problem reveals a fundamental limitation that makes the method vulnerable to simple linguistic variations, while the choice of CNN architecture for text processing appears arbitrary and poorly justified.\n\nAdditionally, the lack of thorough analysis, missing implementation details, and limited validation make it difficult to assess the method's true effectiveness and reproducibility. The combination of theoretical flaws, practical limitations, and methodological gaps warrants a rejection at this point.\n \nI recommend the following:\n1. A complete theoretical overhaul establishing proper connections between differential geometry and their implementation\n2. Justification for architectural choices and comparison with alternatives\n3. Analysis of and solutions for the sequence dependency problem\n4. Comprehensive evaluation across a wider range of linguistic variations and attack types\n5. Complete technical details enabling reproducibility\n\nWhile the empirical results appear promising, they cannot overcome the fundamental theoretical and methodological issues present in the paper." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper presents three classical curvature definitions (osculating circle, Whewell equation, differential geometric) but implements a formula (dθ/(1/||u|| + 1/||v||)) that has no clear mathematical connection to any of them. The reference to Whewell's equation (dφ/ds) lacks proper theoretical grounding. The authors don't define arc length, don't justify how embedding angles relate to tangential angles, and provide no proof of geometric equivalence to Whewell’s equation. The denominator term using inverse vector norms appears ad hoc and unfortunately comes with no geometric justification.Claims about semantic shifts relating to curvature are made without mathematical rigor or proof. The relationship between their geometric measures and adversarial behavior is assumed (lines 265-266) rather than proven. Therefore, there is a massive disconnect between the theory shown in Section 3 and what is finally implemented. \n\n2. The choice of architecture to use a CNN for text classification is unjustified by the authors. Why choose a CNN when there are more appropriate architectures (like Transformers, RNNs etc) available for sequential data?\nThe missing aspects are that: (i) there is no justification as to why CNNs would better capture textual properties. \n(ii) how are variable length sequences handled by a CNN? Is it via padding? \n(iii) How would a CNN capture long-range dependencies in text?\n(iv) CNNs typically rely on “spatial order”, what is the equivalent in this setting?\n\n3. In the multi-class classification definition, there is no proper specification given of the label space Q and its relationship to their task. It looks like it’s just a single label per training example. \n\n4. This paper provides no theoretical time and storage complexity analysis of their defense, especially for their key algorithms. While TextCurv computation is O(nd) for sequence length n and embedding dimension d, they don't discuss this in detail. The fact is that if “true differential geometry” metrics were used then the computational complexity is high and mostly polynomial time. \n\nThe complexity analysis for critical components should be presented. Such as CNN feature extraction, k-NN computation for PormptLID, Overall pipeline complexity including the MLP classifier. How does complexity scale with longer sequences or larger embedding dimensions or increased batch sizes?\n\nThis lack of complexity analysis makes it difficult to assess: i). Practical deployability of their method, ii). scalability to longer prompts, iii). real-time defense capabilities and iv). resource requirements for implementation\n\n\n5. The method relies critically on exact word order, making it vulnerable to simple paraphrasing. If there are basic linguistic variations like active/passive voice conversion, which retains the semantic meaning, this method would end up with two different representations (geometric measures) completely. This limitation isn’t mentioned or addressed. A simple text restructuring can circumvent this defense. \n\n6. This work presents a limited evaluation in terms of variety of attack types, even when focused on just single-step attacks. There is no analysis of the dataset diversity or potential overlaps between the 8 datasets proposed. There are no evaluations on social engineering attacks (as claimed in introduction) where they stop persuasion attacks for example. There is no testing of non-English attacks despite claims of language-agnostic performance. \n\n7. The CNN feature extraction process is not explained properly and leaves a lot of questions. The hyperparameter choices are arbitrary and do not come with a good justification. It would be nice to see training stability, convergence and sensitivity to initialization studies in the experiments. There must be comparisons made to move naive baseline approaches. \n\n8. There are some reproducibility concerns that must be highlighted too. Critical hyperparameters are left unspecified. Preprocessing details are left incomplete. There is no outline of the training environment and no discussion of potential implementation challenges that are faced.\n\n\nThe paper suffers from multiple fundamental flaws that significantly undermine its contribution. Currently, the paper presents a major disconnect between the presented theoretical framework (which is sophisticated and borrowed from differential geometry) and actual implementation (which is a heuristic). The sequence dependency problem reveals a fundamental limitation that makes the method vulnerable to simple linguistic variations, while the choice of CNN architecture for text processing appears arbitrary and poorly justified.\n\nAdditionally, the lack of thorough analysis, missing implementation details, and limited validation make it difficult to assess the method's true effectiveness and reproducibility. The combination of theoretical flaws, practical limitations, and methodological gaps warrants a rejection at this point." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Could the authors address the three weaknesses above?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper is clearly written and the proposed method is described clearly." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an adversarial prompt detection method using the geometric information within the latent representation of an input prompt. Specifically, prompt-level local intrinsic dimensionality using a CNN is proposed as a feature to distinguish the difference between adversarial and benign prompts. TextCurv is also proposed to account for semantic change at the word level, which is another feature to classify adversarial prompts. The proposed method is validated across several adversarial attacks in different LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The size of the data (2500 prompts) is quite limited. \n\n2. Persuasive adversarial prompts [1] will also lead to harmful content but not considered and tested in this work. \n\n3. The proposed detection looks similar to perplexity-based filter. It would be more convincing to include attacks that can bypass perplexity-based filter like AutoDAN [2].\n\n[1] Yi Zeng, Hongpeng Lin, Jingwen Zhang, Diyi Yang, Ruoxi Jia, and Weiyan Shi. 2024. How Johnny Can Persuade LLMs to Jailbreak Them: Rethinking Persuasion to Challenge AI Safety by Humanizing LLMs. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics\n\n[2] Zhu, S., Zhang, R., An, B., Wu, G., Barrow, J., Wang, Z., Huang, F., Nenkova, A. and Sun, T., 2024. AutoDAN: interpretable gradient-based adversarial attacks on large language models. In First Conference on Language Modeling." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "1 Have the authors tried different CNN configurations to assess their impact on the performance of PromptLID and TextCurv?\n\n2. Why did the authors choose to use CNN? If the goal is to capture relationships between textual features, wouldn’t Transformers perform better? Have you considered trying Transformers for this purpose?\n\n3. The method relies only on structural and semantic changes to identify adversarial samples. Have the authors tested against adversarial examples with structural and semantic characteristics similar to benign samples? Specifically, have the authors tried generating adversarial examples with density and curvature that closely resemble benign samples to evaluate CurvaLID’s performance?\n\n4. Have the authors analyzed the impact of text length on PromptLID? Given that longer texts have higher complexity and density, could they lead to benign texts being misclassified as adversarial?\n\n5. Have the authors tested whether non-standard benign samples, such as those containing spelling errors or slang, might be misclassified as adversarial examples?\n\n6. Can the method detect adversarial prompts that rely on contextual association across multiple prompts?\n\n7. The authors trained the model on Orca, MMLU, AlpacaEval, and TQA datasets and then used the same datasets to evaluate whether the samples were benign or adversarial. Given this setup, does the model’s high accuracy have practical significance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The method has good advantages. Currently, most adversarial sample detection methods rely on large models, which makes the training process quite time-consuming. CurvaLID bypasses large language models entirely, directly analyzing text embeddings to identify adversarial inputs, and saving computational resources.\n\nAdditionally, the method is zero-shot, meaning it doesn't require fine-tuning for each LLM, making it even more versatile and efficient." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces CurvaLID, a model that detects adversarial prompts targeting LLMs. CurvaLID can capture the structural characteristics of entire prompts through \"Prompt-Level Local Intrinsic Dimensionality\" (PromptLID) and detects semantic shifts within the text using \"Text Curvature\" (TextCurv) to identify whether a prompt is adversarial. Since this method relies entirely on analyzing textual structure rather than LLM-specific architectures or parameters, it requires no fine-tuning for each LLM. This design significantly accelerates training speed, taking only 15 minutes on a single Nvidia H100 GPU." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1 I think that the innovation and contributions of your method are somewhat limited. The proposed approach resembles a CNN-based text filter, primarily detecting anomalies in embeddings, which isn't a novel concept.\n\n2 The connection between this work and LLMs is relatively weak. As mentioned, CurvaLID functions more like an independent text filter, focusing only on text structure and density features. It's more of a generic text filter than a targeted tool for detecting adversarial samples specific to LLMs. I don't mean to suggest that the detection method has no value, but most LLM-focused adversarial attacks today generate adversarial samples that closely resemble benign samples in terms of structure, density, and even embedding similarity. This would likely limit CurvaLID's effectiveness against subtle adversarial prompts.\n\n3 There are also issues in the experimental design. CurvaLID was trained and evaluated on the same datasets—Orca, MMLU, AlpacaEval, and TQA—which could inevitably lead to overly optimistic results.\n\n4 I noticed that the manuscript mixes American and British spellings in some places (such as \"defence\" and \"defense\"). To maintain a professional and consistent tone throughout, I recommend choosing one spelling style—American or British—and using it uniformly across the article.\n\n5 The formatting of tables is inconsistent. Some tables use a three-line format, while others do not. A unified table style would improve clarity and presentation quality." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "CurvaLID is a unified algorithm that detects adversarial prompts in LLMs using Local Intrinsic Dimensionality and curvature, achieving consistent and over 0.99 accuracy across various models." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024curvalid,\ntitle={{CURVALID}: A Geometrically-guided Adversarial Prompt Detection},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v1qNr99R5n},\nnote={under review}\n}" }, "abstract": { "value": "Adversarial prompts that can jailbreak large language models (LLMs) and lead to undesirable behaviours pose a significant challenge to the safe deployment of LLMs. Existing defenses, such as input perturbation and adversarial training, depend on activating LLMs' defense mechanisms or fine-tuning LLMs individually, resulting in inconsistent performance across different prompts and LLMs. To address this, we propose CurvaLID, an algorithm that classifies benign and adversarial prompts by leveraging two complementary geometric measures: Local Intrinsic Dimensionality (LID) and curvature. LID provides an analysis of geometric differences at the prompt level, while curvature captures the degree of curvature in the manifolds and the semantic shifts at the word level. Together, these tools capture both prompt-level and word-level geometric properties, enhancing adversarial prompt detection. We demonstrate the limitations of using token-level LID, as applied in previous work, for capturing the geometric properties of text prompts. To address this, we propose PromptLID to calculate LID in prompt-level representations to explore the adversarial local subspace for detection. Additionally, we propose TextCurv to further analyze the local geometric structure of prompt manifolds by calculating the curvature in text prompts. CurvaLID achieves over 0.99 detection accuracy, effectively reducing the attack success rate of advanced adversarial prompts to zero or nearly zero. Importantly, CurvaLID provides a unified detection framework across different adversarial prompts and LLMs, as it achieves consistent performance regardless of the specific LLM targeted." }, "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", "Adversarial attacks", "Local Intrinsic Dimension", "Curvature" ] }, "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/05a9568ddf21e7f58c2253821467b1bdd2bb24cb.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": "CURVALID: A Geometrically-guided Adversarial Prompt 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": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No 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 in weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Subgraph FL with inter-connections is an important topic.\n2. Completing the missing L-hop features by learning a L-hop node structure embedding is an interesting idea.\n3. Experiments show the performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper works on subgraph FL for node classification, where inter-connections between different clients is important. \n\nIt first computes a global L-hop neighborhood matrix before training. During training, it uses GNN for node feature embedding and use L-hop matrix multiplying a trainable matrix to calculate node structure embedding. Both embeddings are concatenated to get the final prediction result. Experiments show the performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Privacy leakage. Before training, clients communicate to calculate the L-hop neighborhood matrix $\\hat{A}$. In the 2-hop case, since the client knows 1-hop neighbors and the information during the communication, it is still able to reconstruct the 2-hop graph. Pruning cannot guarantee the privacy.\n2. FedSage+ and FedGCN can outperform FedStruct.\n3. In FedGCN, the server does not require a global adjacency matrix for homomorphic encryption. It only needs to know the node ids for encrypted aggregation and identify which nodes belong to each client for sending the aggregation result back." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See the weakness." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1)\tThis paper studies a significant and interesting problem, and the method can be used in a wide range of real-world applications. \n2)\tThe paper is overall well motivated. The proposed model is reasonable and sound. Theoretical analysis is performed." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel framework, FEDSTRUCT, to tackle the challenge of federated learning on graph-structured data distributed across multiple clients, particularly in scenarios involving interconnected subgraphs, it utilizes explicit global graph structure information to capture inter-node dependencies. The effectiveness of FEDSTRUCT is validated through extensive experiments on six datasets for semi-supervised node classification, demonstrating performance that approaches that of centralized methods across various scenarios, including different data partitioning strategies, levels of label availability, and numbers of clients." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The abstract lacks a description of the background. I recommend briefly outlining the context of the issues addressed in this paper before elaborating on the key problems that are solved.\n\n2) Figure 1 has not been cited and its placement is too early in the text; please adjust this detail. Additionally, Figure 2 is unclear; I recommend adjusting the proportions or border thickness of each subfigure.\n\n3) In the Related Work section, you mention that FED-STAR shares structural knowledge, yet in the conclusion, you state, \"No work has leveraged explicit structural information in SFL.\" Are \"structural knowledge\" and \"structural information\" the same concept? Please provide more clarification in the conclusion.\n\n4) The formula following (1) is missing a comma; please check for similar issues throughout the paper.\n\n5) Privacy is one of the directions addressed in this paper, yet most references are to other works. I suggest including some original proofs or experiments related to privacy to enhance the completeness of the article." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "see weaknesses.\n\nIf the first questions can be well explained, the rating should be higher." }, "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 method is novel, utilizing augmented explicit structure which can be regarded as global knowledge to promote the performance of the SFL model. \n2. Utilizing pruning skills decrease the calculation complexity and communication costs.\n3. Well written and well formulated problem." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a novel SFL method called FEDSTRUCT, which leverages the augmented explicit structure $\\bar{A}$ to promote the SFL model performance. Moreover, they propose HOP2VEC to learn local structure embedding. FEDSTRUCT precalculates the $\\bar{A}$ with privacy protection and prunes the $\\hat{A}$ matrix to decrease the calculation complexity and communication costs, thus balancing the communication-privacy-accuracy trilemma." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Focus on Privacy**. How to obstain the local L-hop combined adjacency matrix while not share the L-hop global Adjacency Matrix maybe play the core role in FEDSTRUCT. In APP D, the equations [30] [31], [32], what does the $\\hat{A}^{[K]}_j$ mean? Should be $\\hat{A}^{[k]}_j$? If so , the next question, for client $i$, how does it know all $\\hat{A}^{[k]}_j$ for $ k \\in [K]$ without sharing the global adjacency matrix in all clients. So another question when computing, the $\\tilde{A}^{[i]}_j \\in \\mathbb{R}^{|\\tilde{V}_i| \\times |{V}_j|}$, the same to $\\hat{A}^{[i]}_j$? So $\\hat{A}^{[i]}_k \\times \\hat{A}^{[k]}_j$ should be $\\mathbb{R}^{|\\tilde{V}_k| \\times |{V}_k|} \\times \\mathbb{R}^{|\\tilde{V}_k| \\times |{V}_j|}$， but according the definition before, the $|\\tilde{V}_k| \\neq |{V}_k| $, how does the computation continue? Maybe I miss something? I really hope you can explain it for me to understand the feasibility of FEDSTRUCT. That's my main concern about this paper.\n\n2. **About the hyperparameters.** **1)** The analysis of $\\beta$ is not enough, an essential parameter in FEDSTRUCT for various homophilic and heterophilic graphs, which directly dominate the performance and affect the judgment of FEDSTRUCT's contributions. **2)** It is so strange that the parameter $L_s$ and $L$ is set to1 for heterophilic graph chameleon in table 5. As the author states in lines 297-299, a heterophilic graph should own multi-hop nodes and a high-frequency filter to augment the local graph representation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024decoupled,\ntitle={Decoupled Subgraph Federated Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v1rFkElnIn},\nnote={under review}\n}" }, "abstract": { "value": "We address the challenge of federated learning on graph-structured data distributed across multiple clients. Specifically, we focus on the prevalent scenario of interconnected subgraphs, where inter-connections between different clients play a critical role. We present a novel framework for this scenario, named FedStruct, that harnesses deep structural dependencies. To uphold privacy, unlike existing methods, FedStruct eliminates the necessity of sharing or generating sensitive node features or embeddings among clients. Instead, it leverages explicit global graph structure information to capture inter-node dependencies. We validate the effectiveness of FedStruct through experimental results conducted on six datasets for semi-supervised node classification, showcasing performance close to the centralized approach across various scenarios, including different data partitioning methods, varying levels of label availability, and number of clients." }, "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": [ "Federated Learning", "Subgraph Federated Learning", "Inter-Connected Graphs", "GNN", "Decoupled GCN" ] }, "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/b3b1d0dc226f6dd7ac30eeb7faae000a011fa159.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": "Decoupled Subgraph Federated 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": null, "code_of_ethics": null, "comment": { "value": "We would like to thank our sincere gratitude to the reviewers and Area Chairs for their invaluable feedback and insights. After careful consideration, we have decided to withdraw our submission to further develop and refine our paper based on these constructive comments. Thank you once again for your time and thoughtful evaluation." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. Many things in the paper are not defined properly. Please clearly specify the definitions of calibration and modularity. \n\n2. Why MAE is not used as a metric in the experiments, as all the loss functions used are surrogates for it. It would be interesting to see comparison results on MAE.\n\n3. Please explain the regularization term in detail as it is still unclear how it promotes unimodularity. A graphical explanation will also help.\n\n4. A theoretical analysis of calibration and unimodularity of the proposed approach is missing. \n\n5. Is the proposed approach rank consistent? If not, then how frequently it violates ranking of thresholds. An experimental study on it will be helpful.\n\n6. How does the performance of the model vary with the variation in thevalue of $t$ used in the regularization term." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. a novel regularization term is used to promote unimodularity.\n2. Paper is well written and easy to read." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, authors propose an approach for calibration of ordinal regression. They propose a loss function that introduces order-aware calibration They use soft ordinal encoding and label-smoothing-based regularization to enforce both calibration and unimodality. To show the efficiency of the proposed approach, authors propose extensive experimental results on benchmark datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Theoretical proofs of calibration and unimodularity are missing." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. In equation 3, we generally adopt a hyperparameter $\\lambda$ to balance the loss and the regularization, like $L = L_1 + \\lambda L_2$. Could you explain why this method does not require such a hyperparameter? And why $t$ can control the strength of this regularization?\n\n2. Could you explain why SCE and ACE are improved a lot, but ECE is 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 problem of calibration in the context of ordinal regression sounds novel and important. As far as I know, this work should be the first work to solve this issue.\n\n2. The improvement is significant empirically. From Table 2, we can observe a great improvement in the calibration of ordinal regression models and the classification accuracy is preserved." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors aim to enhance the confidence calibration of ordinal regression in the training stage. The main challenge of this task is to consider the calibration and unimodality together. To address this challenge, they propose a new loss function for ordinal regression, which combines order-aware calibration with a unimodal regularization term (based on the SORD encoding). In particular, their method enforces both calibration and unimodality by explicitly modeling the ordinal relationships between classes. The effectiveness of their method is validated on three public datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The L_{REG} defined in Equation 2 is not clearly explained. In particular, the design of I(r) is hard to understand for readers. It would be better if the authors could elaborate on how the regularization is constructed.\n\n2. The writing of the gradient analysis in Subsection 3.4 is not clear. The authors may need to improve the writing in this part, or it might be too challenging for readers to follow.\n\n3. The technical novelty of the proposed method is not presented. While the authors claim that the method considers the unimodality compared to current calibration methods and considers the calibration when compared to current Ordinal losses, I am not clear about if this method is newly designed in each aspect. In other words, the authors may need to show the new insight of calibration part compared to calibration methods.\n\ntypos:\n1. Line 49, Oridnal -> Ordinal.\n\nThe major issue of this work is on the writing: readers cannot easily understand why we should design such a regularization and how t works here. I will improve my score if the authors can make it clear in the revised version." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "•\tThe motivation behind the new method is well-articulated, clearly highlighting the limitations of traditional cross-entropy (CE) loss in ordinal tasks and the miscalibration in modern ordinal regression models.\n•\tThe proposed method is straightforward, and easy to implement." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposed method introduces ORCU, a novel loss function designed to ensure calibration and unimodality in ordinal regression tasks. ORCU leverages soft ordinal encoding and order-aware regularization to produce calibrated and unimodal probability distributions, which are particularly valuable in high-stakes applications requiring reliable confidence estimates and accurate predictions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tIt is unclear how the regularization component of ORCU promotes calibration. Lines 240-252 discuss scenarios where the model is under or overconfident, yet this confidence is based on a soft-encoded distribution not directly related to the data, which raises questions about its reflection of \"real\" confidence. Additionally, I would appreciate a more rigorous explanation of how this regularization approach aligns with the standard mathematical definition of calibration. Could the authors provide a clearer mathematical justification for this relationship?\n\n2.\tThe paper claims that CE loss leads to overconfident predictions, yet the reliability diagrams presented indicate underconfident outcomes in the experiments, seemingly contradicting this claim. Temperature scaling, a prominent calibration technique, relies on CE, further challenging the assertion that CE is fundamentally flawed for calibration. Could the authors address this discrepancy and clarify why their results show underconfidence in CE where overconfidence might be expected? Additionally, a nuanced discussion of CE’s strengths and limitations for calibration, especially in light of techniques like temperature scaling, would be valuable.\n\n3.\tThe evaluation is limited to loss function baselines. Including additional non-loss-based methods for ordinal regression, such as the approach presented in https://arxiv.org/pdf/2303.04547 for unimodality could highlight the unique benefits of ORCU more effectively. I recommend incorporating a discussion on why ORCU and loss function-based methods may offer advantages over such approaches.\n\n4.\tThe experiments were conducted on only three datasets, which limits the scope for evaluating the method’s robustness across a wider range of ordinal regression tasks. Incorporating a more extensive dataset selection would allow for a better assessment of the generalizability of the 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": 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": "- Please address the points I raised in the weaknesses part\n- The methods that are not based on CE loss - like optimal transport loss - how are the limitations applied to them? \n- Missing definition of z_{n.k} and intermediate step to get f’=y - p, also r is not defined before it presented \n- Can you elaborate please why the gradient encourages the model to distribute probability across adjacent classes ( as you claim in the sentence in Line 176) compared to the statement in Lines 147-149" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- This work addresses an important overconfidence issue in ordinal regression tasks\n- The proposed loss function is assumed to address both accuracy and confidence of the cross entropy loss based model during optimization without additional post-training calibration\n- The authors justify the unimodality enforcement of the proposed loss by gradient analysis" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors proposed an addition to the cross entropy loss term that corrects overconfident predictions for incorrect labels. The advantage of the proposed loss is that the calibration is done jointly with the accurate prediction learning during the training optimization and doesn’t require additional post-training steps. The method was evaluated on the 3 datasets and compared against CE-based ordinal models and calibration-loss based and showed an improvement in calibration/accuracy metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "**Major**:\n- The main focus of the work is CE-loss based ordinal regression which is not an optimal loss for this task and several methods were proposed without CE loss: [1-4]\n- The motivation in Sec 3.1 is unclear, how the calibration is defined and why it is not implied by CE-loss. The discussion seems to be valid for the ordered nature of classes but not for calibration. It is better to discuss the motivation for each problem separately.\n - $\\mathcal{L}_{SCE}$ - the explanation in L175-177 is unclear, how the defined loss encourages what the authors claim - maybe it is explained by Diaz et. al but the manuscript should be self contained with additional clarification. It is also not clear how it helps to reflect the ordinal relationships.\n- The Sec. 3.3 in unclear, the explanation and derivation of the loss formula should come before presenting the loss term \n - why the authors choose it\n - how it helps to ensure calibration\n - what is r ?\n - The L181-183 is unclear.\n\n- Weak evaluation with only 3 small-sample datasets - overall the improvement is incremental so presenting results on more datasets could be beneficial.\n- Missing unimodality evaluation - the authors claim the model enforces unimodality - please show it in the results as in [5]\n- Missing additional deep ordinal regression baselines [1-5].\n- Sec. 3.4 - while it is clear why the loss term enforces unimodality, I’m not sure how it enforces calibration. By saying “*by increasing the gradient for such incorrect predictions, the model is able to reduce the predicted probability for the incorrect class more effectively*” you can claim the same for the standard CE loss. \n- while it could be seen from the results that calibration metrics improved, I’m not sure it is clear from the manuscript why it works.\n\n**Minor**:\n- Missing additional deep ordinal regression methods in the related work discussion\n- It is better to put Figure 1 closer to the gradient analysis section to make it easier to follow\n\nReferences:\n[1] Liu, X., et al. (2019a). Unimodal-uniform constrained wasserstein training for medical diagnosis. In Proceedings of the IEEE International Conference on Computer Vision Workshops\n\n[2] Beckham, C. et al. (2017). Unimodal probability distributions for deep ordinal classification.\n\n[3] Wenzhi Cao et. al (2020). Rank Consistent Ordinal Regression for Neural Networks with Application to Age Estimation. Pattern Recognition\n\n[4] Xintong Shi et al. (2021). Deep Neural Networks for Rank-Consistent Ordinal Regression Based On Conditional Probabilities.\n\n[5] Cardoso, J. S. et. al (2023). Unimodal distributions for ordinal regression" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a loss function that introduces order-aware calibration in ordinal regression tasks, combining soft ordinal encoding and label-smoothing-based regularization to enforce both calibration and unimodality." }, "_bibtex": { "value": "@misc{\nkim2024calibration,\ntitle={Calibration of ordinal regression networks},\nauthor={Daehwan Kim and Haejun Chung and Ikbeom Jang},\nyear={2024},\nurl={https://openreview.net/forum?id=v27yHgKtMv}\n}" }, "abstract": { "value": "Recent studies have shown that deep neural networks are not well-calibrated and produce over-confident predictions.\nThe miscalibration issue primarily stems from the minimization of cross-entropy, which aims to align predicted softmax probabilities with one-hot labels. In ordinal regression tasks, this problem is compounded by an additional challenge: the expectation that softmax probabilities should exhibit unimodal distribution is not met with cross-entropy. Rather, the ordinal regression literature has focused on unimodality and overlooked calibration. To address these issues, we propose a novel loss function that introduces order-aware calibration, ensuring that prediction confidence adheres to ordinal relationships between classes. It incorporates soft ordinal encoding and label-smoothing-based regularization to enforce both calibration and unimodality. Extensive experiments across three popular ordinal regression benchmarks demonstrate that our approach achieves state-of-the-art calibration without compromising accuracy." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Daehwan_Kim4", "~Haejun_Chung1", "~Ikbeom_Jang1" ] }, "authors": { "value": [ "Daehwan Kim", "Haejun Chung", "Ikbeom Jang" ] }, "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": [ "Ordinal regression", "Calibration", "Deep neural networks", "Unimodality", "Loss function", "Soft ordinal encoding", "Label smoothing", "Order-aware calibration" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "kim|calibration_of_ordinal_regression_networks" }, "pdf": { "value": "/pdf/68595e0558e86c2f45ec3d79e27ab7eb06ef8403.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/1442042c4f8b5c9a9fddfdf0db7ec9f12fc8d9d1.zip" }, "title": { "value": "Calibration of ordinal regression networks" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Will the author build a large-scale dataset that can be released to the public?\nThe author mentioned the model only generates sparse rewards, will the batch RL be affected since only parts of the data have rewards?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The experiments are extensive. This paper compares with many SOTA baselines and makes many useful analyses. It shows the effectiveness of the proposed framework. \n2. This paper is easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a framework, Proposer-Agent-Evaluator(PAE), to boost the performance of agents in web tasks. In more detail, it uses a foundation model to automatically generate task instructions and another model to evaluate task completion and generate a reward. Lastly, RL is used to train a policy model." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The main weakness is still the novelty. The method can be divided into two parts. Utilize existing foundation models to process data or automatically generate data. This part is not novel. Many other works have tried similar pipelines to collect data. The second part is using RL to fine-tune a model, which is also a standard procedure. \n\n\n[1] Navigating the Digital World as Humans Do.UNIVERSAL VISUAL GROUNDING FOR GUI AGENTS" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 PAE, why sample tasks uniformly rather than prioritizing by difficulty or learning progress?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Demonstrates an increase in performance for open-source model weights compared to previous open-source models.\n- PAE framework leverages self-generated tasks, avoiding static human-annotated instructions." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents Proposer-Agent-Evaluator (PAE), a framework enabling autonomous skill discovery for vision-language model (VLM) internet agents. By employing a context-aware task proposer and an autonomous evaluator, PAE allows VLM agents to independently explore and refine skills for web navigation tasks, demonstrating improvements in success rates on benchmarks like WebVoyager and WebArena Easy compared to other open-source models. However, the model still relies on closed-source VLMs for effective task generation and evaluation, thus partially constraining the proposed framework’s open-source utility." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The term \"model-based task proposer\" is used ambiguously. In reinforcement learning (RL), \"model-based\" generally implies use of a dynamics model for planning, whereas in this paper it refers to LLM prompting, which can lead to confusion in terminology.\n- The novelty and motivation of the contributions are limited. Although the paper claims state-of-the-art (SoTA) for open-source models, it requires closed-source models for generating and evaluating tasks. This reliance on closed-source models highlights that the bottleneck remains the quality of these closed-source models, rather than any specific methodological improvement introduced by PAE.\n- The paper makes an assumption that the ground-truth task distribution and reward functions are inaccessible, even though in simulated environments, these can often be directly obtained (e.g., verifying if a button was clicked). If the goal is to improve policy generalization skills, and use ground-truth distributions and rewards only as a way to evaluate the method, the paper would benefit from a clearer rationale and explanation for this assumption (Section 3.1).\n- The paper lacks references to relevant prior work in autonomous task proposal and RL, and some claims of novelty are inaccurate, given that similar approaches have been explored [1, 2, 3, 4].\n\n[1] Zhang, J., Zhang, J., Pertsch, K., Liu, Z., Ren, X., Chang, M., ... & Lim, J. J. (2023). Bootstrap your own skills: Learning to solve new tasks with large language model guidance. arXiv preprint arXiv:2310.10021.\n[2] Zhang, J., Lehman, J., Stanley, K., & Clune, J. (2023). Omni: Open-endedness via models of human notions of interestingness. arXiv preprint arXiv:2306.01711.\n[3] Faldor, M., Zhang, J., Cully, A., & Clune, J. (2024). OMNI-EPIC: Open-endedness via Models of human Notions of Interestingness with Environments Programmed in Code. arXiv preprint arXiv:2405.15568.\n[4] Colas, C., Teodorescu, L., Oudeyer, P. Y., Yuan, X., & Côté, M. A. (2023, November). Augmenting autotelic agents with large language models. In Conference on Lifelong Learning Agents (pp. 205-226). PMLR." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "How does PAE differ from other methods for skill discovery via foundation models such as OMNI?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "This paper demonstrates several notable strengths. The authors conducted comprehensive evaluations of their PAE framework across multiple challenging benchmarks, including WebVoyager and WebArena. They compared their method against a range of baselines, including proprietary models, state-of-the-art open-source vision-language models, and supervised fine-tuning approaches. This thorough evaluation provides a clear picture of PAE's performance in relation to existing methods.\n\nThe paper also presents a detailed analysis of the results, including error analysis and generalization studies. The authors break down different types of errors made by the models, providing insights into where improvements are made and what challenges remain. They also examine how well the skills learned through PAE transfer to unseen websites, demonstrating the framework's ability to develop general web browsing capabilities. Additionally, the authors investigate how PAE scales with larger base models and explore the impact of different context information on performance. This level of analysis and discussion significantly strengthens the paper by providing a nuanced understanding of the method's capabilities and limitations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces Proposer-Agent-Evaluator (PAE), a novel framework for autonomous skill discovery in foundation model agents, particularly focusing on Internet agents. PAE consists of three key components: a context-aware task proposer, an agent policy, and an autonomous evaluator. The framework enables agents to autonomously discover and practice skills without human supervision, potentially leading to a more diverse and scalable skill repertoire. The authors validate PAE on challenging vision-based web navigation tasks using both real-world and self-hosted websites. Results show that PAE significantly improves the zero-shot generalization capability of VLM Internet agents. Notably, the PAE-trained model outperforms other state-of-the-art open-source VLM agents. The authors claim this work represents the first attempt to apply autonomous task proposal with reinforcement learning for agents, achieving state-of-the-art performance among open-source models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am a little unclear about the novelty of the actual algorithm underlying PAE in comparison to something like OMNI (https://arxiv.org/pdf/2306.01711) and subsequent work. I think applying this kind of open-ended task selection for skill discovery to the internet agent case is definitely unique, but how does PAE differ from other open-ended skill discovery algorithms in other RL domains? If the authors could clarify this either in the related work section or after introducing their method in section 3.3 that would be really helpful.\n\nOverall the results of the paper are strong, but framing the algorithmic novelty can be done with more clarity. If the novelty arises from the domain then this point and the difficulties of internet-based environments should be emphasized more rather than as just a potential application case as section 4 treats it as." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Response to the questions and concerns mentioned in the Weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The paper is well written and easy to follow. The problem of foundation model agents and the PAE method is formally defined and clearly illustrated with figures. The idea of the Proposer-Agent-Evaluator framework seems reasonable in that the proposer should approximate the real-world task distribution and the agent learns to maximize the autonomous reward function.\n2. Extensive and systematic experiments are carried out to verify the effectiveness of PAE compared with untrained VLMs and SFT training. Error analysis and the study of context complexity also show that the context-aware method can discover low-level web skills." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new framework called Proposer-Agent-Evaluator (PAE) which allows foundation model agents to autonomously discover and practice skills in the wild. The framework consists of a context-aware task proposer that suggests tasks for the agent to practice with website context information, an agent policy that attempts these tasks in the real world, and an autonomous model-based success evaluator that evaluates the resulting trajectories. The success evaluation serves as the reward signal for the agent to refine its policies. The authors validate PAE on challenging vision-based web navigation using real-world and self-hosted websites and demonstrate significant improvements in the zero-shot generalization capability of VLM Internet agents compared to state-of-the-art open-source VLM agents." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the PAE framework seems reasonable, the method itself is naive and lacks novelty. Basically, the authors train a VLM with data containing instructions and successful trajectories relying on the proprietary VLMs, which is a common approach and does not include inspirational techniques. Although human annotation may overlook some long-tail real-world tasks, it's unclear why VLMs provided with web context can approximate the real-world task distribution. Also, behavior cloning the successful trajectories labeled by a VLM should not be described as reinforcement learning (line 117).\n 2. It's not explained why behavior cloning is adopted instead of other training methods including SFT to discover web skills. Is the LLaVa-SFT baseline trained on the same trajectories as the proposed PAE model? What are the performances of other leading-edge VLMs such as GPT-4o? Comparisons with other trained web agents on Web Voyager or Web Arena are also missing.\n3. Several typos in line 91 ('a foundation model agents') and line 191 (task distribution R)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": { "value": "Do you have the right permissions in place to sample from the user demos on the websites (Section 4.2 for the context aware task proposer)" }, "flag_for_ethics_review": { "value": [ "Yes, Legal compliance (e.g., GDPR, copyright, terms of use)" ] }, "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": "Clarification Questions:\n\n-\tFor all components, the task proposer, autonomous evaluator and agent policy, could you elaborate more on the choice of VLMs: Claude-3-Sonnet and the LLaVa models? Why those and not others?\n-\tFor evaluation, could you elaborate on your choice of open-source VLMs? Why are they chosen as opposed to others?\n-\tIn the main results, under the scaling paragraph, line 400, I’m not quite clear on the following statement “Again, LLaVa-34B PAE beats LLaVa-7B SFT on 12 out of 13..”. Is that the intended comparison, or did you mean to compare across PAE models only?\n-\tIn the main results, under the generalization paragraph, line 405, why is LLaVa-7B PAE highlighted, when actually you present results for the larger LLaVa-34B PAE as well in Table 3? Also in Table 3, why did you drop some of the other models from Tables 1 and 2 (Claude 3.5, InternVL and vanilla LLaVa-7/34B)?\n-\tIn the alignment with human judgements, could you elaborate on how many human annotators were included in the user study and if possible how many hours they allocated to the task?\n\nMinor comments/suggestions:\n\n-\tThere are a few minor typos: on line 91 “for foundation model agents” (remove “a”), on line 191 in section 3.2, task distribution should be C instead of R, on line 485 “LLaVa-7B SFT knows that”\n-\tFor consistency across the entire paper, use Claude-3-Sonnet – on line 269 in section 4.3, it is referred to as Claude-Sonnet-3.\n-\tLine 317 the LLaVa model name needs to be in bold to align with the others." }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The author(s) highlight and prove the feasibility of using much smaller VLMs and thus significantly lowering the test-time compute through their method, as opposed to finetuning larger models.\n\n- The author(s) are willing to opensource the code and models to encourage further development in the research community.\n\n- Overall, the paper is easy follow and clearly motivates the research question addressed and the main contribution in relation to existing literature. Concepts introduced in the main paper nicely point to corresponding appendix sections when needed which provide .\n\n- The author(s) are doing a great job explaining how their work differentiates itself from existing research along 3 separate dimensions: foundation model agents, self-generated instructions and unsupervised skill discovery in RL. I particularly appreciate the focus on enabling the community to use much smaller sized models in terms of parameters, and therefore significantly less test-time compute.\n\n- I appreciated the analysis on failure modes in Section 6, highlighting the value added by the PAE method, as opposed to SFT.\n\n - It is very good to see that each action contains chain of thought element, aiding interpretability of the model choices. The author(s) also take preventive measures, as highlighted in their Ethics statements." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a framework for autonomous skill discovery for real-world Internet agents, called PAE (Proposer Agent Evaluator). Compared to existing approaches, this method allows for the agent to collect, explore and refine new skills automatically, without relying on a limited set of predefined human-annotated instructions. The author(s) show how the additional PAE components introduced (the task proposer, the agent policy and the autonomous evaluator) allow for zero-shot SOTA generalization performance, as well as better success rates against opensource VLMs and finetuned opensource VLMs on 2 environments, WebVoyager and WebArena." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- To strengthen the submission/contribution, it would be good to elaborate on why proprietary VLMs outperform all opensource methods.\n\n- It would be great to add some details/discussion on inference times, comparing the opensource VLMs, the SFT variation and PAE.\n\n- It wasn't very clear to me how you define the reward for the agent policy, could you elaborate on that?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024proposeragentevaluator,\ntitle={Proposer-Agent-Evaluator ({PAE}): Autonomous Skill Discovery For Foundation Model Internet Agents},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v2D1ASk5MT},\nnote={under review}\n}" }, "abstract": { "value": "The vision of a broadly capable and goal-directed agent, such as an Internet-browsing agent in the digital world and a household humanoid in the physical world, has rapidly advanced, thanks to the generalization capability of foundation models. Such a generalist agent needs to have a large and diverse skill repertoire, such as finding directions between two travel locations and buying specific items from the Internet. If each skill needs to be specified manually through a fixed set of human-annotated instructions, the agent's skill repertoire will necessarily be limited due to the quantity and diversity of human-annotated instructions. In this work, we address this challenge by introducing Proposer-Agent-Evaluator (PAE), a novel framework that enables foundation model agents to autonomously discover and practice skills in the wild. At the heart of PAE is a context-aware task proposer that autonomously proposes tasks for the agent to practice with context information of the websites such as user demos or even just the name of the website itself. Then, the agent policy attempts those tasks in the real world with resulting trajectories evaluated by an autonomous model-based success evaluator. The success evaluation serves as the reward signal for the agent to refine its policies through RL. We validate PAE on challenging vision-based web navigation, using both real-world and self-hosted websites from WebVoyager and WebArena. Our results show that PAE significantly improves the zero-shot generalization capability of VLM Internet agents (more than 30\\% relative improvement) to both unseen tasks and websites. Our model also achieves an absolute advantage of over 10\\% (from 22.6\\% to 33.0\\%) comparing to other state-of-the-art open source VLM agents including Qwen2VL-72B. To the best of our knowledge, this work represents the first attempt to apply autonomous task proposal with RL for agents, achieving SOTA performance among open-source models. We plan to release our models and code to facilitate further research." }, "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": [ "VLM Agent", "Web/GUI Agent", "VLM", "Reinforcement Learning", "Skill 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/0aabb0e488a607064496d67722adb71ad368919a.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/48560904762adc22d5ec38b5486cf459a69bf043.pdf" }, "title": { "value": "Proposer-Agent-Evaluator (PAE): Autonomous Skill Discovery For Foundation Model Internet 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": { "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": "- In section 1.3, what is the definition of \"surprise\"?\n- It is stated in Section 1.3 that DRL requires \"fixed environments\", which I believe is not true. Could you please clarify this?\n- The equation in page \"v\" is vague" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- The NetAIF framework, especially the application of random pullback attractor to it is interesting" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes NetAIF, a network-based adaptive inference framework, based on random attractor dynamics and Free Energy Principles (FEP), to improve adaptive control problems in robotic systems, and PV panel inspection. The experimental results show that NetAIF performs well on a PV panel inspection and a robotic tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The introduction contains several statements and justifications without clear evidence or relevant citation.\n- The number of references used in this study is low (only 21), several of which are not peer-reviewed.\n- There is a rich body of literature that supports using (deep)RL for similar applications. The application of these approaches is not studied in detail. \n- I expected the authors to use SOTA (deep)RL and AIF approaches as baselines for comparison with NetAIF. Therefore, the experiment section is limited\n- The paper lacks adequate background and theory about different components of the method such as AIF, FEP, RDS, and random attractor dynamics.\n- The equation numbers are missing\n- Section 2 lacks effective connections between its subsections, which results in disjointed and potentially misleading descriptions. This section should be thoroughly revised to improve readability and flow. Additionally, the relationship between the equations and Algorithm 1 is unclear." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "The re-use of considerable parts of text between the two papers may be problematic; it may be worth checking it.\n\nThere seems to be strong self-plagiarism." }, "flag_for_ethics_review": { "value": [ "Yes, Research integrity issues (e.g., plagiarism, dual submission)" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Can the authors please elaborate on the differences between this paper and the connected one?" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- If verified thoroughly, the reduced computational requirements and real-time capabilities are promising.\n- The method is designed for practical applications, which is nice." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes Network-based Active Inference (NetAIF), a framework for trajectory generation and control in robotics. NetAIF integrates random attractor dynamics and the Free Energy Principle (FEP) to create a system that can adapt in real-time with low computational requirements. The authors focus on a practical application in PV panel inspections, demonstrating the system's efficiency, adaptability, and robustness. Evaluation is performed in a simulated Mujoco environment, and on a very simple task on a real robot." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The implementation details are not very clear; from what I understand, the method does not seem particularly novel.\n\n- Experimental Validation: The evaluation is limited to simulations and controlled lab conditions, which makes the results not very convincing. In particular, no baseline is provided.\n\n- Figures 1 and 2 are not particularly informative, nor are their captions (e.g., for figure 1, \"parameters that determine the network\nstructure such as number of layers, strides were determined through hyper parameter search\").\n- Table 4 seems overkill to report the results; the values would be more appropriate just mentioned in the text (as they are) and in the caption of figure 11.\n\n- Minor: is the template correct? This is the only one among the ICLR papers I reviewed that used roman numerals for the page numbers.\n- Minor: reporting network sizes in bytes is not very useful, and it's the first time I see it in a paper. Reporting the number of parameters is better.\n\n- Possibly major: The authors cite a paper from the same group in concurrent submission to ICLR ( https://openreview.net/pdf?id=Hm7RYDspQP ); the two papers seem to have major overlap (including using many of the same figures and whole sections / parts of the text). It feels like the authors have tried to write two papers on the same method, splitting evaluations and 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": 1 }, "primary_area": null, "questions": { "value": "- What do you mean with \"Unlike DRL, which requires fixed environments,\". DRL methods can generalize quite well to unstructured unseen environments, see for instance https://arxiv.org/abs/2109.11978, https://arxiv.org/pdf/2010.11251\n\n- The paper does not address how the model would adapt to obstacles in the environment. How does the current approach account for or avoid collisions?\n\n- The statement “the random attractor dynamically explores within the constraints set by the control law” raises some questions:\n - How are these constraints practically defined?\n - What control law ensure that exploration remains controlled and the robot behaves predictably during this process?\n\n- The pseudo algorithm suggests that prediction errors are minimized by randomly sampling new weights, with updates only occurring if the input surpasses a certain threshold. A few points to clarify:\n - How is this threshold determined?\n - Why is this threshold-based sampling considered an optimal method for weight updates?\n - What does it mean to reset a weight? Is it reset to a default value? If so, which value?\n\n - “The joint pose was directly fed into the system, and the attractor calculated waypoints for a smooth and efficient trajectory to the specified pose,”:\n - What specifically is the simple control law—is it a linear law?\n - How are waypoints computed? Is it a linear interpolation between current and target joint angles?\n - Why is there not a single goal-based attractor, and does the choice affect stability?\n\n- The text states that “noise affecting the system at time t is related to past noise.” Does this formulation assume a colored noise model?\n\n- Based on Table I, the robotic arm with 6 DOF has a position accuracy of 14 mm from a fixed base. While this may suffice for inspecting large panels, it is insufficient for precision tasks, like assembly, which require sub-millimeter accuracy." }, "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 a novel robot control algorithm, and evaluate both in simulation and on a real robot arm." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents \"Network-based Active Inference\" (NetAIF), a novel framework for trajectory generation and control in robotics. The main novelty is replacing traditional activation functions with a discrete weight-assigning mechanism, especially for a system focused on achieving and maintaining stability within a NESS framework. The authors mainly focus on a photovoltaic panel inspection task, where a robot arm needs to reach a pre-determined distance from a panel, perpendicular to the panel's surface. The experiments evaluate error and planning time of the system both in simulation and on a real robot arm." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper focuses on the particular task of PV panel inspection in title and abstract. However, the actual task considered is controlling a robotic arm to a particular end-effector pose and orientation. I would think there are many other approaches to do PV panel inspection, and wouldn't necessarily require a robot arm (e.g. using a drone for instance). Also the actual \"visual inspection\" is not addressed. \n\n- The introduction discusses deep RL systems, but the paper does no comparison against any of those, but refers to another paper in the conclusion.\n\n- The methodology section lacks detail, making it challenging to reproduce the work. Comparative analysis with traditional control methods in particular for visual servoing, which are well-established for inspection tasks, is also absent.\n\n- The paper lacks a formal proof of system stability. To verify stability, the authors might consider using Lyapunov’s stability criteria. Identifying a well-defined Lyapunov function could reveal whether trajectories converge to an attractor, A(ω), suggesting stability." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": 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 questions in weaknesses" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper proposes an alternative to reinforcement learning by adopting the Active Inference framework from cognitive neuroscience, in which perception, action, and learning are obtained through the minimization of variational free energy." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces NetAIF, a method that integrates random attractor dynamics and the Free Energy Principle (FEP) to improve trajectory generation and control in robotics. The paper shows the performance of NetAIF in applications of PV Inspection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Overall, the paper is not well-written and hard to follow. It gives a superficial idea without going through technical details. I do not understand the mathematics behind NetAIF from this paper. More explicit mathematical equations/formulas should be written: what are the explicit formulas of (variational) Free Energy and surprise? What is the structure of networks? Are they multilayer perceptions? What are the parameters and how are they being updated? how are the explicit feedback loops sent to the previous layers?\n\nThe algorithm 1 is not clearly explained and hard to reproduce. Please elaborate more what do each variable represents and how are they obtained/calculated.\n\nThe algorithm 1 used “desired state” which means it required the ground truth of optimal trajectory. Would this not be equivalent to having a reward or a set of demonstrations of experts in Deep reinforcement learning?\n\nExperiments lack benchmarks and comparisons with DRL methods. It is not clear whether the results are good. Are PRM and Hybrid RRT-PRM tested in this experiment?\n\nIt is also not clear how your work is different from “Deep active inference as variational policy gradients (Millidge, 2020)” and “Deep active inference (”K. Ueltzhoffer, 2018”)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "no concerns on ethics" }, "flag_for_ethics_review": { "value": [ "No 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 and sentence structure : the paper could use a read through and improve upon sentence structure and layout, also several typos, omissions could be rectified\nex. Active Inference (AIF) = Active Inference Framework \nIn my opinion, given the complexity of the proposed solution the paper could benefit from a more rigorous mathematical treatment.While the work uses a novel approach the paper lacks mathematical rigor and this paper relies very heavily on published work forcing the reader to read several papers before extracting value from this work. While citations are needed and not everything needs to be reintroduced, the paper should have enough content to stand on its own. \nMathematical Model Integration: can the authors incorporate mathematical models demonstrating how NetAIF functions in practice ? This could involve equations showing how the network computes trajectories, adjusts weights, or minimizes prediction errors dynamically.\nFormal Optimization Framework: Strongly suggest that authors introduce a formal optimization framework to enhance enhance the credibility of claims regarding efficiency and adaptability. For instance, showing how free energy is minimized using a variational approach or Bayesian inference would connect the theoretical claims with a concrete mathematical foundation.\nSensitivity Analysis and Stability Proofs: This is another weakness of the paper. Suggest authors to include a mathematical analysis of system stability and sensitivity to changes in parameters or inputs. This would validate the robustness of NetAIF and its applicability to real-world scenarios beyond the initial PV panel inspection.\nStability analysis: it would be very beneficial to see a formal treatment on the intentional use of controlled instability and/or some plots in that regard." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Novelty:The use of FEP with random attractor dynamics, for real-time control, is innovative. This combination offers an efficient alternative to Deep Reinforcement Learning (DRL) by minimizing computational requirements and avoids extensive pre-training​.\nReal-world Applicability: The application of NetAIF in PV panel inspections is practical and addresses significant challenges in the clean energy sector. The use of a physical 6-DoF robotic arm for experiments shows a commitment to real-world validation​\nEfficiency: The framework’s low computational footprint and rapid adaptability in dynamic environments are highlighted as key benefits, making it suitable for industries needing quick, cost-effective automation solutions​." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces Network-based Active Inference (NetAIF) as a novel framework that leverages the Free Energy Principle (FEP) and random attractor dynamics for efficient and adaptive robotics. \nThe authors demonstrate their ideas through the use case of a photovoltaic (PV) panel inspection. The authors claim that \"NetAIF optimizes the intrinsic dynamics of neural networks and enables robots to quickly adapt to dynamic and complex real-world environments with minimal computational re- sources and without the need for extensive pre-training. Unlike traditional learning methods that rely on large datasets and prolonged training periods, NetAIF offers a more efficient alternative.\"\nThe authors also provide supplementary material which is a slide show of the paper but the experiments with the 6DOF robotic arm show performance against a few experiments in a lab setting. \n\nWhile the work uses a novel approach the paper lacks mathematical rigor and the paper relies very heavily on a few papers forcing the reader to read several papers before extracting value form this work. While citations are needed and not everything needs to be reintroduced the paper should have enough content to stand on its own. This makes it challenging to evaluate and reproduce the proposed NetAIF framework. Without any formal mathematical treatment, the claims remain at a high level, reducing the work’s overall scientific and practical value. Integrating a comprehensive mathematical model would significantly enhance the credibility and applicability of the approach. For example, the use of stochastic processes and attractors should be backed by a comprehensive set of equations that demonstrate how these elements interact dynamically within the network. Without this, it is unclear how the system transitions between states, adapts to sensory inputs, and minimizes free energy in a precise manner, additionally , intentionally introducing instability has risks and managing instability should be critical for convergence, the authors mention controlled instabilities and cite a paper but do not provide any formal treatment or plots showing what this looks like." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Clarity: The explanation of the technical mechanisms, such as the use of random attractors and the free energy landscape, can be complex and may not be accessible to readers unfamiliar with advanced robotics or neural dynamics. Visual aids, while present, could have been better integrated to explain these concepts more intuitively​\nReproducibility: The paper lacks detailed implementation specifics, such as parameter settings and hyperparameters used during the experiments, which limits the reproducibility of results. Clearer code snippets or references to an open-source implementation would enhance its utility for other researchers​.\nStrong Assumptions: The paper assumes that the Free Energy Principle (FEP) is applicable to all dynamic robotic systems without extensive empirical validation outside the PV panel inspection case. Additionally, the scalability of the model to other industries or tasks is presumed but not explicitly tested​.\nCitations: Citations lack depth and specificity, especially in sections where novel methods are introduced.\nAuthors often rely on generic references rather than recent, more relevant studies directly supporting the claims made in the paper.\nSome citations are self-referential, reducing the overall credibility. The paper would benefit from a more thorough literature review, inclusion of detailed empirical comparisons from other studies, and references to supplementary or reproducible materials that validate the methods described.\nMathematical rigor: While the work uses a novel approach the paper lacks mathematical rigor. This makes it challenging to evaluate and reproduce the proposed NetAIF framework. Without any formal mathematical treatment, the claims remain at a high level, reducing the work’s overall scientific and practical value." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 Algorithm 1, what is the difference between Hidden_signals_prev and Hidden signals in lines 12-13?\n\n2. Is there some drawbacks (prices) to induce the random attractors?\n\n3. The function of feedback is still a bit confusing. If its implicit structure can be described in Fig 1?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The scheme of interrogating active inference and neural network.\nThe experimental demonstrations." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "By integrating active inference and neural network learning schemes, this paper presents a novel control framework, aiming to adapt to an unknown dynamical environment, without pre-training and extensive computations. The focused topic is important considering the shortcomings of traditional deep reinforcement learning. Several numerical simulations and experiments are conducted to demonstrate the proposed control strategy. Although some initial valuable results have been provided in this paper, there are still many important improvements that need to be considered, including clear presentations, solid theorem, and extensive comparisons." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Experimental comparisons** are lacking in this paper. The authors claim the proposed method has robust and lightweight advantages, in comparison with DRL, while there are no compared results quantitatively. Maybe some of the results are mentioned in the companion paper, but this paper needs to be self-contended. Moreover, the proposed strategy can avoid the traditional planning module and some feedback scheme is integrated into the proposed strategy. It is curious to see how the proposed method compares with the traditional approach (planning + feedback control or MPC).\n\n2. Improve the presentation quality. The used symbols in this paper are unfamiliar and lack illustrations. The symbols employed in Fig 3 need to be illustrated, otherwise, the purpose of intuitional expression will be lost. \n\n3. The key idea of the Net AIF is to introduce the controlled instabilities by random attractors, while the whole system is within the safety region. More rigid convergence or stability analyses are needed.\n\n4. Experimental details need to be added. For example, the accurately measured states, noised states, and unknown states in Section 3.1 should be clearly provided.\n\n5. An ablation study is needed. It seems that the proposed framework consists of several parts. The functionality of each part needs to be quantified. For example, the efficiency of the replaced discrete weight-assigning mechanism is unknown. \n\n6. The limitations should be analyzed finally. It is highly recommended to provide the source code." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "This paper introduces Network-based Active Inference (NetAIF), a game-changing framework that merges Active Inference with network dynamics, enabling adaptive real-time robotic control while dramatically reducing computational costs and time." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024networkbased,\ntitle={Network-based Active Inference for Adaptive and Cost-efficient Real-World Applications: {PV} Panel Inspection},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v2NuTf6Kww},\nnote={under review}\n}" }, "abstract": { "value": "This paper introduces Network-based Active Inference (NetAIF), a novel framework that integrates random attractor dynamics and the Free Energy Principle (FEP) to improve trajectory generation and control in robotics. NetAIF optimizes the intrinsic dynamics of neural networks, enabling robots to quickly adapt to dynamic and complex real-world environments with minimal computational resources and without the need for extensive pre-training. Unlike traditional learning methods that rely on large datasets and prolonged training periods, NetAIF offers a more efficient alternative. \n\nIn real-world scenarios, such as Photovoltaic (PV) panel inspections, NetAIF demonstrates its ability to execute dynamic tasks with both high efficiency and robustness. The system excels in unpredictable environments while maintaining a low computational footprint. These capabilities make NetAIF a promising solution for industrial applications, offering cost-effective, adaptive robotic systems that can reduce operational expenses and enhance performance, particularly in sectors like energy, where adaptability and precision are crucial." }, "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 Inference (AIF)", "Free Energy Principle (FEP)", "Robotics", "Trajectory generation", "Random dynamical systems", "Random attractor dynamics", "Non-Equilibrium Steady State (NESS)", "Adaptive control", "Industrial automation", "Computational efficiency", "Cost-efficient solutions" ] }, "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/0cf50653eb955fc5e6acd39b5559912d783b4cd0.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": { "value": "/attachment/6d2ad4a3f3c4ebb719dfea5cd195d75ec469a1d5.zip" }, "title": { "value": "Network-based Active Inference for Adaptive and Cost-efficient Real-World Applications: PV Panel Inspection" }, "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": "- How to compute the specturm? It appears to rely on time-consuming SVD operations.\n- The feasibility of applying this method to large-scale graphs is uncertain.\n- Runing time comparsion need to be added\n- Notable baseline methods are missing from the evaluation, particularly:\n - SGC (Simple Graph Convolution)\n - SSGC (Simple Spectral Graph Convolution)" }, "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 transformation from scalar-to-scalar to set-to-set methodology, building upon the Spectral Former framework. By introducing a learnable parameter W to capture relationships between different frequency domain eigenvalues, the authors aim to enhance model performance through better consideration of inter-frequency domain relationships." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose SSGNN, a simple yet effective spectral-based Graph Neural Network that captures rich spectral information through an adaptive set-to-set filtering approach, offering a more efficient alternative to transformer-based methods. The method introduces a parameter-free Relative Gaussian Amplifier (ReGA) module for robust spectral filtering, and demonstrates superior or comparable performance to state-of-the-art models while using significantly fewer parameters (55x reduction) and computational resources (100x reduction in GFLOPs) across 20 real-world datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- However, the methodology lacks clarity regarding the eigenvalue computation process, which appears to rely on time-consuming SVD operations, raising concerns about computational efficiency.\n\n- Given that your proposed method emphasizes simplicity and reduced learnable parameters, it would be particularly valuable to demonstrate its effectiveness on large-scale graphs. While the reduction in parameter count is noteworthy, the real advantage of a simpler model should be its ability to scale effectively to larger, real-world graph applications. Therefore, I strongly recommend including comprehensive experiments on large-scale graph datasets to validate the method's practical utility. This would not only strengthen your contribution but also clearly differentiate your work from existing methods that may struggle with scalability.\n\n- From a comparative standpoint, although the spectral approach shows promise, the evaluation lacks comprehensive comparisons with important baseline methods, particularly SGC and SSGC. These baselines are especially relevant as they also prioritize simplicity and efficiency. To make your contribution more compelling, consider expanding the experimental section to include: (1) comparisons with these relevant baselines, (2) clear documentation of the eigenvalue computation process and its efficiency, and (3) thorough scalability analysis on large-scale graphs that would demonstrate the practical advantages of your simplified approach. This would help readers better understand the unique benefits of your method in real-world applications where scalability is crucial." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1、\tThe motivation of this paper is significant and valuable for Spectral GNNs.\n2、\tSSGNN achieves significant efficiency in terms of computation and parameters.\n3、\tExperimental results in most cases seem promising." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the improvements of spectral GNNs, which aims to design an expressive filter based on spectral graph theory for effective graph representations. \nThe paper points out that existing SOTA spectral GNNs bring more computational burden, though they can learn the filters better. Thus, the paper proposes a novel efficient framework, namely SSGNN, which only applies simple linear transformation instead of Transformers on the spectrum. Moreover, SSGNN incorporates a parameter-free Relative Gaussian Amplifier to the decoder to enhance adaptive filter learning and maintain stability. The paper conducts extensive experiments on synthetic and real-world datasets to demonstrate the effectiveness of SSGNN." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The novelty of the method is somewhat limited, especially the direct application of existing eigen-correction, eigenvalue encoding, and convolution framework without any transfer challenge.\n2. The description of “set” of “set-to-set filtering” is ambiguous. Specformer applies Transformer on eigenvalue encodings, which enables filters to capture relative dependencies among the eigenvalues. Thus, the “set” of “set-to-set” in Sepcformer means the set of eigenvalues. However, SSGNN learns the spectral filter through linear transformations, so eigenvalues don’t interact with each other. Thus, what’s the meaning of “set”?\n3. The roles of the two linear transformations (namely W_{eig} and W_1W_h) respectively playing in encoder and decoder are not clear. In other words, why do authors include W_1 and W_h in the decoder instead of the encoder?\n4. The paper ignores some essential experiments.\n（1）As mentioned in Line 217, different heads allow the decoder to learn diverse spectral filtering patterns. However, there is no visualization of the diverse spectral filters learned by different heads to verify this conclusion.\n（2）There is no ablation study on the effectiveness of re-center adjustment in Equation 4 and the effectiveness of Relative Gaussian Amplifier in Equation 6.\n（3）Authors don’t verify the stability of SSGNN on OOD benchmarks such as DrugOOD[1], where many model stability studies are validated on this dataset.\n5. The symbols are ambiguous. For example, \\epsilon is used in Line 182, Line 240-241, and Line 307 simultaneously, making the paper more difficult to read. Moreover, it’s not clear on which \\epsilon the ablation experiments reported in Figure 4 are conducted.\n\n[1] Ji, Yuanfeng, et al. \"Drugood: Out-of-distribution dataset curator and benchmark for ai-aided drug discovery–a focus on affinity prediction problems with noise annotations.\" Proceedings of the AAAI Conference on Artificial Intelligence. Vol. 37. No. 7. 2023." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "See weaknesses." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The design of the ReGA component in SSGNN has some novelty. It not only avoids negative eigenvalues but also improves the robustness of SSGNN.\n\n2. This paper conducts extensive experiments and SSGNN also shows competitive performance. For example, in the ZINC dataset, SSGNN has a RMSE value of 0.0592." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a Simple Yet Effective Spectral Graph Neural Network (SSGNN), which simplifies the set-to-set graph filter, e.g., Specformer, without performance degeneration. The key component is a parameter-free Relative Gaussian Amplifier (ReGA), which not only improves model performance but also maintains the robustness against graph perturbations. Extensive experiments on both node-level and graph-level datasets demonstrate the superiority of SSGNN in terms of effectiveness and efficiency over baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. This paper claims that SSGNN uses a **simplified set-to-set approach** to capture key spectral features. However, there is no interaction between different eigenvalues in SSGNN. Specifically, the matrix $Z_{eig} \\in \\mathbb{R}^{N \\times (d+1)}$ indicates the $(d+1)$ dimensional representation of each eigenvalue. The transformation $W_{eig}$ is applied on the channels of a single eigenvalue, i.e., $Z_{eig}W_{eig}$. Therefore, SSGNN is not a set-to-set approach.\n\n2. SSGNN involves a lot of tricks in the training process, such as eigen-correction. However, both Specformer and Polynormer do not use eigen-correction for data preprocessing. In this case, it is necessary to make a comprehensive ablation study to validate the roles of each trick. We need to verify whether the performance improvement mainly comes from ReGA rather than eigen-correction.\n\n3. It would be better if the authors could provide a comparison of the time and space overhead between different methods.\n\n4. How many parameters does SSGNN have in Table 5? Generally, we will control the number of parameters around 50K in the ZINC dataset." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "In the abstract, “Our analysis indicates that applying Transformers to these filters provides minimal advantage in the spectral domain.” While such analysis seems to be missing in the content. Could you elaborate it more?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1-\tThe method is simple and effective on multiple graph downstream tasks.\n\n2-\tThe experiments are comprehensive and solid.\n\n3-\tThe theoretical analysis of ReGA is interesting and novel." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes SSGNN, a simple and effective GNN model which can achieve good performance with much reduced parameters compared to transformers. With basic spectral encoder-decoder structure, it further incorporates a REGA module to strengthen the representational capabilities and the robustness against spectral perturbation. Experiments demonstrate its effectiveness and superiority on model parameters." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1-\tThe time and space cost in pre-computation and training stage of SSGNN seems to be a bottleneck for large graphs. Though top-k techniques can be applied, it’ll lose spectral frequencies which is essential for down-streaming tasks. When comparing parameter amounts and GFLOPS, it’ll be fair to show the training and pre-computation cost for baselines, transformers and other GNNs.\n\n2-\tIt needs experiments to demonstrate the contribution of ReGa, which is the most novel part in SSGNN. Will the removal of it greatly influence the performance?\n\n3-\tPresentation can be improved. “. .” appears in line 373 and line 398 appears incomplete sentences “* means” what?" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Transformer-free spectral GNNs for graph representation learning achieving SOTA performance with significantly fewer parameters and computations." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024ssgnn,\ntitle={{SSGNN}: Simple Yet Effective Spectral Graph Neural Network},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v2nEL42Pvb},\nnote={under review}\n}" }, "abstract": { "value": "Spectral GNNs leverage graph spectral properties to model graph representations but have been less explored due to their computational challenges, especially compared to the more flexible and scalable spatial GNNs, which have seen broader adoption. However, spatial methods cannot fully exploit the rich information in graph spectra. Current Spectral GNNs, relying on fixed-order polynomials, use scalar-to-scalar filters applied uniformly across eigenvalues, failing to capture key spectral shifts and signal propagation dynamics. Though set-to-set filters can capture spectral complexity, methods that employ them frequently rely on Transformers, which add considerable computational burden. Our analysis indicates that applying Transformers to these filters provides minimal advantage in the spectral domain. We demonstrate that effective spectral filtering can be achieved without the need for transformers, offering a more efficient and spectrum-aware alternative. To this end, we propose a $\\textit{Simple Yet Effective Spectral Graph Neural Network}$ (SSGNN), which leverages the graph spectrum to adaptively filter using a simplified set-to-set approach that captures key spectral features. Moreover, we introduce a novel, parameter-free $\\textit{Relative Gaussian Amplifier}$ (ReGA) module, which adaptively learns spectral filtering while maintaining robustness against structural perturbations, ensuring stability. Extensive experiments on 20 real-world graph datasets, spanning both node-level and graph-level tasks along with a synthetic graph dataset, show that SSGNN matches or surpasses the performance of state-of-the-art (SOTA) spectral-based GNNs and graph transformers while using significantly fewer parameters and GFLOPs. Specifically, SSGNN achieves performance comparable to the current SOTA Graph Transformer model, Polynormer, with an average 55x reduction in parameters and 100x reduction in GFLOPs across all datasets. Our code will be made public upon acceptance." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Spectral Graph Neural Networks", "Graph 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/8f99fd24f552579b17bb2a3b3a6ffa5f79f5849e.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": "SSGNN: Simple Yet Effective Spectral Graph Neural Network" }, "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. The paper could be divided into two parts: Gaussian smoothing and the Bregman distance method for the nonsmooth outer problem, which currently feel loosely connected. Could the authors further clarify why the nonsmooth problem necessitates the use of zeroth-order gradient descent?\n2. The convergence relies heavily on sufficiently large values of $B$ and $B'$. What happens if $B$ and $B'$ are of $\\mathcal{O}(1)$? This might provide a fairer comparison with existing methods.\n3. In lines 283 and 285, it appears that Gaussian smoothing is applied twice, converting the second-order gradient to first and then to zeroth order. This may introduce significant errors; could the authors elaborate on how they mitigate this error?\n\nI would be open to reconsidering my grade if all of my concerns are addressed." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper presents a query-efficient bilevel optimization method tailored for nonsmooth stochastic black-box problems, achieving competitive convergence with lower query complexity and outperforming existing methods in data hyper-cleaning and hyperrepresentation learning tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a solution for nonsmooth bilevel optimization using Bregman distance, with a focus on zeroth-order gradient approximation via Gaussian smoothing." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The assumption of Lipschitz continuity in Assumption 2 appears inconsistent with strong convexity; further clarification is needed on why this assumption holds in the given 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": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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 notice that the property of in the Bregman distance used in the convergence analysis is only the strong-convexity. How does the selection of the Bregman distance function impact the performance of the algorithm? Is there a systematic approach to choosing the optimal function for a given problem? What kind of is used in the numerical experiment?\n\n2.The paper applies Gaussian smoothing approximation techniques to replace the gradient updates. Does the algorithm in this paper have any special structure that supports the feasibility of this technique? Can this technique be adopted by other gradient-based algorithms?\n\nTypo：\n\n1. $\\mathcal{G}_t=\\frac{1}{\\alpha} ( x^t-x^{t+1} )$ if you want to get (64), the coefficient of $\\mathcal{B}_{\\Psi_t} ( x, x_t )$ in (63) should be $\\frac{1}{\\alpha}$." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.The algorithm BreZOSBA proposed in this paper combines zeroth-order gradient estimations and Bregman distances to solve the challenging black-box bilevel problem. This idea is both novel and inspiring.\n\n2.BreZOSBA can converge to the stationary point within $\\mathcal{O}\\left(\\frac{d_1\\left(d_1+d_2\\right)^2}{\\epsilon^2}\\right)$ queries, which is outstanding for the zeroth-order gradient method.\n\n3.The experimental results show that the proposed method outperforms several baseline methods, demonstrating its effectiveness in practical applications." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a gradient-free algorithm, named BreZOSBA, for solving stochastic black-box bilevel optimization problems with nonsmooth upper-level and smooth lower level. To deal with the nonsmoothness and the black-box nature of the problem, the algorithm updates the upper and lower-level variables iteratively using stochastic zeroth-order gradient estimates and a Bregman distance-based proximal term. The theoretical analysis demonstrates the query efficiency of the algorithm, showing that it can achieve a certain accuracy level with a finite number of queries. \nThe paper also presents experimental results on four datasets (MNIST, FashionMNIST, Cifar10, and SVHN) to evaluate the performance of the proposed method. The results show that BreZOSBA outperforms several baseline methods in terms of accuracy and convergence speed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The subproblem related to updates of $x$ may be difficult for some Bregman distance, especially when the structure of $h(x)$ is unknown or complex.\n\n2.The theoretical convergence analysis of this method relies on several strong assumptions like strong convexity of the lower-level, which may not be met in most practical applications. This can consequently affect the actual performance of the algorithm.\n\n3.Although Gaussian smoothing can approximate the gradient, the additional errors introduced at each iteration may accumulate. Although it can converge to a solution in expectation, it may still be far from the solution after multiple iterations due to a large variance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "See weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- Theoretical Convergence Guarantees: The convergence analysis looks comprehensive, providing non-asymptotic convergence results that are theoretically solid and highlight the advantage of BreZOSBA in terms of query complexity. \n\n- BreZOSBA’s use of a single-loop ZO framework with Bregman distance introduces computational savings compared to double-loop structures." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a query-efficient algorithm, BreZOSBA, designed to solve nonsmooth stochastic black-box bilevel optimization (BO) problems by leveraging Bregman distance and Gaussian smoothing. By adopting a single-loop, zeroth-order (ZO) framework, the authors claim improved query efficiency, theoretically achieving $ O(d_1(d_1 + d_2)^2 \\epsilon^{-2})$ query complexity to reach an $ \\epsilon$ -stationary point. The paper validates BreZOSBA on two small scale applications—data hyper-cleaning and hyper-representation learning—and compares it to two baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The practical applicability of bilevel optimization with both inner and outer levels as black-boxes was largely not motivated by the authors. While some specific bilevel problems can have partial black-box structure, many practical applications in machine learning (such as hyperparameter optimization, meta-learning, etc) involve inner-level variables that represent the parameters of deep neural networks. In such cases, using zeroth-order methods to solve the inner problem would be extremely slow and may not even converge. \n\n- The choice to employ ZO techniques to solve both the inner and outer levels restricts the scalability of the approach, which may explain why the experimental validation was limited to small-scale toy tasks. Thus, the method’s suitability for real-world BO problems remains questionable, especially in deep learning context. \n\n- The experiments are very limited with only two compared baselines (that are not particularly strong for these experiments). In fact, HOZOG [1] was introduced especially for hyperparameter optimization (where $x$ are the hyperparameters and $y$ is the parameters of possibly a deep neural network), in which it can be a strong baseline. However, the authors used it here in settings for which HOZOG is not the most adequate baseline, which I believe is unfair and undermines the validity of the results. \n\n- The literature review on bilevel optimization is sparse and does not adequately cover recent, relevant works. For instance, PZOBO [2], another hessian-free ZO method, is neither discussed nor compared against, despite its direct relevance. \nThe paper also lacks a discussion on extensions or limitations of the proposed approach in the broader context of BO, such as bilevel problems without lower-level singleton constraint. This would be essential to clarify where the proposed method fits within the current landscape of BO approaches and where it may fall short or require adaptation. \n\nOther minor issues: \n\n- The inner objective $g$ is smooth and this should be explicitly said after the problem definition, just like the authors did it for the outer objective function $f$. \n\n- In the text, the authors keep mentioning that using ZO gradient for the inner level problem is not the most efficient thing, but their algorithm did exactly use that (equation 14). \n\nReferences: \n\n[1] Bin Gu, Guodong Liu, Yanfu Zhang, Xiang Geng, and Heng Huang. Optimizing large-scale\nhyperparameters via automated learning algorithm. arXiv preprint arXiv:2102.09026, 2021. \n\n[2] On the convergence theory for hessian-free bilevel algorithms. D Sow, K Ji, Y Liang - Advances in Neural Information Processing Systems (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": 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": "What is the benefit of using the Bregman distance in the proposed algorithm?" }, "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 chooses a good perspective to study BLO methods, where there is still some gaps in the complexity theory.\n\n2.The paper improves the theoretical complexity bound to an almost optimal one." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper considers nonsmooth black box Bi-level optimization. The lower problem is assumed to be strongly convex and have Lipschitz continuous gradients and hessians, and the upper level function can be nonconvex. Under this setting, the paper improves the query complexity of zero order method for finding an approximately stationary point of BLO. Numerical experiments are conducted to show the competitive performance of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The strong convexity assumption has been extensively studied for BLO. The authors should lay importance on the difference of the used techniques for establishing complexity bound for black box BLO.\n\n2.The stationarity measure $\\|G^t\\|$ lack explanation. It is important for the authors to justify that the used measure is equivalent to the compared methods." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a query efficient method for nonsmooth stochastic black-box bilevel optimization." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024query,\ntitle={Query Efficient Nonsmooth Stochastic Black-Box Bilevel Optimization with Bregman Distance},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v2uPdQDwSz},\nnote={under review}\n}" }, "abstract": { "value": "Bilevel optimization (BO) has recently gained significant attention in various machine learning applications due to its ability to model the hierarchical structures inherent in these problems. Several gradient-free methods have been proposed to address stochastic black-box bilevel optimization problems, where the gradients of both the upper and lower-level objective functions are unavailable. However, these methods suffer from high query complexity and do not accommodate more general bilevel problems involving nonsmooth regularization. In this paper, we present a query-efficient method that effectively leverages Bregman distance to solve nonsmooth stochastic black-box bilevel optimization problems. More importantly, we provide a non-asymptotic convergence analysis, showing that our method requires only $\\mathcal{O}({d_1(d_1+d_2)^2}{\\epsilon^{-2}})$ queries to reach the $\\epsilon$-stationary point. Additionally, we conduct experiments on data hyper-cleaning and hyper-representation learning tasks, demonstrating that our algorithms outperform existing bilevel optimization 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": [ "zeroth-order gradient", "bilevel 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/9b61c2968b66a5459900496a3c077fb9c6b49c9c.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/ea98fd95487c994f55cfdff499f80857424942b0.zip" }, "title": { "value": "Query Efficient Nonsmooth Stochastic Black-Box Bilevel Optimization with Bregman Distance" }, "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": "The questions here are related to the three points described as 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 proposes a novel approach to estimating blurriness by utilizing the gradient of metric respect to kernel parameters, which serves as a guide for the sampling of x_ {t-1}.\n2. The paper introduces an auto-scale gradient guidance strategy that automatically calculates the guidance scale corresponding to different forecast time periods, models, and datasets. This strategy enables the model to adaptively denoise the prediction results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed a universal post-processing method, PostCast, designed for precipitation nowcasting. This method achieves denoising of convolution based prediction models using unconditinal DDPM through two key innovative points: zero shot blur estimation mechanism and auto scale gradient guidance strategy. The model was trained on a combination of five different datasets and prediction results generated by different models, and outperforms other conditional denoising models in terms of CSI metrics." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. CasCast and DiffCast are both designed for **predicting precipitation**, but not **postprocessing tasks**. It is crucial for the authors to compare with other generative models such as GAN. This would provide a more comprehensive evaluation of PostCast's performance against a wider range of methodologies.\n2. The authors have compared PostCast with CasCast and DiffCast only in **out-of-distribution datasets** but did not compare them on HKO7, SEVIR, etc.\n3. Relying on CSI as the only metric may not sufficiently judge the quality of precipitation prediction. The average intensity of the precipitation predictions can significantly impact CSI. For instance, using simple histogram matching on predictions can also significantly improve CSI. The authors should clarify whether the CSI improvement is due to **increased intensity** from the model or other factors to provide a more accurate assessment of the model's performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. **Evaluation Metrics**: Could you add results for POD, FAR, and image quality metrics like PSNR and SSIM in the appendix to provide a more comprehensive evaluation of PostCast?\n\n2. **Related Work**: Could you expand the related work section to discuss GAN and Transformer methods in precipitation nowcasting to better contextualize the background and contributions of PostCast?\n\n3. **Comparison with GAN Methods**: Could you include comparisons with GAN-based methods in precipitation nowcasting to further validate PostCast's effectiveness?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. **Originality**: The primary innovation of PostCast lies in employing an unconditional diffusion model (DDPM) to remove blurriness from precipitation predictions without relying on paired data. By introducing zero-shot blur kernel estimation and an auto-scale denoise guidance strategy, PostCast adapts to various datasets, prediction lead times, and blur modes across different models. This unsupervised deblurring approach appears to be the first of its kind in the domain of precipitation nowcasting, demonstrating creative thinking.\n\n2. **Completeness of Experiments**: The paper provides a well-designed experimental setup, covering multiple datasets and forecast models to showcase the generalizability and robustness of PostCast. The experiments include diverse evaluation metrics and extensive visual results, effectively validating the method’s performance in extreme precipitation event prediction." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel postprocessing approach for precipitation nowcasting, named PostCast. The core innovation of PostCast lies in utilizing an unconditional diffusion model (DDPM) to remove blurriness from precipitation predictions without requiring paired data of real observations and blurry predictions. The method introduces a zero-shot blur kernel estimation mechanism and an auto-scale denoise guidance strategy, enabling the model to adapt to various datasets, prediction lead times, and blur modes to generate sharper predictions. The authors conducted experiments on multiple precipitation datasets and forecast models, demonstrating PostCast’s effectiveness in improving prediction accuracy for extreme precipitation events. The paper proposes a highly generalizable and adaptive postprocessing framework that shows strong potential for broad applications in precipitation nowcasting tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Incomplete Evaluation Metrics**: The paper relies only on CSI, lacking POD, FAR, and similar precipitation metrics, as well as image quality metrics like PSNR and SSIM. Including these in the appendix would strengthen the evaluation’s comprehensiveness.\n\n2. **Insufficient Related Work Discussion**: The discussion of GAN methods in precipitation nowcasting is limited, and the introduction to Transformer-based approaches is also lacking. Expanding these sections would better contextualize PostCast’s contributions.\n\n3. **Lack of Comparison with GAN Methods**: There is no direct comparison with GAN-based approaches in precipitation nowcasting. Adding such comparisons would provide a more complete assessment of PostCast’s 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": 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 current metric/score mainly evaluates whether the small-scale features are recovered and extreme events are captured well after deblurring. However, isn’t it also important to assess how closely the nowcasting outputs, after deblurring with this methodology, match the ground truth, generally? If we don’t threshold the pixel values in this case and compute an error metric such as MSE and PSNR, that could also be informative. The authors can consider adding these scores." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The problem is well-motivated. An unsupervised method alleviates the need for labels and the model is no longer limited to the blur modes seen during training. Given that this framework is independent of the spatiotemporal prediction model and can be added on top of any prediction model, it makes this work very useful and widely applicable. \n- The results strongly demonstrate the ability of their proposed method. The experiments are extensive and prove PostCast’s effectiveness in recovering weather patterns across multiple prediction models, datasets, and lead times." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes an unsupervised postprocessing method for precipitation nowcasting outputs to remove blurriness from these predictions, especially at longer lead times. This can be important to obtain accurate extreme precipitation events. The blurriness is eliminated with an unconditioned denoising diffusion probabilistic model guided by blurry predictions. The authors introduce a zero-shot kernel estimation method and an auto-scale denoise guidance strategy and show that their proposed framework generalizes to many blurriness modes across different datasets and varying lead times." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The \"method\" section could be more organized, creating subsections could help. Moreover, the “zero-shot kernel estimation mechanism” and “auto-scale denoise guidance strategy”, should be explained in more detail, as these are the main contributions this paper makes. \n- The authors should justify using \"CSI\" (Critical Success Index) scores as the only evaluation metric. Adding another metric such as “Recall” (True positive rate) can complement the current evaluation and make it more comprehensive." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024postcast,\ntitle={PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v2zcCDYMok},\nnote={under review}\n}" }, "abstract": { "value": "Precipitation nowcasting plays a pivotal role in socioeconomic sectors, especially in severe convective weather warnings. Although notable progress has been achieved by approaches mining the spatiotemporal correlations with deep learning, these methods still suffer severe blurriness as the lead time increases, which hampers accurate predictions for extreme precipitation. To alleviate blurriness, researchers explore generative methods conditioned on blurry predictions. However, the pairs of blurry predictions and corresponding ground truth need to be given in advance, making the training pipeline cumbersome and limiting the generality of generative models within blurry modes that appear in training data. By rethinking the blurriness in precipitation nowcasting as a blur kernel acting on predictions, we propose an unsupervised postprocessing method to eliminate the blurriness without the requirement of training with the pairs of blurry predictions and corresponding ground truth. Specifically, we utilize blurry predictions to guide the generation process of a pre-trained unconditional denoising diffusion probabilistic model (DDPM) to obtain high-fidelity predictions with eliminated blurriness. A zero-shot blur kernel estimation mechanism and an auto-scale denoise guidance strategy are introduced to adapt the unconditional DDPM to any blurriness modes varying from datasets and lead times in precipitation nowcasting. Extensive experiments are conducted on 7 precipitation radar datasets, demonstrating the generality and superiority of our method." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "AI for Science; Precipitation Nowcasting; Diffusion Model; Zero-shot Blurriness Kernel; Auto-scale Denoise 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/4626fa65af0515195be96e1ad39b36336cfaa023.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "PostCast: Generalizable Postprocessing for Precipitation Nowcasting via Unsupervised Blurriness Modeling" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 minor and major questions:\n1. Abstract: Concrete examples would be better, such as which model did it beat despite being smaller etc..\n2. L196: Please give examples, what happens for various ways of writing floats. How are the European and US/UK numbers treated 1,43 vs 1.43. Mixed numbers and digits, other mathematical symbols. To me, its not so clear from the writing.\n3. L210: The architecture description seems incomplete, given its a section. You have mentioned decoders elsewhere, but you should complete this here, mentioning how many layers of decoders (or ranges), how many dense layers, and some block diagrams, referred from the section. This is supposed to be the most important section.\n4. L249: Whats the perplexity of other models, especially mathematics specialist or science specialist ones? Can you show a table comparing them? Otherwise, the standalone numbers do not make sense to me." }, "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. I appreciate the background and the description of how the model is trained. The idea of targeting mathematics is important and building LLMs specializing in math (at least some part of it) is important. \n\nThe dataset is an important contribution, however I am not sure whether the authors plan to make it public." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Authors present a LLM specializing in mathematics, called Paramanu-Ganita. It is quite smaller in size and exhibits interesting performance benefits in several math and logical datasets, compared some LLMs with bigger size. Authors also trained tokenizers from scratch, curate a new dataset for pretraining and show that Paramanu-Ganita outperforms several general-purpose LLMs and some domain-specialist LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I feel the paper explores an interesting direction, but there are some concerns:\n\n1. Firstly, GSM8k tests basic math word problem skills and given the model's GSM8k performance is pretty poor, I do not feel the model is ready yet. I think more experimentation is required. Also, how are Table 2 values computed? It seems the MetaMath paper reports GSM8K performance to be 82.3. Why is it 66.5 here? [1]\n\n2. What is mostly missing from the paper are proper motivations and justification as to what \"contributes\" or what is expected to contribute to the \"improved\" performance? \n\n - Looking at this from a different point of view, why did the authors not start with MetaMath, then say change the tokenizers or change the dataset? Then, slowly demonstrate how all the innovations are truly necessary. At the least such ablations would have showed the necessity of new models. \n\n - Secondly, given the model's performance is not so great, what are we gaining by spending so much training time and cost?\n\n3. One more important aspect is, what are the domains that the model targets? What are the grade levels? Is it the expectation that we will also do IMO problems starting from GSM8k? Or, are we targeting sub-disciplines algebra, pre-algebra, calculus etc.? I think this depth is also missing, so is related papers that investigate the need of such models [2].\n\n[1] https://openreview.net/forum?id=N8N0hgNDRt\n[2] MATHSENSEI: A Tool-Augmented Large Language Model for Mathematical Reasoning" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "- Good empirical results despite smaller model size, demonstrating the effectiveness of their approach\n- Demonstrates that smaller, more efficient models can achieve good mathematical reasoning performance" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces PARAMANU-GANITA, a 208 million parameter mathematics-focused language model trained from scratch. The authors demonstrate that effective mathematical reasoning capabilities can be achieved with smaller, more efficient models when trained specifically for the domain. This approach offers significant advantages in terms of computational costs and environmental impact while maintaining good performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The overall presentation of the paper still needs much improvement. The paper is not in ready-to-review or ready-to-submit status. The figures are pretty rough and unclear for what the authors want to express. For example, Figure 2 shows GPU Power Usage during pretraining of Paramanu-Ganita. But what conclusions do the authors want to make here? How does it illustrate the environment friendly nature of the model? For the figure 1, what does the blue line mean here?\n- Limited Ablation Studies. The paper doesn't analyze the relative importance of different components of their training data (web text vs. code vs. lecture notes). It is unclear why the authors want to utilize these data sources and why the data mixture should be adopted as it is in the paper.\n- Contamination issues. The model achieves good performance on GSM8K and MATH with 200M parameters. It is unclear whether there is data contamination issue." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "Address the weaknesses of the paper mentioned above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. A novel decoder model, that is 34 times smaller than existing LLMs and can outperform them by a huge margin\n2. A detailed explanation of the training process required\n3. Detailed benchmarking on GSM8K, MATH and other datasets.\n4. Emphasis on the training time required and compared it to other existing LLMs, showing computation and environmental prowess in training an exclusive tiny model from scratch." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The author presents a small decoder-based language model on mathematics called Paramanu-ganita. They trained this model from scratch using the existing public mathematical corpus and also performed CoT instruction finetuning on top of it. They also train their own tokenizer specialised in math and code. Despite their model only having 208 million parameters, it outperforms general LLMs by approximately 30% points, and even math-specialised LLMs by 3-23% points in GSM8K, 6-8% on MATH. The 208 million parameter model outperformed LLaMa-1 (33B, 13B, 7B), LLaMa-2 (7B, 13B), Falcon (40B, 7B), PaLM (62B, 8B), MPT (30B, 7B), Vicuna 13B, and math-specialised LLMs like Minerva 8B, LLEMMA-7B on GSM8K, MATH, AGIEVAL-AQuA-RAT benchmarks. They also showed the reduced time and computation requirement to train this model as compared to existing LLMs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper uses Qwen-72B to label the corpus and use a score >= 0.6 for training the model, ensuring only a high-quality dataset is used. However, apart from this, the training process used is not novel. Specifically, there is no novelty in the model architecture or training paradigm used that can justify the complete novelty of the paper and also puts into question the improved performance of a 208 million parameter model over LLMs\n2. The paper does not touch upon, newer and difficult mathematical datasets such as MATHBENCH or JEEBENCH. These are some datasets that were released after training cutoff time for some models, ensuring they are not part of their training data. These datasets are also much more difficult as compared to gms8k. This will ensure that the proposed model is robust in solving difficult problems that it hasn't seen before.\n3. Will the checkpoint-filtered corpus used for training be publicly available?\n4. How does the model perform on out-of-distribution data points, this can be checked by first doing a sanity check of data memorization/contamination [1]. Performing simple algorithms 1 and 2 from the paper will ensure that the model has not seen the evaluation dataset, making the results more robust.\n5. The empirical analysis is missing from the paper. A thorough qualitative comparison of reasoning chains produced by Paramanu-Ganita versus other models on a few representative problems from the benchmark datasets. For example, what errors are made by existing LLMs vs. Paramanu-ganita and in which area does it improve?\n\nReference\n[1] Golchin, Shahriar, and Mihai Surdeanu. \"Time travel in llms: Tracing data contamination in large language models.\" arXiv preprint arXiv:2308.08493 (2023)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "An efficient language model for mathematics, which is pretrained from scratch on a custom corpus on 31.5 billion tokens; despite having only 208 million parameters, it outperforms several large and very large LLMs on standard benchmarks" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024paramanuganita,\ntitle={Paramanu-Ganita: An Efficient Pre-trained Generative Mathematics Language Model with Chain-of-Thought Instruction Fine-Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v3DwQlyGbv},\nnote={under review}\n}" }, "abstract": { "value": "In this paper, we present PARAMANU -G ANITA, a 208 million-parameter novel\nAuto Regressive (AR) decoder based language model on mathematics. We per-\nformed pretraining from scratch on 31.5 billion tokens using a context size of 4096\non a mixed mathematical corpus consisting of mathematical web pages, mathe-\nmatics related source code such as AlgebraStack, mathematical textbooks, Chain-\nof-Thought (CoT) templatised mathematical StackOverflow question answers\npairs, and mathematical lecture notes in LATEX curated by us. We also trained a\nmath and code specialised BPE tokenizer. We proposed and performed Chain-of-\nThought instruction fine-tuning of Paramanu-Ganita on the MetaMathQA dataset.\nWe evaluate our model on GSM8K and MATH mathematical benchmarks, and on\nlogical deductive reasoning (LogiQA) and multiple choice high school and col-\nlege level math questions from SAT (AGIEVAL-SAT-Math), GRE/GMAT ques-\ntions (AGIEVAL-AQuA-RAT), college and high school level math questions from\nMMLU. Our model Paramanu-Ganita, despite being 34 times smaller than the\n7B LLMs, outperforms general LLMs by approximately 30% points, and even\nmath-specialised LLMs by 3-23% points in GSM8K test accuracy metric. On\nMATH benchmark, Paramanu-Ganita outperformed the various models by 6-8%\npoints. On other benchmarks such as LogiQA logical deductive reasoning bench-\nmark, mathematical high school level multi-choice questions (MMLU-math-high-\nschool), GRE-GMAT level quantitative questions (AGIEVAL-AQuA-RAT), SAT\nlevel math questions, Paramanu-Ganita was better than the others by about 1-4%\npoints. The large significant margin improvement in performance of our math\nmodel over the existing LLMs signifies that reasoning capabilities of language\nmodels are just not restricted to those with humongous number of parameters.\nParamanu-Ganita took only 170 hours of A100 training whereas large LLMs such\nas the math-specialised LLM, LLEMMA 7B, was trained for 23,000 A100 equiv-\nalent hours. Thus, our approach of pretraining powerful domain-specialised lan-\nguage models from scratch for domain adaptation is much more cost-effective and\nenvironmental friendly than performing continual training of LLMs." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "reasoning", "language models", "pretraining", "CoT fine-tuning", "AI4Math" ] }, "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/28346be59c5d9b389bc43395e3522fd1ca72b143.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": "Paramanu-Ganita: An Efficient Pre-trained Generative Mathematics Language Model with Chain-of-Thought Instruction Fine-Tuning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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 definition of the connectivity loss in Eq. (7) requires an integration over $\\alpha$ which I believe is not possible to implement practically. The authors should comment on how they approximate this loss in practice.\n\n2. What is the neural network used in Figure 2?\n\n3. Can we extend the definition of group connectivity to work with general weights instead of fixing the weights as $1/K$ for each of the group models?\n\n4. For overparameterized models (every global optimum is also a local optimum), it appears that the value of $\\Gamma$ in Definition 3.7 would be zero, implying no effect of heterogeneity on the bound in Theorem 3.8. Can the authors comment on this?\n\n**Typos**\n1. There appears to be a typo in Line 95 in the statement \"refer to each client's distribution $D_i$\".\n2. It should be $D^t$ and not $D$ in Line 113." }, "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-structured and easy to read and intuitively the idea of using anchor models to improve LMC cross the client local models makes sense.\n\n2. Experimental results look promising and show that the proposed FedGuCci and FedGuCci+ can outperform vanilla FedAvg and other baselines across a wide range of settings.\n\n3. Ablation studies are provided showing that the proposed algorithms are generally robust and require lower computation cost to reach a target accuracy compared to FedAvg and other baselines." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposed FedGuCci, an algorithm to improve merging of local client models in federated learning by ensuring that the client models are in the same loss basin, motivated by findings in the linear mode connectivity (LMC) literature. To do so, authors propose to add a connectivity loss to the standard client objective which tries to ensure that the client's local model is mode connected to an anchor model. The authors theoretically motivate their approach by proving the transitivity property of LMC when merging two layer neural networks. This is followed by experiments on practical FL training tasks, which shows that the proposed FedGuCci and it's extension FedGuCci+ can outperform FedAvg and other baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Lemma 3.3 statement and implication** \n\n* Firstly, I don't understand what is random about $D_\\\\epsilon(w_{anc}^*)$ defined in Lemma 3.3. If $w_{anc}^*$ is a deterministic anchor model (as defined later in Theorem 3.5 and Theorem 3.8) then $D_\\\\epsilon(w_{anc}^*)$ is also deterministic so the bound on the probability of $D_\\\\epsilon(w_{anc}^*)$ does not make sense to me. Intuitively, I believe Lemma 3.3 is trying to bound the probability of LMC between a deterministic anchor model $w_{anc}^*$ and a random model $w$ such that $||w-w_{anc}^*|| \\leq d/2$. So the probability here is over some uniform distribution of $w$ and not $w_{anc}^*$. If so, Lemma 3.3 should be re-written to clarify this.\n\n* There is no understanding of how large $d_{\\epsilon}$ can be. For instance if $d_{\\epsilon} \\geq d$ then the probability bound is just vacuous. Therefore authors need to either assume with some justification or prove that $d_\\epsilon < d$ for the bound to make sense.\n\n2. **Theorem 3.5 and 3.8 statement and implications.**\n\n* It is not clear to me why we need randomness in these theorem statements. Suppose we are given three models $w_{anc}, w_1, w_2$ such that i) $w_{anc}$ and $w_1$ are LMC ii) $w_{anc}$ and $w_2$ are LMC and iii)$||w_{anc} - w_1|| \\leq d$ and $||w_{anc} - w_2|| \\leq d$. Why can't we use just these assumptions to bound the loss barrier between $w_1$ and $w_2$? In other words why do we need the additional assumption that $w_1$ and $w_2$ are sampled from the uniform distribution?\n\n* Following up on my point in Weakness 1), authors need to show that $d_\\epsilon$ is bounded for the bound in Eq. (6) and Eq. (10) to not be vacuous. In addition, authors also need to show that $\\delta > 0$ that this there is a non-zero probability that $w_1$ and $w_2$ are sampled from the uniform distribution and are also LMC with $w_{anc}$. \n\n* Currently I don't see a dependence on $d$ in the bounds in Eq. (6) and Eq. (10) which is a bit surprising to me since $d$ is defined in the Theorem statement. \n\n3. **Inconsistent experimental results and performance of baselines**\n\n* It appears to me that the authors have either not implemented FedProx correctly or have not tuned the regularization parameter $\\mu$ in FedProx correctly. If we set $\\mu \\rightarrow 0$ then FedProx just becomes same as FedAvg so the performance of FedProx should at least be as good as FedAvg. Therefore it is surprising to see the consistently poor performance of FedProx across all the experiments.\n\n* In Table 2, for the Tiny ImageNet column with Non-IID hyper. being 100, the performance of SCAFFOLD seems inconsistent with previous results. \n\n* In Table 3, for the STS-B column, the performance of FedDyn seems inconsistent with previous results.\n\n* Authors should provided information on how they tuned hyperparameters for the experiments and also provide graphs showing test accuracy vs number of rounds for all the baselines. Currently only the final accuracy numbers is reported for each of the experiments \n\n* Why is FedDyn and SCAFFOLD not compared against in Table 4 and Table 5?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Q1: In Section 2.3, the way your method leverages the global model is different from FedProx, but it’s unclear what specific advantages your approach offers. While your method integrates the connectivity loss and replays N historical global models, FedProx only uses the current global model for regularization. A more detailed explanation of why your method is superior would be helpful. \n\nQ2: The proposed algorithms do not have theoretical convergence guarantees. Many of the methods you compare against do provide convergence analyses, and including such an analysis would be essential for the theoretical foundation of your work.\n\nQ3: When discussing group connectivity varying K, the range of client numbers is too small to reflect the truth in practice. The left plot in Figure 4 may not support the author’s conclusion that “the increase of barriers may converge to a point lower than vanilla training”. If it finds the transitivity of group connectivity may be weakened for larger K, an interpretation of the reason should be provided.\n\nQ4. If FedGuCci+ is designed to enhance FedGuCci by incorporating additional techniques, it’s important to evaluate the individual contribution of each technique to the overall improvement, like conducting ablation studies." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper adapts a connectivity perspective to use LMC improving the connectivity among the local models. It shall be a novel tool in the analysis of FL." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "FedGuCci(+) enhances federated learning generalization by improving group connectivity, inspired by linear mode connectivity, to better fuse local models and achieve stronger performance across diverse tasks and architectures." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1. The comparison with prior work is limited\n\nW2. The proposed algorithms, including FedGuCci, do not have theoretical convergence guarantees.\n\nW3. The comparisons and conclusions may not have sufficient evidence. \n\nW4. The evaluation needs to be enhanced." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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.\tCan authors provide insights into storage challenges in very large-scale scenarios? Furthermore, I would be happy if I could see 1-2 experiments on large clients (M > 500) or (M > 1000) if possible.\n2.\tTheoretically, can authors intuitively explain how the model behaves and challenges when applied to models like GPT-3? Is there a way to expand this for such large models in the near future?\n3.\tCould the authors elaborate more on how the method would behave in low-bandwidth, high-latency federated learning environments? Would there be trade-offs in performance?\n4. Please respond to the weaknesses above." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1)\tThe paper provides a creative adaptation of LMC to FL, which is an underutilized idea in the FL domain.\n2)\tWell written paper and easy to follow.\n3)\tSound theoretical foundations with supporting proofs to enhance the credibility." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an interesting concept in FL often underexplored. Specifically, this paper leverages the concept of group connectivity, drawn from linear mode connectivity (LMC), to better fuse local models in parameter regions into a global generalized model, which could come with many benefits. The issue of model drift and heterogeneity seems to be tackled well. The authors validate the methods on vision and NLP datasets. Overall, the idea looks interesting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am particularly concerned with communication costs and computational overhead of the method when applied to large-scale FL scenarios. Below are few weaknesses:\n\n1) Regarding the communication costs, I agree that there are no additional costs compared to FedAvg, but this remains an issue. There have been many advancements in communication efficiency over FedAvg in recent years. To mention a few recent works, DisPFL[1] and SSFL[2] have demonstrated better performance than FedAvg, even at very high sparsity levels. I believe the baselines are insufficient and suggest comparing the method with at least these sparse baselines on a few vision datasets to evaluate whether the performance gains justify the communication cost to validate its worth.\n\n2) The authors did not consider incorporating techniques to prune or compress unimportant weights, which could reduce communication overhead without sacrificing performance. While the core idea of the paper is appreciated, there is room for optimization and further improvements in terms of communication efficiency.\n\n2) Although the idea is novel, it’s a lot of computations and could pose practical challenge in large-scale FL applications. Authors should think of addressing this in the future works so that this method can be practically applied.\n\nReferences:\n\n[1] https://doi.org/10.48550/arXiv.2206.00187\n\n[2] https://doi.org/10.48550/arXiv.2405.09037" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nli2024improving,\ntitle={Improving Group Connectivity for Generalization of Federated Deep Learning},\nauthor={Zexi Li and Jie Lin and Zhiqi Li and Didi Zhu and Tao Shen and Tao Lin and Chao Wu},\nyear={2024},\nurl={https://openreview.net/forum?id=v3W9tdTGx5}\n}" }, "abstract": { "value": "Federated learning (FL) involves multiple heterogeneous clients collaboratively training a global model via iterative local updates and model fusion. The generalization of FL's global model has a large gap compared with centralized training, which is its bottleneck for broader applications. In this paper, we study and improve FL's generalization through a fundamental \"connectivity'' perspective, which means how the local models are connected in the parameter region and fused into a generalized global model. The term \"connectivity'' is derived from linear mode connectivity (LMC), studying the interpolated loss landscape of two different solutions (e.g., modes) of neural networks. Bridging the gap between LMC and FL, in this paper, we leverage fixed anchor models to empirically and theoretically study the transitivity property of connectivity from two models (LMC) to a group of models (model fusion in FL). Based on the findings, we propose FedGuCci(+), improving group connectivity for better generalization. It is shown that our methods can boost the generalization of FL under client heterogeneity across various tasks (4 CV datasets and 6 NLP datasets) and model architectures (e.g., ViTs and PLMs)." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Zexi_Li1", "~Jie_Lin7", "~Zhiqi_Li4", "~Didi_Zhu1", "~Tao_Shen4", "~Tao_Lin1", "~Chao_Wu1" ] }, "authors": { "value": [ "Zexi Li", "Jie Lin", "Zhiqi Li", "Didi Zhu", "Tao Shen", "Tao Lin", "Chao Wu" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Deep learning", "federated learning", "generalization" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "li|improving_group_connectivity_for_generalization_of_federated_deep_learning" }, "pdf": { "value": "/pdf/e1e45b4f520c317aa6d6dc06e9187fd3efe19db9.pdf" }, "presentation": null, "primary_area": { "value": "transfer learning, meta learning, and lifelong learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/33a2b677607abdfeed4fb37ef180a282d6b3e487.zip" }, "title": { "value": "Improving Group Connectivity for Generalization of Federated Deep Learning" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "N/A" }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "The paper's application area is an important problem of long-sequence reconstruction, which could enable the capture of essential clinical information in the latent space." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper aims to build a VAE-based model for long ECG segments and considers a benchmark ECG dataset for the analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper's presentation is poor and does not meet the standards of ICLR or other relevant AI/ML venues. The authors are advised to proofread the paper carefully, as there are numerous grammatical errors, including missing commas, throughout the text, including in the abstract. Many of these errors could be easily fixed by using available grammar checkers, suggesting that the manuscript may not yet be ready for submission.\n2. The proposed method lacks novelty, as it primarily focuses on splitting signals in the input space.\n3. The results are not well-presented, making it unclear what the main contributions of this work are." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "# Suggestions\n\n- Lack of Quantitative Metrics for Reconstruction Quality: --> Idea: Including quantitative reconstruction metrics (e.g., Mean Squared Error, Structural Similarity Index) would make it easier to assess the improvement over standard VAE architectures objectively.\n- Suboptimal Sleep Stage Classification Results: --> Idea: Exploring methods to enhance inter-segment information sharing could improve performance. Incorporating recurrent layers (RNN, LSTM) in the Parameterizer module could capture temporal dependencies between ECG segments, potentially boosting classification accuracy.\n- Over-Reliance on Folding Technique Without Comparison to Alternative Approaches: --> Idea: Including a baseline comparison with alternative architectures for long-sequence encoding would provide a clearer picture of the folding method’s relative effectiveness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Comprehensive Experimentation: The paper includes experiments with multiple datasets (MESA and MIT-BIH), demonstrating the generalizability of the proposed architecture across different signal sources and classification contexts.\n- Practical Applications: The application of this model to sleep stage classification is valuable and opens doors for future research on ECG-based sleep monitoring systems, which has relevance in healthcare." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a folded VAE architecture designed to improve the reconstruction quality of long ECG segments, aiming to retain interpretable features for downstream tasks like sleep stage classification. The authors present a CNN-based VAE architecture that encodes ECG signals by folding long segments into smaller sub-segments and processing them through a shared backbone encoder and decoder. The manuscript explores this architecture's reconstruction quality, highlights the challenges of using VAEs for long physiological signals, and evaluates the model's performance on two ECG datasets. Additionally, the folded-VAE framework's latent space representation is leveraged for sleep stage classification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack of Quantitative Metrics for Reconstruction Quality: The reconstruction results rely on visual analysis without quantitative metrics, which may lead to subjective conclusions.\n- Suboptimal Sleep Stage Classification Results: The model’s classification accuracy (mean of 65% across subjects) is lower than other baseline methods.\n- Over-Reliance on Folding Technique Without Comparison to Alternative Approaches: While folding offers improved reconstruction, the authors do not compare this method against alternative architectures (e.g., hierarchical or multi-scale VAEs, or LSTM-based methods) that could also encode long signals effectively. \n- Clarity and Language: Some sections, particularly in the methodology, could benefit from improved clarity. For instance, equations used to describe the encoding and decoding of ECG segments could be elaborated on to better illustrate the folding 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": 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. To more clearly demonstrate the advantages of the presented method, please explicitly compare the folding approach to a sliding window baseline, both conceptually and empirically. \n\n2. Please clarify the questions raised in bullet 2 in the weakness section. Furthermore, please provide a step-by-step explanation of how the folding and merging operations are implemented, including the mathematical justification for each step.\n\n3. Please provide a diagram or explicit description of how the VAE from section 2.4 is integrated into the overall architecture described in section 2.8. This would help clarify the relationship between these components." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The problem of reconstructing long ECG sequence is interesting.\n\n2. The experiments were conducted on 2 datasets and consider not only reconstruction but also classification tasks for performance evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a VAE for reconstructing long (10-30s) ECG sequences. This is achieved with a so-called folding scheme to process short sequences during the encoding and then combine the folds before a decoder mirroring a similar architecture. Experiments are conducted on sleep ECG from two different datasets, and the benefit of the approach is investigated in both reconstruction and its effect on downstream sleep classification." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The premise of the study can be made more clear. Currently It is not clear what is the benefit of the presented method against simply using a sliding window on a long ECG sequence by reconstructing short sequences across the windows\n\n2. Some of the technical components can be better described. For instance, it is not clear what the summation means in equations (1-2) — does it really mean summation by averaging over the representation obtained from different folds? If yes, why? Similarly, it was not well justified why we can merge 30 8x4 features into 8x120 future map.\n\n3. The relation between the VAE described in 2.4 and the specific model described in 2.8 is not clear.\n\n4. There were baselines or comparative studies conducted evaluating the presented method with existing approaches to handle this, especially by simply learning to reconstruct for short sequences and apply to long sequences with a sliding window." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "see above" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The question of whether variational auto-encoders can generate a flexible continuous latent space for long electrocardiogram (ECG) segments and reconstruct the input is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work explores the use of a folded Variational Autoencoder (VAE) architecture to reconstruct long electrocardiogram (ECG) signals. A folded VAE architecture is proposed to address the limitations of traditional VAEs in handling long ECG sequences. The proposed method involves splitting long ECG segments into smaller folds, processing them sequentially, and concatenating them for reconstruction, within a VAE framework. The authors evaluate the proposed method's performance in a sleep stage classification task using two datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-\tRelated work section is missing completely. Can the authors comment on what work is out there that investigated similar problems as they do? How do for example the following works relate to the work proposed here (just to name a few):\n1. Comparison of Autoencoder Encodings for ECG Representation in Downstream Prediction Tasks. Christopher J. Harvey, Sumaiya Shomaji, Zijun Yao, Member, IEEE, Amit Noheria , 2024\n2. Multi-Domain Variational Autoencoders for Combined Modeling of MRI-Based Biventricular Anatomy and ECG-Based Cardiac Electrophysiology , Marcel Beetz, Abhirup Banerjee and Vicente Grau, Frontiers in Physiology, 2022\n3. Joint optimization of a β-VAE for ECG task-specific feature extraction. Viktor van der Valk, Douwe Atsma, Roderick Scherptong, and Marius Staring, arXiv 2023\n4. Feasibility of ECG Reconstruction From Minimal Lead Sets Using Convolutional Neural Networks, Maksymilian Matyschik; Henry Mauranen; Pietro Bonizzi; Joël Karel, IEEE 2020 Computing in Cardiology\n\n\nAs no related work is mentioned, also no baselines or comparisons are performed. I think the simplest comparison would be to take the vanilla VAE and take short sequences, and then concatenate. How does this compare to the proposed folded VAE?\n\nFurther the presentation of results is very poor. There are long vectors of numbers put into the text. Please create e.g. tables that show what method you use, what the result is for different architectures or modifications, and display this in a structured way." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A manifold encoding decoding architecture of CNN VAE for long ECG signals which facilitates interpretable model design." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024cnn,\ntitle={{CNN} Variational autoencoders' reconstruction ability of long {ECG} signals},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v3XabZsB7j},\nnote={under review}\n}" }, "abstract": { "value": "Can variational auto-encoders (VAEs) generate flexible continuous latent space for long electrocardiogram (ECG) segments and reconstruct the input? A folded VAE architecture is introduced in this study which is able to encode long ECG segments by splitting an input segment into folds and process them in sequence using a narrow field-of-view in the encoder and concatenate them at the end, instead of processing the long segment at a time. The VAE decoder follows similar folding and concatenation strategy for reconstruction of the original ECG segments. The proposed folded VAE architecture is able to generate better reconstruction of long 30-second ECG segments compared to unfolded classical VAE approach which often produce trivial reconstruction of long ECG segments. Experimental results show that the latent representation generated by our folded VAE architecture not only retains rich compressed information but also aids designing interpretable models by providing decision-making insights." }, "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": [ "VAE", "CNN", "electrocardiogram", "reconstruction", "compression", "interpretability" ] }, "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/9d57ab156227dd7f49f132e5d6010291d8569d02.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": "CNN Variational autoencoders' reconstruction ability of long ECG signals" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "My major concerns are related to the clarity and the performance." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Enhancing GNN efficiency is an important topic.\n\n2. Theoretical justification is provided to demonstrate how POD can preserve graph information.\n\n3. Using POD to improve GNN efficiency is a novel approach." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper aims to develop a scalable training approach for GNNs based on the proper orthogonal decomposition (POD) technique. It proposes the PGNN approach, which includes a preprocessing stage to sketch the input node feature matrix, sketch the convolution matrix, and generate count-sketch matrices to obtain the sketches. Then by using the sketches in the training stage, it greatly reduces the complexity as the dimensionality is reduced. The paper provides theoretical justification as well as the empirical study results to justify the power of PGNN." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Clarity - The paper is not very easy to follow. \n- The methodology section is hard to understand, bringing the pseudocode from Appendix to the main paper may help with this. \n- The experiment section applies the proposed PGNN approach to various GNN backbone (SGC, GCN, SAGE), each compares to different baselines and on different benchmark datasets, which looks a bit confusing to me.\n- The conclusion of each experiment is hard to find in the paper.\n\n2. The performance improvements looks marginal." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "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. This paper is easy to understand.\n\n2. The idea is novel." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a sketch-based GNN training method that does not require updating the sketches during training. PGNN uses POD sketches to approximate the update rules of GNNs. The effectiveness of the PGNN method is validated by experimental results on five datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. As a work on GNN training, the current amount of experimentation is far from sufficient. The authors only presented very few experimental results, which greatly undermines the solidity of this paper. The results from Table 2 to Table 6 are scattered, missing many important results, such as the results of GCN and GAT on Products.\n2. There seems to be a problem with the format of this paper. I am unable to select any text.\n3. The author mentions in Section 3 that the proof can be found in the appendix A, but I was unable to find the appendix A." }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 would expect the use of billion scale datasets instead of datasets such as Cora and Citeseer, which are not reliable datasets to compare model quality. Do the authors have any results for such datasets?" }, "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 makes GNN training a little bit more efficient.\n* The theorems make the proposed method theoretically sound." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes approximating the feature and the adjacency matrices into a lower dimensional subspace to increase computational efficiency of model training. Theorems about the quality of the approximations are proved. Experiments show how effective the proposed approach is in terms of trained model quality." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The speedup one gets by sacrificing model quality is not great.\n* For a proposed method claiming to improve efficiency, the datasets used are small. ogbn-papers100M, ogbn-mag240M, igb-het, igb-hom datasets would be more appropriate to prove the real worth of the method.\n* Figure 3 y-axis has no reference numbers.\n* Experiments against GraphBolt's [1] fully GPU-enabled GraphSAGE implementation [2] should be made if the authors want to compare against one of the best available GraphSAGE implementations, when it comes to runtime efficiency.\n* The experimental evaluation is focused against the SGC baseline, reducing the impact of the work. When the method is compared against nonlinear baselines, it does not fare well (Table 6).\n\n\n[1] https://www.dgl.ai/dgl_docs/api/python/dgl.graphbolt.html\n\n[2] https://github.com/dmlc/dgl/blob/master/examples/graphbolt/pyg/node_classification_advanced.py" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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": "See weaknesses." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "-The method is able to achieve a high-degree of compression on ogbn-arxiv and reddit with comparable accuracy with compared against sketch-gnn which is using less-aggressive compression. \n\n-The POD ostensibly has not been considered in this type of training regime before. \n\n-Wide set of experiments -- citeseer, cora, ogbn-arxiv, reddit are all classical GNN datasets along with comparisons against sketch-gnn, Graph-SAINT, VQ-GNN." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose an algorithm to shrink the memory and computational complexity of forward passes in a message-passing based setting for GNNs. The authors rely on a proper orthogonal decomposition (POD) to achieve compression, mixed with some classical sketching techniques such as locality-sensitive hashing and count sketches. The method is able to achieve high compression over some graph datasets while remaining competitive against the uncompressed setting and baselines such as Sketch-GNN. Theory is provided on the optimality of the POD along with error bounds on the approximation." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "-Novelty is limited: the approach is simply using a low-rank projector and still relies on count-sketches per the update rule (5). The sketched update rules look remarkably similar to those of Sketch-GNN minus the POD which is taking care of the non-linearity (more on this below).\n\n-Sketch-GNN uses polynomial tensor sketches to avoid blow-up in the backwards pass when examining the derivative with respect to the activation function. The authors of this work are using LSH (per SLIDE) to avoid this. A few issues with this: (1) Is this really less complicated or computationally-less efficient than using learnable sketches? LSH (the SimHash) relies on dense Gaussian matrices and even SLIDE acknowledges learnable projections must be used. (2) The usage is not appropriate: observe that the LSH in SLIDE is used in the *final* layer as the magnitude of the dot product directly corresponds to the magnitude of the logit, i.e., class probability. When used in intermediate layers, all you are doing is ignoring smaller activations -- but these can be very important in the update, which is why the SLIDE strategy is nearly exclusively used for the final layer of massive, extreme multi-label compression tasks. (3) Sneaking in LSH for optimized forward and backwards passes is a non-trivial engineering task. The audience would appreciate seeing computational run-times associated with this overhead. \n\n-The theory is weak. In Theorem 1, the authors should clarify what they mean by \"optimal projection matrix\" as you have to look in the Appendix to gather it. The result is a close cousin of the Eckhart-Young-Mirsky theorem and the result is a few obvious inequality simplifications followed by a citation. The appendix recycles several lemmas from Ding et al., 2022 and the error bound, again, follows routine calculations from sketching theory. \n\n-The experimental results are weak. Table 3 shows the results are within-error equivalent to Sketch-GNN thus showing non-trivial improvement. In Table 4, the accuracies mostly lag or minimally improve accuracy. The authors should increase the sketch-ratio until the accuracy exceeds their competitors so the audience can understand the performance curves better. Table 6 has the same issue -- just increase the ratio until the PGNN outcompetes Sketch-GNN so we can understand at which ratio this will occur.\n\nMinor: Please fix the citations. They read as normal text -- parenthesize them." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We introduce PGNN, a novel sketch-based method utilizing Proper Orthogonal Decomposition to train Graph Neural Networks efficiently, achieving sublinear training time and memory usage relative to graph size." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024proper,\ntitle={Proper Orthogonal Decomposition for Scalable Training of Graph Neural Networks},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v44CUwEeDY},\nnote={under review}\n}" }, "abstract": { "value": "As large-scale graphs become ubiquitous in real-world applications, there is growing concern about \nthe memory and time requirement to train a graph neural network (GNN) model for such datasets.\nStoring the entire adjacency and node embedding matrices in memory is infeasible in such a scenario. Standard sampling-based methods for addressing the memory constraint suffer from the dependence of the number of mini-batches on the graph size. Existing sketch-based methods and graph compression techniques operate at higher sketch ratios, with the graph compression techniques showing poor generalization, implying that different GNNs trained on the same synthetic graph have performance gaps. Sketch-based methods necessitate online learning of sketches, further increasing the complexity. In this paper, we propose a new sketch-based algorithm, PGNN, employing the Proper orthogonal decomposition (POD) method to craft update rules to train GNNs, improving the memory requirement and training time without the complication of updating the sketches during training. Experiments on standard graph datasets show that PGNN can reach much lower sketch ratios without compromising the performance. We prove the optimality of the POD update rule for the linearized GNN (SGC). Empirical findings validate our approach, demonstrating superior performance at reduced sketch ratios and adaptability across various GNN architectures." }, "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", "Scalability", "Proper Orthogonal Decomposition", "Sublinear Complexity" ] }, "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/8a83662ca8166b418d88ca1e1ebe6502caebebf6.pdf" }, "presentation": null, "primary_area": { "value": "learning on graphs and other geometries & topologies" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/642544e83614a460f5d1ff05cce52ee2cb4005e7.pdf" }, "title": { "value": "Proper Orthogonal Decomposition for Scalable Training of Graph 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": null, "code_of_ethics": null, "comment": { "value": "I appreciate the reviewer's and AC's time and effort. I decide to withdraw the submission at this step." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "In addition to the questions in the weaknesses part, I have the following questions:\n* The authors explore different model depths, which I appreciate; however, I am curious about the necessity of this exploration. What specific insights or conclusions can be drawn from varying the model depth?\n* The competition between autoregressive and diffusion models remains strong. I would like to know the authors' perspective on the advantages of using an autoregressive approach for conditional generation compared to the more mature diffusion methods." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* The alternating prediction of image tokens and control tokens seems new.\n* Experiments show the effectiveness of the proposed methods.\n* The visualization is clear and helps illustrates the framework of the proposed method and the generation results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces ControlVAR, a novel framework for controllable visual autoregressive modeling, which integrates pixel-level controls to enhance conditional image generation. By transferring pixel-level controls like mask into the same RGB space like images, the control and the image can be tokenized using the same approach. The authors also leverage teacher-forcing guidance to enhance the sampling quality. Extensive experiments demonstrate the effectiveness of the proposed approach." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The organization of the paper lacks clarity, with some confusing aspects. For instance, at the beginning of Section 3, the notation is unclear: $C$ represents pixel-level control, while $c$ stands for token-level control, but the distinction between these two is not fully explained. Additionally, the problem formulation is introduced without any examples, making it challenging to follow. The first example of control only appears on page five, and the tokenization method is explained on page six. Before this point, it’s unclear why the number of control tokens must match the image tokens. Including examples or preliminaries earlier in the section would help readers understand these concepts. Moreover, the notation in Equation (6) is confusing—it's not immediately apparent how $x$ and Equation (6) are derived.\n* This approach also seems computationally intensive, as ControlVAR sequences are twice as long as those in VAR, leading to at least double the training and inference time. Although the authors have compared the training speed of ControlVAR with ControlNet and T2I-Adapter, the model configurations for the comparison are not provided, which makes the comparison unconvincing.\n* The citation format should be revised. Instead of relying solely on \\cite, please use \\citep or \\citet as appropriate for the 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": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. ControlVAR currently requires fine-tuning when changing base models, which could limit its adaptability. Are there plans or ongoing efforts to reduce the dependency on model-specific tuning to improve scalability, particularly for scenarios where retraining large models is not feasible?\n\n2. The theoretical framework for achieving control, introduced in Section 3.2, could be expanded to provide more depth. Could you elaborate on how control is achieved within the AR setup, particularly at the pixel level, and what makes this approach effective?\n\n3. ControlVAR’s reliance on control conditions might limit its use in scenarios that require unconditioned generation. Is there a potential for adapting or extending ControlVAR to handle unconditioned tasks, and if so, could you describe how that might be approached?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. **Promising Direction for AR Model Control**: The paper addresses an important and interesting challenge—how to control autoregressive (AR) models effectively—which is valuable for the community as AR applications grow.\n2. **Well-Designed Experiments**: The experiments are thoughtfully set up for various tasks, providing a clear view of the framework’s capabilities and its performance compared to popular models.\n3. **Clear Writing**: The paper is well-written and easy to follow, making the technical aspects and contributions understandable to a broad audience." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces **ControlVAR**, a “controllable autoregressive” (AR) image generation framework designed to enhance flexibility and efficiency in conditional generation tasks, particularly as an alternative to diffusion models (DMs) due to their computational costs and latency. \n\nThe work jointly models image and pixel-level conditions, using a teacher-forcing guidance (TFG) strategy that improves controllable generation by substituting predicted values with ground truth during inference. \n\nThis approach enables adapted VAR to handle a wide range of tasks—such as control-to-image and image-to-control generation—and even unseen tasks like control-to-control generation. \n\nThe work demonstrates good performance, compared with ControlNet and T2I-Adapter across multiple pixel-level control tasks, including mask, canny, depth, and normal control." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Resource-Intensive Tuning and Limited Flexibility**: My major concern is the motivation for the work. The limitation of ControlVAR is its requirement for fine-tuning the pre-trained VAR model, which reduces the method’s flexibility and scalability. Unlike diffusion models such as Stable Diffusion, where ControlNet adds control without altering the base model’s weights, ControlVAR necessitates modifications to the underlying VAR model to enable control. This limitation makes it less practical for applications that demand flexibility across diverse base models, as switching to a new foundational model would still require retraining or extensive fine-tuning, resulting in substantial computational and resource overhead. \n2. **Lack of Adaptability**: ControlVAR’s requirement for retraining or fine-tuning when switching base models significantly limits its adaptability, particularly in settings where retraining large models is impractical. This limitation could hinder ControlVAR's adoption in real-world applications, as it restricts the flexibility needed for various scenarios. Furthermore, ControlVAR's dependence on control conditions makes it challenging to function independently of them, reducing its suitability for tasks that might require unconditioned generation. Although extensive experiments validate the framework's performance, the primary motivation for ControlVAR should be clarified further to avoid potential confusion within the community.\n3. **Delayed Method Introduction and Lack of Theoretical Depth**: The core methodology of ControlVAR is not introduced until Section 3.2, and even then, the theoretical grounding is somewhat shallow. This late introduction and limited depth in the explanation of controllable modeling make it challenging for readers to fully understand the framework’s design and motivations. The paper’s presentation could be strengthened by expanding the network modeling section with more rigorous and specific theoretical insights. A more structured approach to explaining the controllability mechanism would enhance the clarity of ControlVAR’s contributions and make the work more accessible to a broader audience. Without a thorough explanation, the framework may appear conceptually fragmented, and readers may struggle to appreciate the full scope of its methodological innovations.\n4. **Limited Novelty in Teacher-Forcing Guidance**: While teacher-forcing guidance is a core element of the ControlVAR framework, it does not introduce substantial novelty and appears to be similar to existing guidance methods, such as classifier-free guidance. Its application within ControlVAR is not notably distinct from previous uses in controllable generation, which may lessen its perceived value and impact within the paper. This similarity raises questions about the uniqueness of the guidance method as it relates to enhancing control in AR models. Without a more innovative or tailored approach, teacher-forcing guidance may come across as a standard technique rather than a breakthrough for controllable generation.\n5. **Focus and Theme Ambiguity**: The paper’s inclusion of various tasks, such as image-to-control and control-to-control predictions, diverges from its primary focus on control-to-image generation. This range of tasks blurs the paper’s central theme, making it challenging for readers to pinpoint the core contribution and focus of ControlVAR. While demonstrating versatility is valuable, the inclusion of these peripheral tasks risks diluting the main message and may confuse readers about the intended purpose of the model. Furthermore, the name “ControlVAR” suggests a potentially misleading emphasis on control-focused, multi-task learning, which may not align with the actual scope and main objectives of the framework. This ambiguity in the framework’s focus and naming could hinder its positioning within the broader field of controllable generation." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "None" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Although the field of controllable image generation using autoregressive models is novel, and the experiments showcase the strong controllability of the proposed ControlVAR, the main concern for reviewer lies in the method’s scalability to other autoregressive models. If a much larger and more powerful base model emerges in the future, would it also require fine-tuning to achieve controllability? (ControlNet, after all, doesn’t require retraining Stable Diffusion.) Perhaps freezing the base model could be a solution worth exploring, and we would appreciate the authors’ perspective on this point." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1.\tThis paper focuses on controllable image generation using autoregressive models, a forward-looking area with substantial application potential, providing valuable insights for the research community.\n2.\tThe paper introduces ControlVAR, which employs pixel-level controls in autoregressive modeling for controllable image generation, and employs several innovative mechanisms, such as teacher-forcing guidance (TFG) for controllable sampling.\n3.\tThe paper provides some empirical results, outperforming popular DMs, such as ControlNet and T2I-Adapter, across different conditional generation tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces ControlVAR, an autoregressive framework for conditional image generation that jointly models images and pixel-level controls, enabling flexible and efficient visual generation. By employing a teacher-forcing guidance strategy, ControlVAR facilitates controlled testing across various tasks, including control-image generation, inpainting, and image-to-control prediction. Experimental results indicate that ControlVAR performs competitively with diffusion-based models, such as ControlNet, suggesting its suitability as an alternative for controllable visual generation with potentially lower computational costs." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.\tControlVAR requires tuning the pre-trained VAR model, which limits the flexibility of the proposed method. Switching to a new base model still requires retraining, making this approach resource-intensive. In diffusion models like Stable Diffusion, ControlNet does not modify SD’s weights. However, if there is a mature autoregressive generation model with a parameter size similar to SD in the future, ControlVAR would require retraining or fine-tuning it, which would be unacceptable in most application scenarios.\n2.\tIn Section 3.1, the author provides extensive details on a particular method of control extraction, but this content may not be suitable to occupy a significant portion of the core methodology section; it might be more appropriate to place it in the experiments section or supplementary materials for a more detailed discussion. \n3.\tSimilarly, the Tokenization part merely states that the same tokenizer as VAR is used, offering minimal informative value. \n4.\tThe concept of methods is only introduced in Section 3.2, resulting in a lack of detailed information throughout the methods section, and the interpretation is not sufficiently in-depth. The overall network modeling section could be described in a more specific and theoretically grounded manner." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nli2024controlvar,\ntitle={Control{VAR}: Exploring Controllable Visual Autoregressive Modeling},\nauthor={Xiang Li and Kai Qiu and Hao Chen and Jason Kuen and Kai Hu and Jiuxiang Gu and Zhe Lin and Bhiksha Raj},\nyear={2024},\nurl={https://openreview.net/forum?id=v46TPwU0Uy}\n}" }, "abstract": { "value": "Conditional visual generation has witnessed remarkable progress with the advent of diffusion models (DMs), especially in tasks like control-to-image generation. However, challenges such as expensive computational cost, high inference latency, and difficulties of integration with large language models (LLMs) have necessitated exploring alternatives to DMs. This paper introduces ControlVAR, a novel framework that explores pixel-level controls in visual autoregressive (VAR) modeling for flexible and efficient conditional generation. In contrast to traditional conditional models that learn the conditional distribution, ControlVAR jointly models the distribution of image and pixel-level conditions during training and imposes conditional controls during testing. To enhance the joint modeling, we adopt the next-scale AR prediction paradigm and unify control and image representations. A teacher-forcing guidance strategy is proposed to further facilitate controllable generation with joint modeling. Extensive experiments demonstrate the superior efficacy and flexibility of ControlVAR across various conditional generation tasks against popular conditional DMs, \\eg, ControlNet and T2I-Adaptor." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Xiang_Li35", "~Kai_Qiu2", "~Hao_Chen15", "~Jason_Kuen1", "~Kai_Hu2", "~Jiuxiang_Gu2", "~Zhe_Lin1", "~Bhiksha_Raj1" ] }, "authors": { "value": [ "Xiang Li", "Kai Qiu", "Hao Chen", "Jason Kuen", "Kai Hu", "Jiuxiang Gu", "Zhe Lin", "Bhiksha Raj" ] }, "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": [ "Autoregressive generation", "Controllable image generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "li|controlvar_exploring_controllable_visual_autoregressive_modeling" }, "pdf": { "value": "/pdf/3ec653aed0cc1c35ab95dd6d2c609bdaec2b183e.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": "ControlVAR: Exploring Controllable Visual Autoregressive Modeling" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See Weaknesses above" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper is well-written, and the idea is easy to follow. The authors did a very nice job in the narrative of the paper. The explanation of their new assumption, applications of where this assumption appears, and the theoretical analysis are clearly presented." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on solving non-convex problems and, in particular, aims to design methods for finding Nash Equilibrium. The paper proposes the n-sided PL condition and provides convergence guarantees for different block coordinate descent methods (BCD, IA-RBCS, and A-RBCD) under this new assumption. Applications of where the assumption is satisfied and some preliminary experiments are presented." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I believe there are a few issues in terms of using specific terminology that are not standard, and the author might need to revise. \n\n1. The notion of Nash Equilibrium (NE) is very popular in min-max problems and multi-agent games. However, the paper focuses on pure minimization problems and defines NE as a standard concept, which is not typically the case. How is their NE equilibrium related to the minimizer of the function they are trying to optimize? This i believe was not clearly presented in the paper. \n\n2. I believe the notion G_f(x) should be more carefully presented and explain why this is a valid measure of convergence. Why is showing that f(x^t) - G_f(x^t) reduces enough to guarantee convergence to global minima (again, this is related to the notion of NE)? I would appreciate it if the authors provide more details on this. \n\n3. Why does Assumption 3.4 make sense? This looks like a very artificial condition to have just to make the proof work. This shows that one example (in Fig 1) satisfies, but that does not mean that this is a relaxed condition. More explanation is needed\n\n4. The paper is clearly theoretical, but I would appreciate it if the authors compared their proposed method with state-of-the-art approaches for training Linear Residual Networks and Infinite Horizon n-player Linear-quadratic (LQR) Game. In my opinion, this is very important for the paper as then it would convey the message that through the new assumption, we not only provide new convergence guarantees but that via the new analysis and algorithms, we can more efficiently solve these practical problems." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 2 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. It would be helpful to highlight the differences between IA-RBCD and A-RBCD that arise due to using approximate best responses.\n2. The Figure 5d is used to claim that the third case does not occur while executing the A-RBCD algorithm. I suppose the point being made is that $\\rho\\leq \\gamma-\\gamma\\frac{\\alpha^3}{13}$. A clarification regarding the same would be appreciated. If the graph could indicate the line $y=\\gamma-\\gamma\\frac{\\alpha^3}{13}$ that would make it much easier to understand." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "Overall, the paper is well written and uses simple examples to provide intuition for important concepts. The strengths of the paper are as follows:\n1. The assumption $n$-sided PL condition introduced in the paper is a weaker notion than two-sided PL condition making the results more general. The assumption is weaker in the sense that under this the stationary points are not guaranteed to be global minimums.\n2. Linear convergence guarantee is proven for BCD and GD under $n$-sided PL condition and gradient alignment assumption." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper focuses on the problem of finding Nash Equilibrium for function $f(x)$ having a block form $f(x)=f(x_1, …, x_n)$. The paper introduces $n$-sided PL condition which extends the notion of PL condition to functions with block structure. Under this weaker assumption the algorithm Block Coordinate Descent is analyzed and its convergence to set of NE points is proven. Under additional assumption on the alignment of gradients of $f$ and $G_f$, linear convergence of BCD and vanilla GD is proven. Furthermore, the algorithms Adaptive randomized Block Coordinate Descent and its oracle version are proposed and their convergence rate is analyzed without the additional assumption on gradient alignment." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Insufficient empirical evidence presented for demonstrating the benefits of the proposed variant: a. Comparison with BCD method is missing for the experiments with Linear Residual Network and Infinite Horizon $n$-player LQR game. It gives empirical comparison of convergence rate of BCD and A-RBCD which would be helpful. b. The variant A-RBCD requires significantly more gradient computation than BCD. It would be beneficial to have comparison keeping the gradient computation budget constant across different methods. c. Furthermore, adding SGD and Adam as baselines would make the experiments more persuasive." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 you explain further about what I wrote in W1? When could it be important to find Nashes in minimization problems? Or, in a slightly different perspective, could these algorithms generalize to multi-player games with adversarial components?\n- Have you checked if the proposed algorithms are capable of further avoiding saddles if possible either explicitly, or in some sort of an ‘implicit-bias’ sense? We don’t have any evidence proving/disproving this since, in the experiments in the paper, for the first quadratic case the only option is a saddle, while for the rest of the cases, we can only see $f(x) - G_f(x)$ from which we cannot distinguish saddles from minima.\n- Case 3 happens when the vectors $\\nabla f(x_t)$ and $\\nabla G_f(x_t)$ are not well-aligned (or $\\| \\nabla G_f(x_t) \\|$ is much larger than $\\| \\nabla f(x_t) \\|$… which probably won’t happen). While by definition of $G$ it seems intuitive that the two gradients will eventually align together near the stationary points (while both converging to zero vectors), are there any ideas to quantify or theoretically analyze this part a bit further?\n\nTYPO. Page 6, converges sublinearly for $f_1$ (instead of $f_2$)\n\nTYPO. Algorithm 4, $x_{-j}^t \\rightarrow x_{-j}$ (while we use $x = x^t$ in Algorithm 3)" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- I believe that extending ideas from convex/nonconvex or minimax optimization to multi-player games is a valuable direction to study and develop further. The $n$-PŁ condition might be useful in many cases, especially in handling stuff on nonconvex (and nonconcave) games.\n- Experiments are well-aligned with the theorems. The results all demonstrate and support exactly what’s in the theoretical statements.\n- The overall writing is clear and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper focuses on the convergence rates of block coordinate descent algorithms under an $n$-sided PŁ condition. Based on an adaptive, theoretical algorithm IA-RBCD that provably converges to the NE of a minimization problem of an $n$-sided PŁ objective, the paper proposes a novel algorithm A-RBCD that is implementable via computing the approximate best response and shows the same convergence speed as the ideal version. The paper also shows empirical results in various practical settings involving $n$-sided PŁ functions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **W1.** I don’t get why and when could finding an NE for a *minimization* problem be an important thing. In Figure 1, the set of stationary points (which is equal to the set of Nashes for $n$-PŁ functions) might contain saddle points. Indeed we would want to find Nashes for minimax or multi-player game settings, but for (nonconvex) minimization problems it’s usually also important to *avoid* possible saddles among stationary points (Lee et al., 2016). I get that there might be cases when we won’t care about saddles and just want to find stationary points, maybe as in the example function with only strict saddles points in Section 4, but still, I think it’s not clear in the paper about why this search for NE’s and the $n$-PŁ assumption could be a really necessary or interesting thing.\n- **W2.** In the theorems, everything related to Case 3 makes the theoretical contribution of this paper quite weak. While empirical results in Figure 5 suggest that Case 3 never really happens here, there should have been a better explanation than just saying ‘rigorously verifying these cases is intractable’. There are no lower bound results, i.e., which means that there could be pathological cases that get stuck in Case 3 for a long time, and the $n$-PŁ assumption isn’t powerful enough. In fact, we already need a stronger $(\\nu, \\theta)$-PŁ assumption to get non-asymptotic results for Case 3, which is even sublinear.\n- **W3.** Continuing from **W2**, the main results all state no more than case-specific one-step contraction inequalities. These should have been conclusively rewritten in terms of iteration complexities for a complete understanding of the convergence rate of this algorithm. This will include quantifying how rare Case 3 happens (or maybe when exactly can we make it rare), and seeing what convergence rate we get after fixing a single step size (or possibly a schedule) considering all three cases." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "-See the previous section.\n-Lines 251-252 you mentioned \"The above result ensures the convergence of BCD to the NE set, but it does not necessarily indicate whether the output converges to a point within the NE set.\" can you give an example where this is the case? i.e. we have the convergence to the NE set but not the convergence to a point in the set. \n-Line 292: can you explain why f(x) = G_f(x) if and only if x is NE.\n-Assumption 3.4. you require the condition (5) for a given set of points. But how can one know this set of points beforehand? you mentioned just after that for instance, for f_0 this assumption is valid for some domain, but the iterations of a given algorithm applied to this problem may leave this domain in the middle of the algorithm.\n-Can you tell why Ass 3.4 for instance holds for strongly convex functions?\n-In thm 3.6 in the linear rate of convergence, since \\mu, \\alpha and \\kappa \"do not depend\" on \"n\" this rate can be negative for large \"n\"?\n-Definition 3.8, do you really need the minimum to be zero?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The main strengths of this work are:\n-The authors extended the PL condition to the n-sided PL condition. \n-The authors proposed adapted variants of the gradient descent (GD) and block coordinate descent (BCD) algorithms and demonstrated the convergence to Nash Equilibrium of these methods.\n-The authors provided theoretical proofs for the convergence rates of the proposed algorithms and established conditions under which linear convergence can be guaranteed.\n-The authors provide some numerical tests to validate their claims." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work explores the optimization of finding a Nash Equilibrium for nonconvex functions using first-order gradient-based algorithms and their variations, such as block coordinate descent. The authors introduce the n-sided PL condition, an extension of the PL condition. Then, under this condition, they analyze the convergence of various variants of gradient descent algorithms. They provide theoretical proofs of convergence rates for these algorithms and examine conditions under which linear convergence can be guaranteed." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main weaknesses of this work are:\n-Dependence on Strong Assumptions: I understand that the authors gave some toy examples and linear NN that satisfy the assumptions however in general these assumptions remain strong and hard to check.\n-The empirical validation is focused on simple toy examples and linear NN. The performance of the proposed algorithms on large-scale or more complex problems is not deeply explored, making it uncertain how well the algorithms scale in practice. The authors may consider including some tests on bigger problems to assess the numerical efficiency of the proposed variants, even if the conditions to converge are not necessarily satisfied.\n-The paper lacks a detailed computational complexity analysis of the proposed variants (the cost per iteration compared to the classical GD or BGD). The added steps (like approximating best responses in certain cases) introduce computational overhead, especially for high-dimensional data, and it’s unclear how these adaptations perform against standard or simpler gradient-based methods in terms of cost/runtime.\n-The adaptive algorithms, particularly with the introduction of conditions for selecting step sizes and updating directions, could be challenging to implement efficiently in practice." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024convergence,\ntitle={Convergence Analysis of Gradient Descent under Coordinate-wise Gradient Dominance},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v49jqwmGtM},\nnote={under review}\n}" }, "abstract": { "value": "We consider the optimization problem of finding Nash Equilibrium (NE) for a nonconvex function $f(x)=f(x_1,...,x_n)$, where $x_i\\in\\mathbb{R}^{d_i}$ denotes the $i$-th block of the variables. \nOur focus is on investigating first-order gradient-based algorithms and their variations such as the block coordinate descent (BCD) algorithm for tackling this problem. \nWe introduce a set of conditions, termed the $n$-sided PL condition, which extends the well-established gradient dominance condition a.k.a Polyak-{\\L}ojasiewicz (PL) condition and the concept of multi-convexity. This condition, satisfied by various classes of non-convex functions, allows us to analyze the convergence of various gradient descent (GD) algorithms. \nMoreover, our study delves into scenarios where the objective function only has strict saddle points, and normal gradient descent methods fail to converge to NE. In such cases, we propose adapted variants of GD that converge towards NE and analyze their convergence rates." }, "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": [ "Non-convex Optimization", "Nash Equilibrium", "Gradient Dominance", "Strict Saddle" ] }, "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/399883408091903871689fdd8baf5eca00946b2c.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/1650d25759bad38898dcc696a6e17784c654f2bf.zip" }, "title": { "value": "Convergence Analysis of Gradient Descent under Coordinate-wise Gradient Dominance" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": { "value": "Line 146 reveals the author identities, which breaks double blind review." }, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": { "value": "Submission Desk Rejected by Program Chairs" }, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@misc{\nmeo2024bayesianlora,\ntitle={Bayesian-Lo{RA}: Lo{RA} based Parameter Efficient Fine-Tuning using Optimal Quantization levels and Rank Values trough Differentiable Bayesian Gates},\nauthor={Cristian Meo and Ksenia Sycheva and Carlo Saccardi and Anirudh Goyal and Pietro Lio and Justin Dauwels},\nyear={2024},\nurl={https://openreview.net/forum?id=v4Bl6tfaaO}\n}" }, "abstract": { "value": "It is a common practice in natural language processing to pre-train a single model on a general domain and then fine-tune it for downstream tasks. However, when it comes to Large Language Models, fine-tuning the entire model can be computationally expensive, resulting in very intensive energy consumption. As a result, several Parameter Efficient Fine-Tuning (PEFT) approaches were recently proposed. One of the most popular approaches is low-rank adaptation (LoRA), where the key insight is decomposing the updated weights of the pre-trained model into two low-rank matrices. However, the proposed approaches either use the same rank value across all different weight matrices, which has been shown to be a sub-optimal choice, or do not use any quantization technique, one of the most important factors when it comes to a model's energy consumption. In this work, we propose Bayesian-LoRA, a new method that approaches low-rank adaptation and quantization from a Bayesian perspective by employing a prior distribution on both quantization levels and rank values. As a result, B-LoRA is able to fine-tune a pre-trained model on a specific downstream task, finding the optimal rank values and quantization levels for every low-rank matrix. We validate the proposed model by fine-tuning a pre-trained DeBERTaV3 on the GLUE benchmark. Additionally, we fine-tune Phi-2 and Qwen, and evaluate them on few-shot and zero-shot MMLU. We compare our proposed method with relevant baselines and present both qualitative and quantitative results, showing its ability to learn optimal-rank quantized matrices. B-LoRA performs on par with or better than the baselines while reducing the total number of bit operations by roughly 70\\% compared to the baseline methods." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Cristian_Meo1", "~Ksenia_Sycheva1", "~Carlo_Saccardi1", "~Anirudh_Goyal1", "~Pietro_Lio1", "~Justin_Dauwels1" ] }, "authors": { "value": [ "Cristian Meo", "Ksenia Sycheva", "Carlo Saccardi", "Anirudh Goyal", "Pietro Lio", "Justin Dauwels" ] }, "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": [ "PEFT", "LORA" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "meo|bayesianlora_lora_based_parameter_efficient_finetuning_using_optimal_quantization_levels_and_rank_values_trough_differentiable_bayesian_gates" }, "pdf": { "value": "/pdf/9f0636ccc5703170c04ccaa3bef771f5c5cfbc1b.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": "Bayesian-LoRA: LoRA based Parameter Efficient Fine-Tuning using Optimal Quantization levels and Rank Values trough Differentiable Bayesian Gates" }, "venue": { "value": "ICLR 2025 Conference Desk Rejected Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Desk_Rejected_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- While the theoretical analysis shown in Section 4.3 demonstrates that the Q value learned with Equation (3) has the bounds independent from the hyperparameter \\gamma, I wonder what the ablating effect \\gamma brings in practice: Will certain configuration of \\gamma contribute the learning of a better Q-value estimation? This question has another motivation: In Equation (3), some of the probability mass is allocated to the dummy actions a', and the weight ratio between the actions in the demonstration trajectory and the dummy ones has one degree of freedom determined by \\gamma. It would therefore be great to sweep \\gamma to study the ablation effect of the weight ratio.\n - And by the way, how is the \\gamma set in the demonstrated experiments?\n\n- According to Equation (3), learning the Q-value function requires a learned behavior policy \\pi_\\beta in prior. How will the quality of the behavior policy impact the learning of the Q-values, and the performance of the ultimate policy? While I appreciate the experiments in Figure 4 where only 10% of the training data are used to get the \\pi_\\beta, I wonder how the performance grows with the quality of the initial behavior policy. For example, if the initial behavior policy is near perfect, will it be that the ultimate policy, although weighted with the learned Q-values, performs similarly with the initial one?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- The work targets at the multi-turn RL scenario for LLMs, which is vital for the development of complex reasoning and agentic use in LLMs.\n- The proposed algorithm is backuped with a solid theoretical motivation.\n- The experimenting scenarios have a broad coverage, and the performance is well-demonstrated." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed Q-SFT, a method that integrates Q-learning with SFT for LLMs in the multi-turn RL setting. To exploit both the representation and the logits in LLMs learned in the pretraining stage and bypass the poor scaling effect of TD-style objectives, the authors proposed a novel weighted cross-entropy loss that embeds the learning of the Q-value into the weights. In this way, the logits prior in the pretrained LLMs can be leveraged, and no new head is needed to learn the Q-values. Theoretical analysis has shown the guarantee of the new learning objective leading to a conservative estimation of the value function. Experiments on several scenarios have demonstrated the effectiveness and efficiency of the proposed method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Please see the Questions below." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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 see the previous section." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Using probability prediction in the Q-learning objective is cool, and it sounds particularly beneficial for large vocabulary/action space. \n- The implementation of the method is straightforward: learning two models using different objectives, and use both at inference time" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a method that leverages the LLM logits learned via supervised fine-tuning to approximate the Q-learning objective. Authors argue that traditional Q-learning setup suffers from discarding the logit head, and learning a new head for Q values, so the proposed method directly translates the learned logits to Q values, with theoretical analysis on its bound and approximation. Experiments include text-based games, image-based navigation and robotic manipulation, and show some benefits of the proposed method over the prior value-based RL method." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I found the pitch very confusing: it emphasizes multi-turn problems like dialogue, but the actual math formulation treats each turn as a single token in an utterance (i.e., the action space is the vocabulary space). The generation process is to generate an utterance token by token, which is conventional, and it’s confusing where the multi-turn setup plays a role other than using this generative model repeatedly to generate different utterances at different parts of the dialogue. In other words, the value function here is not the reward per utterance as stated in the intro (line 074), and question answering is also not a single-step problem as the answer is also generated token by token.\n- Experimental setup seems simplified in terms of 1) for language-based tasks, the game nature makes the effective vocabulary very small and unambiguous, 2) the well-defined reward function setup in text games do not translate into practical problems, 3) why no supervised learning results from prior work as a direct comparison?; 4) most of the datasets already have RL method outperform supervised learning which is contradictory to the intro “offline RL fine-tuning of LLMs does not reliably outperform supervised fine-tuning”. Therefore, it’s unclear whether the improvement brought by the proposed method should be contrasted to what exactly.\n- The writing can be further improved in terms of clarity, accuracy, and fixing typos. A few examples below: \n - Table 1: number from the proposed method is bold in the Wordle column, but it’s not the best performance method\n - Define h used in line 162\n - Prime should be on a, instead of argmax in line 218\n - Consistently use “RL” instead of having “RI” occurred \n - Citation should be within a pair of parentheses: line 037, 039" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- Some typos: L39 citation, L84 sentence, L194 “performing”, Fig2 top left, \n- Eq2 should have a negative sign? Also tab1 -2.08 should be bold, not -2.11. \n- The figure 1 was bit too small and hard to read.\n- L153 defines RL as MDP, but what about POMDP?\n- About the theoretical guarantee, how good is the lower bound? A probability from a policy can be quite low for large vocabulary, in which case then the bound is not that tight?\n- How many models needed to be loaded during training? There is the main model, also a target model and the original model. So is that 3? How does that affect memory usage?\n- Given that some baselines use different base models, it would be nice to put them in the table. This will help if they perform better because their base model is stronger or not.\n- In fig3, the error bars correspond to different seeds?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- Novelty: the method seems to be quite novel. The way it works is quite different from a typical Q-learning where L2 loss is used. Instead it relies on a cross-entropy loss that is modified. Given the importance of RLHF in LLMs and the need of multi-step interaction, a study like this can be impactful and important.\n- The experiments are quite diverse, ranging from LLM-based text tasks to VLM-based robotic manipulation. It is also compared to a diverse set of baselines, which helps to solidify the claims.\n- There is a proof about a theoretical guarantee, which is always nice to have. The method also leads to an improvement in performance in practice too, where the gap increasing as the model scales. Given the lack of value-based approaches in LLMs, it would be interesting to see if this method will adopted widely.\n- Paper is well written and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes a new RL training method for LLMs that take advantage of the pretraining. It is based on Q-learning, but instead of attaching a new value head, the method uses the existing token probabilities as value function. There is some theoretical guarantees and diverse experimental results involving multi-steps interactions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- I felt like the writing of the experimental results were a bit too short. It would have been good to have more ablation studies to understand why the method works, and in which situations it is more optimal. For example, since the main motivation is that the value head is not initialized from scratch, what will happen if we do that but keep the loss the same. Also there is not much on the training details, such as the number of training steps etc. I think the other parts can be condensed to make space. \n- Another important part that was too short was the method itself. The method is introduced between L230-261, which is about only half page. It doesn’t give enough explanation about why this method should work, so it was quite hard to understand it. For example, can you give more insight into why the proposed method should be more stable? \n- Reuse of the pretrained weights is emphasized as the main motivation, but some of the tasks actually train the model from scratch and others are not in natural language, which is a bit conflicting. There is still an improvement when trained from scratch, which is good, but there is a lack of explanation why it works better." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The motivation is clear: the instability of the Q-learning objective is a well-known challenge in the RL literature.\n- The proposed method is straightforward and demonstrates strong empirical performance, with improvements in sample efficiency and scalability over previous methods.\n- The evaluation covers a wide range of benchmarks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Q-SFT, a new approach that enables fine-tuning of LLMs and VLMs within an offline RL framework but using a supervised learning-style loss function. The method matches or outperforms previous approaches across a range of LM fine-tuning benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "It seems there are some limitations to the proposed approach that are not addressed in the paper.\n- During inference, the approach requires double the number of forward passes—one for the original pretrained LM and one for the fine-tuned LM. This can be challenging in practical scenarios where real-time decision-making is essential, such as robotic manipulation.\n- In Theorem 4.1., the lower bound for the learned p_theta is Q* multiplied by pi_beta. If pi_beta is small, the resulting pi_theta can be overly conservative compared to the true Q*." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We present a new offline RL algorithm specifically to fine-tune pretrained LLMs and VLMs better." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024qsft,\ntitle={Q-{SFT}: Q-Learning for Language Models via Supervised Fine-Tuning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v4MTnPiYXY},\nnote={under review}\n}" }, "abstract": { "value": "Value-based reinforcement learning (RL) can in principle learn effective policies for a wide range of multi-turn problems, from games to dialogue to robotic control, including via offline RL from static previously collected datasets. However, despite the widespread use of policy gradient methods to train large language models for single turn tasks (e.g., question answering), value-based methods for multi-turn RL in an off-policy or offline setting have proven particularly challenging to scale to the setting of large language models. This setting requires effectively leveraging pretraining, scaling to large architectures with billions of parameters, and training on large datasets, all of which represent major challenges for current value-based RL methods. In this work, we propose a novel offline RL algorithm that addresses these drawbacks, casting Q-learning as a modified supervised fine-tuning (SFT) problem where the probabilities of tokens directly translate to Q-values. In this way we obtain an algorithm that smoothly transitions from maximizing the likelihood of the data during pretraining to learning a near-optimal Q-function during finetuning. Our algorithm has strong theoretical foundations, enjoying performance bounds similar to state-of-the-art Q-learning methods, while in practice utilizing an objective that closely resembles SFT. Because of this, our approach can enjoy the full benefits of the pretraining of language models, without the need to reinitialize any weights before RL finetuning, and without the need to initialize new heads for predicting values or advantages. Empirically, we evaluate our method on both pretrained LLMs and VLMs, on a variety of tasks including both natural language dialogue and robotic manipulation and navigation from images." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "offline reinforcement learning", "language models", "dialogue", "robotics" ] }, "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/6efdf310ad2d2f7cc8e6ef31b91b7ef9abe7ef7f.pdf" }, "presentation": null, "primary_area": { "value": "reinforcement learning" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": { "value": "/attachment/75144dc83b59a601431eb3b36d3d86a7f3d7e097.zip" }, "title": { "value": "Q-SFT: Q-Learning for Language Models via Supervised Fine-Tuning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See above." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The paper studies an interesting problem of the tradeoff between privacy and fairness in fine-tuning.\n- As a mitigation, the paper further proposes a training-free method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents an approach to improve privacy and fairness simultaneously by deactivating the neurons that react to both objectives." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Theorem 1: The theoretical model assumes that they share the same activation on certain representations, abstracted through a single random variable $Z$. This is a strong assumption as the activation values for both privacy and fairness data might correlate but almost unlikely to lead to the same value. Does the same theoretical result hold if one consider the model that considered two different yet correlated representations $Z_1$ and $Z_2$ respectively for privacy and fairness data, i.e., $I(Z_1;Z_2)>0$? \n\n- Proposition 1: The proposition statement seems problematic. In inequality (2), the terms inside mutual information is an expectation, which has no randomness.\n\n- The paper discusses an interesting finding yet it will be great to further examine this finding in different setups. E.g., is it possible to improve privacy and fairness simultaneously by simply balancing the ratio between privacy and fairness data? How do privacy and fairness vary with different number of samples in SFT?" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Refers to the weakness and questions." }, "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 investigates an interesting phenomenon in supervised fine-tuning methods, where their approach enhances the LLM's privacy awareness but decreases fairness awareness.\n\n2. The presentation is well done." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes new methods to deactivate fairness and privacy neurons, improving fairness awareness compared to Supervised Fine-tuning methods such as Full Finetuning (FFT) and LoRA." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The baseline is somewhat confusing. Why don’t the authors use privacy-related and fairness-related methods? The baseline methods mentioned in this paper do not seem to provide fairness and privacy analysis. There are many privacy-enhanced algorithms, such as DP-LoRA [1]. Do these algorithms also exhibit this phenomenon?\n\n2. Identifying the related neurons seems time-consuming. Can the authors provide an efficiency comparison with other methods? The proposed method, DEAN, appears to generate masks that decouple fairness and privacy-related neurons and then apply the mask to the weights. However, it is unclear when the masking occurs.\n\n3. The definitions of fairness and privacy differ somewhat from what I know in related fields. For example, how do these methods guarantee privacy and fairness (typically, in my field, we use differential privacy and group fairness definitions)? What are the formal definitions of privacy and fairness here, and how is effectiveness measured?\n\n4. There is inconsistency in the notations. For instance, in line 208, $D_f$ and $D_p$ represent the fairness and privacy-related datasets, but in line 215 and Alg. 1, $D_b$ becomes fairness related and $D_f$ becomes privacy related dataset.\n\n5. The theory part is too simplistic just like an inequality application. \n\n[1]Yu, Da, et al. \"Differentially private fine-tuning of language models.\" arXiv preprint arXiv:2110.06500 (2021)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "The paper uses the Hilbert-Schmidt Independence Criterion (HSIC) to estimate mutual information. Does it provide a detailed explanation of the parameter choices for HSIC? How might variations in these parameters affect the identification of neurons coupled with fairness and privacy?\n\n\nDoes directly deactivating coupled neurons lead to information loss? Is there a significant impact on model performance from this approach? Has the paper explored alternative deactivation methods, such as partial suppression or weight scaling, to further reduce any negative effects on overall model performance?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The authors introduce DEAN, a training-free method based on information theory that decouples fairness and privacy without needing extra fine-tuning, making it both effective and easy to understand.\nThe experiments show DEAN’s strong performance across multiple models and challenging scenarios, with noticeable improvements in fairness and privacy. It’s also robust in cases where data is limited or biased, making it practical for real-world situations where high-quality data can be hard to come by.\nThe paper is clearly structured, with straightforward explanations of the problem, DEAN’s approach, and helpful illustrations. The step-by-step breakdown makes it easy to follow and replicate.\nBy addressing both fairness and privacy at the same time, DEAN is a valuable tool for LLMs in sensitive areas. Its training-free design means it can be widely applied and easily integrated into existing frameworks without high computational costs." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "It addresses the challenge of balancing fairness and privacy in large language models (LLMs). The authors observe a trade-off where enhancing privacy awareness in LLMs through standard supervised fine-tuning (SFT) often diminishes fairness awareness, and vice versa. To mitigate this conflict, the paper introduces a novel, training-free method called DEAN (DEActivate the fairness and privacy coupled Neurons). Inspired by information theory, DEAN identifies and deactivates neurons that are coupled to both fairness and privacy awareness, thereby reducing mutual information between these representations. Experimental results demonstrate that DEAN effectively eliminates the fairness-privacy trade-off, achieving significant improvements in both areas, such as a 12.2% increase in fairness and a 14.0% increase in privacy awareness for the Qwen-2-7B-Instruct model. Additionally, DEAN shows robustness even with limited annotated data or when fine-tuning data is potentially biased. The authors suggest that DEAN could be integrated into broader frameworks to develop more ethical and responsible AI systems." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While the paper compares DEAN with several fine-tuning methods, it lacks a detailed performance analysis against other advanced fairness and privacy protection techniques. Including such comparisons would help clarify DEAN’s relative strengths and weaknesses.\n· The paper relies on mutual information and HSIC to identify neurons coupled with fairness and privacy, but the accuracy of this method depends heavily on the quality and representativeness of the dataset. Limited or biased datasets could lead to inaccurate identification, potentially affecting DEAN’s effectiveness.\n· The paper uses a simple, threshold-based binary classification to identify neurons associated with fairness or privacy. This approach may miss neurons that contribute to multiple tasks or whose importance varies by context. A more refined scoring system could better capture these nuanced roles, improving DEAN’s accuracy in targeting the most relevant neurons.\nDirectly deactivating coupled neurons may unintentionally disrupt other important model functions, as these neurons could also contribute to additional tasks. To minimize this risk, the authors might consider techniques like partial suppression or dynamic weight adjustments, which would allow DEAN to address the fairness-privacy conflict without sacrificing the model’s general performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. See weakness 1.\n2. Did the authors consider other, more precise methods of neuron selection? Why did they ultimately choose this mechanism based on importance scores? Were other selection mechanisms tried and their effects compared?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The proposed method can simultaneously improve the model's awareness of fairness and privacy protection by identifying and de-activating, without sacrificing the overall performance of the model. While traditional fine-tuning methods usually require a large amount of computational resources, especially in resource-poor scenarios, DEAN does not require additional training, which is highly efficient and innovative, and provides a new way of thinking for model optimization under resource-constrained conditions.\n2. The authors tested the DEAN approach on several LLMs of different sizes and architectures (e.g., Qwen2, Vicuna, Llama2, etc.) to ensure its broad applicability and reliability. By using different types of datasets such as Beavertails, Salad-bench and Alpaca, the authors were able to evaluate the performance of DEAN in terms of fairness and privacy preservation, and to examine its effectiveness for modeling different data environments and task requirements.\n3. The paper uses several charts to visualize the effects of DEAN, for example, to show the conflict between fairness and privacy, the performance of DEAN on different models, and so on. The table also lists the fairness and privacy enhancement of each model, which makes the experimental results more intuitive and convincing." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an approach called DEAN that aims to mitigate the conflict between fairness and privacy-consciousness in large language models (LLMs) by de-activating coupled fairness and privacy neurons. It is found that enhancing privacy awareness through traditional fine-tuning methods leads to a decrease in fairness and vice versa. DEAN utilizes information theory to reduce the interplay between fairness and privacy, thereby enhancing the independence of the two. The authors conducted experiments using DEAN on several models (e.g., Qwen2-7B-Instruct, Llama2, etc.) as well as on three datasets with different dimensions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The de-activation operation may have side effects on other features of the model, such as affecting generation fluency or multitasking ability. Although the article evaluates the impact of DEAN on the overall performance of the model through several benchmarks (e.g., HellaSwag, Race, MMLU) in Table 3, generation fluency and multitasking capabilities were not specifically tested. It is recommended that the authors provide more specific assessments of generative fluency and multitasking, possibly through more specific generative tasks or benchmarks (e.g., generative coherence or cross-task accuracy metrics), to better understand the potential impact of DEAN on model performance.\n2. Although the importance scoring mechanism is simple and efficient, the lack of comparative analysis with other neuron selection methods may affect the optimal performance of DEAN. It is recommended that the authors further experiment with other neuron selection methods, e.g., using clustering algorithms (e.g., K-means or DBSCAN) to cluster neuron activation patterns and group neurons belonging to sensitive feature mappings into the same group; and using interpretive methods, such as SHAP or LIME (Local Interpretable Model-agnostic Explanations), to compute the contribution of each neuron in the fairness and privacy tasks. The authors can consult relevant paper references to compare the effects of different selection strategies on the de-activation effect to confirm whether the choice of importance scores is optimal. This will help to understand the differences in the performance of different methods in decoupling operations and may further enhance the performance of DEAN.\n3. The experiments mainly focused on generative tasks and lacked tests on common task types such as classification, question and answer, and sentiment analysis. May limit the effectiveness of DEAN in real-world applications. It is recommended to test on a wider range of task types (e.g., classification, Q&A, sentiment analysis), and it is recommended to use publicly available benchmark datasets, such as IMDB, SQuAD (Q&A), and AG News (Classification), to help assess the fairness and privacy-preserving ability of DEAN more comprehensively. Despite the challenges of obtaining real data, data from real-world scenarios can better model DEAN's performance in real-world applications." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024dean,\ntitle={{DEAN}: Deactivating the Coupled Neurons to Mitigate Fairness-Privacy Conflicts in Large Language Models},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v4PnwdA056},\nnote={under review}\n}" }, "abstract": { "value": "Ensuring awareness of fairness and privacy in Large Language Models (LLMs) is critical. Interestingly, we discover a counter-intuitive trade-off phenomenon that enhancing an LLM's privacy awareness through Supervised Fine-Tuning (SFT) methods significantly decreases its fairness awareness with thousands of samples. To address this issue, inspired by the information theory, we introduce a training-free method to \\textbf{DEA}ctivate the fairness and privacy coupled \\textbf{N}eurons (\\textbf{DEAN}), which theoretically and empirically decrease the mutual information between fairness and privacy awareness. Extensive experimental results demonstrate that DEAN eliminates the trade-off phenomenon and significantly improves LLMs' fairness and privacy awareness simultaneously, \\eg improving Qwen-2-7B-Instruct's fairness awareness by 12.2\\% and privacy awareness by 14.0\\%.\nMore crucially, DEAN remains robust and effective with limited annotated data or even when only malicious fine-tuning data is available, whereas SFT methods may fail to perform properly in such scenarios. We hope this study provides valuable insights into concurrently addressing fairness and privacy concerns in LLMs and can be integrated into comprehensive frameworks to develop more ethical and responsible AI systems. Our code is provided in the supplementary materials." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Large Language Models", "Fairness", "Privacy" ] }, "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/cc64dafae7fb1e196576a18c4f7417363c10563d.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/1cbc82ba3db8d77d76f728d2bd3f949d0e01c1dc.zip" }, "title": { "value": "DEAN: Deactivating the Coupled Neurons to Mitigate Fairness-Privacy Conflicts in Large Language Models" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1. How would you extend D-TSNs to use for solving stochastic decision making problems? What modification would be required in the current work to accommodate stochastic transitions and rewards?\n\n2. Could you provide more details on how computationally expensive D-TSN is and how does the method scale w.r.t. the action space size and the search tree depth?" }, "rating": { "value": 10 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "1. Novel Architecture: The paper proposes a novel neural network architecture, D-TSN, which embeds the inductive bias of a best-first search algorithm, allowing for the end-to-end learning of planning components from demonstration sequences. \n\n2. Joint Learning of Planning Components: D-TSN jointly learns the encoder, value function, and world model. This is advantageous when the world model is not given but is needed to be learned from data. \n\n3. Variance Reduction Technique: Authors use an effective variance reduction technique using a telescoping sum in the REINFORCE algorithm to addresses the high variance associated with policy gradient methods. \n\n4. Comprehensive Experiments: The method is applied to a wide variety of tasks, such as reasoning problems, navigation, and game environments, supporting the claim that it is versatile and effective across domains. \n\n5. Improved Performance: The authors show that D-TSN outperforms baselines, showing its problem solving performance in challenging tasks with limited supervision, especially in jointly learned world model settings." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces the Differentiable Tree Search Network (D-TSN), a novel neural network architecture designed to learn search strategies from sequences of demonstrations without access to explicit search trees. D-TSN integrates the inductive bias of a best-first search algorithm into its structure, enabling the joint learning of essential planning submodules, including an encoder, value function, and world model. To construct the search tree, the authors employ a stochastic tree expansion policy, formulating it as a decision-making task optimized via the REINFORCE algorithm. They introduce a variance reduction technique using a telescoping sum to address high variance in gradient estimates. D-TSN is applicable in scenarios where the world model is known or needs to be jointly learned with a latent state space. The authors evaluate D-TSN on tasks from both scenarios, including the Game of 24, a 2D grid navigation task, and Procgen games. Experimental results demonstrate that D-TSN is effective, particularly when the world model with a latent state space is jointly learned, outperforming baselines in terms of success rate and generalization capabilities." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Limited to Deterministic Environments: The current implementation is restricted to deterministic decision-making problems with discrete action spaces.\n\n2. Computational Complexity: The computational complexity for the approach might be high, because it consists of constructing search trees and performing REINFORCE updates. This can be a problem especially when applying for deeper trees or larger action spaces.\n\n3. Scalability: scalability is not thoroughly analyzed for longe-horizon tasks or higher dimensional state space." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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": "- Why did the authors choose to use REINFORCE? Did the authors experiment with using methods like Gumbel-Softmax for sampling?\n- How are the compute budgets between model-based search, model-free Q network and D-TSN equalized for a fair comparison?\n- In Line 22, the authors say they “introduce an effective variance reduction technique”. Isn’t this just Guez et al. (2018)? The way this is written in some places suggests that this is novel.\n- Line 45,46 seems contradictory with lines 48,49. If learning from demonstration sequences fails due to compounding errors due to the agent getting into states not during training, then the training distribution of the proposed method is also not sufficiently covered by the training distribution. I understand the CQL term attempts to address this issue.\n- What are the scores of the PPG sub-optimal policy?\n- Table 3 reports Mean Scores and Mean Z-Scores, but no standard deviations or error bars?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "- Builds on TreeQN and improves a significant limitation of the prior work, i.e., having a fixed tree structure. In reality, search algorithms should attempt to filter large action spaces and focus computation on promising variations in the tree. The proposed work gets around this limitation by sampling from the action space, and using REINFORCE to differentiate through the discontinuity of sampling.\n- Strong empirical evidence that the proposed method improves on TreeQN, and having the modules trained separately.\n- Strong ablation results showing the effectiveness of the proposed method." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel differentiable neural tree search architecture that can be learned directly from trajectories of data in a supervised manner. It essentially introduces a search-like inductive bias within the weights of the neural network. The proposed algorithm builds on TreeQN (Farquhar et al., 2018), and is crucially different since the proposed method allows building a tree structure that can stochastically sample from the action space, and not just be a fixed tree structure as in TreeQN. The authors empirically evaluate the strength of their approach by comparing against a Model-free Q network, a Model-based search method that trains the individual modules separately, and TreeQN, and report performance gains in various RL environments." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Seeing how the approach handles large action spaces remains an empirical question, since currently, there is no policy that directly outputs a distribution over actions, instead the method requires the application of the transition network and the reward network for every action, which is not scalable to settings like, say, Go.\n- The proposed method is mostly applicable to discrete action spaces with deterministic environments. Improving on this remains a future empirical question." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 is the full expansion mechanism in this work not a shallow tree (in comparison to the TreeQN) that you mentioned in the RELATED WORKs section since you mentioned that you have a fixed depth in your search protocol?\n\n+ When the encoder, value, reward, and transition functions are unavailable, how do they get jointly optimized with the search algorithms?\n\n+ In paragraph III, the justification for why end-to-end learning improves the reliability of world-models used in preconditioning the search process is not given. Please address this.\n\n+ Lines 124-126: \"Dividing the network into these submodules reduces\ntotal learnable parameters and injects a strong search inductive bias into the network architecture, preventing overfitting to an arbitrary function that may align with the limited available training data.\" Where is the proof for this?\n\n+ Is there no way the illustration in Figure 2 can be moved to the main text from the Appendix? Probably as a subfigure on page one or so would be really nice! Also, I think the algorithm should justifiably be compressed (e.g. as a pseudocode or flow chart) somewhere in the main text. An elaborate version can be embedded in the Appendix if space is an issue.\n\n\n+ What metric are you using to report the measured quantities in Table 1? Can you include it in the heading/sub-headings?\n\n\n+ Appendix B, Lemma B.3: I'm sorry, is there supposed to be a running sum over all of x in the equation you wrote? \n\n+ Line 915, I think you meant Lemma B.3, not Theorem B.3, no?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ This paper introduces a differentiable treee seach and planning network to alleviate the suboptimal search results that arise when using only action and observation sequences to learn to plan in many modern data-driven plannning problems. The tree search basically synergizes submodules in the form of an encoder, value, reward and transition functions from a network by inducing the bias of a best first search into the neural architecture.\n\n+ I love the motivation stated for constructing an search tree expansion policy that is stochastic in nature. But the justification for why it ensures the continuity of the loss function when the search tree is not fixed is missing. I am referring to lines 55-56.\n\n+ I love the conceptualization and the synergy of REINFORCE, mathematical mechanisms to reduce variance in the REINFORCE Loss owing to possibly biased estimates, the continuity proof (though I have questions hanging over the proof of Lemma B.3 to be fully satisfied with this poposition) .\n\n+ I love that the conclusion section meticulously summarizes the problem, contributions, and shortcomings. Kudos to the authors." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a differentiable treee seach and planning network to alleviate the suboptimal search results that arise when using only action and observation sequences to learn to plan in many modern data-driven plannning problems. The tree search basically synergizes submodules in the form of an encoder, value, reward and transition functions from a network by inducing the bias of a best first search into the neural architecture.\n\nArguing that for a slight perturbation in network parameters, the implementation of the loss function in wuation (2) could generate a tree structure that causes the loss function to become noisy, the authors equated this to a lack of continuity in the loss function space. I think they mistook stochasticity in gradient propagation with discontinuity. The whole premise of the contribution of the paper is based on this assumption that is barely proven to be true or false before the authors dived into a host of mathematical analysis that resulted in the loss function on Line 272 (please number all your equations to allow for easy referencing in the future). As a result of this key oversight, it is not clear to this reviewer that the whole contribution of the paper is warranted. \n\nAs a result, I am assigning a score of fair to this contribution until I see a well-laid out argument for why this new invention is necessary." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "While I do love the mathematical contributions of the paper, I think there are essentials that are missing in the logic, organization, and flow of arguments that require a thorough review before this paper makes it into an acceptance. A principal one is the following (mentioned in the summary box but repeated here. The authors would do well to alleviate my concerns): Arguing that for a slight perturbation in network parameters, the implementation of the loss function in wuation (2) could generate a tree structure that causes the loss function to become noisy, the authors equated this to a lack of continuity in the loss function space. I think they mistook stochasticity in gradient propagation with discontinuity. The whole premise of the contribution of the paper is based on this assumption that is barely proven to be true or false before the authors dived into a host of mathematical analysis that resulted in the loss function on Line 272 (please number all your equations to allow for easy referencing in the future). \n\n+ The claim in the last paragraph of Theorem 3.1 that slight changes in the network parameters could cause discontinuity in the tree structure seems anecdotal and not backed up by a solid proof. I would love to see a concrete reasoning (analytical proof or abundant empirical proofs) behind this claim that warrants section 3.5 and Appendix C.\n\n+ Grammatical errors fly out of the page hither and yon throughout the paper; the uthors would do well to carefully organize their arguments, present their logic convincingly throughout the paper, and punctuate and label every equation appropriately!\n\n+ The logic in the paper could use a more thoughful presentation. Here is an example critique:\n - In the \"introduction\", it is stated in the first paragraph that constructing a search tree from expert data is infeasible due to lack of practicality or scarcity. The authors make an assumption that a search tree is a principal prerequisite for information retrieval (IR) without any justification as to why it may be better than alternative IR methods. Then in the second paragraph, they mentioned how search and planning could be better executed in the presence of a simulator or world mode. While I find this premise alluring, I find it disingenious that the authors claimed that the search could be incomplete because the search process may visit regions unexplored during training. I think the reasoning here is incomplete and should be revisited by the authors." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 differences in the convergence rates of submodules exist? If present, could these differences impact the overall network's performance?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "i) This work introduces the integration of the algorithmic inductive bias from the best-first search algorithm into a differentiable network, enabling automatic guidance of the tree search process through gradient backpropagation.\n\nii) This work underscores the significance of maintaining continuity of both the parameter space and the loss function which are dependent on the tree structure. To address this, the authors advocate for the adoption of a stochastic expansion policy to fulfill these prerequisites. \n\niii) The experiment results are compelling. And it is particularly noteworthy to see the success achieved in tasks involving LLM." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This study presents a new differentiable tree search network comprising multiple submodules to estimate Q values. The network is optimized by minimizing the distance between the estimated Q values and the ground-truth Q values within a provided dataset. To address the substantial changes in the tree structure resulting from updates to the Q value function, a stochastic expansion policy is introduced to guide expansions in the search process. This policy ensures the continuity of the parameter space irrespective of changes in the tree structure." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "i) Previous methods have introduced diverse differential network architectures to integrate different search algorithms into networks, as mentioned in the related works. It is unsurprising that integrating the best-first algorithm into the network has also yielded success. Thus, it would be beneficial to compare the architectural variances between this method and previous methodologies.\n\n\nii) This work trains the overall network using an offline dataset. However, as extensively deliberated in preceding offline RL studies, this paradigm may get stuck when facing out-of-distribution states or actions. Thus, a comparative analysis between online training and offline training for the newly proposed network architecture could provide valuable insights." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a method that constructs search tree in a differetiable manner, and can be trained from just sequence demonstrations." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning to Search from Demonstration Sequences},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v593OaNePQ},\nnote={under review}\n}" }, "abstract": { "value": "Search and planning are essential for solving many real-world problems. However, in numerous learning scenarios, only action-observation sequences, such as demonstrations or instructions sequences, are available for learning. Relying solely on supervised learning with these sequences can lead to sub-optimal performance due to the vast, unseen search space encountered during training. In this paper, we introduce Differentiable Tree Search Network (D-TSN), a novel neural network architecture that learns to construct search trees from just sequences of demonstrations by performing gradient descent on a best-first search tree construction algorithm. D-TSN enables the joint learning of submodules, including an encoder, value function, and world model, which are essential for planning. To construct the search tree, we employ a stochastic tree expansion policy and formulate it as another decision-making task. Then, we optimize the tree expansion policy via REINFORCE, and introduce an effective variance reduction technique for the gradient computation. D-TSN can be applied to problems with a known world model or to scenarios where it needs to jointly learn a world model with a latent state space. We study problems from these two scenarios, including Game of 24, 2D grid navigation, and Procgen games, to understand when is D-TSN more helpful. Through our experiments, we show that D-TSN is effective, especially when the world model with a latent state space is jointly learned." }, "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": [ "planning", "reasoning", "learning to search", "reinforcement learning", "large language model" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/f607d2ad853f8a143e1081061c8daa0635beeadf.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": "Learning to Search from Demonstration Sequences" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Causality Understanding:\n- How does the paper differentiate between correlation and causation, particularly in the first paragraph's examples?\n- Can we truly claim a direct causal relationship between sporting events and electricity usage, or is this more of a correlation due to viewer behavior?\n2. Terminology Clarity:\n- What specific meaning does \"dissimilar information\" convey in Figure 1?\n- What evidence supports the superiority of causal approaches over similarity-based methods?\n- How are these differences quantified and demonstrated?\n3. Generalizability Claims:\n- What specific types of time series data does this approach effectively handle?\n- What evidence supports the claim of applicability to \"a large class of time series data\"?\n- What are the limitations or boundaries of this approach?\n4. Model Evaluation:\n- How do individual components, particularly the residual connections, contribute to the model's performance?\n- What would ablation studies reveal about the preservation of time step information?\n- Which components are essential versus optional for model success?\n5. System Complexity:\n- How does the model account for indirect relationships in weather systems, such as the cloud cover-precipitation-temperature chain?\n- How are feedback loops and mutual influences between variables handled?\n- Does the model oversimplify complex environmental systems?\n6. Causality Validation:\n- What methods were used to establish true causality rather than correlation?\n- How were potential confounding factors identified and controlled?\n- What role did temporal sequences play in establishing causal relationships?\n- Were controlled experiments or alternative validation methods considered?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper introduces SPACE (Series Prediction Augmented by Causality Estimation), a novel model for multivariate time series forecasting. SPACE addresses three key characteristics of multivariate time series: causal relationships rather than mere similarities, information across multiple independent factors, and inherent temporal dependencies. The model integrates several components, including a Sequence Enhancer using attention mechanisms, a Cross-TE module that computes transfer entropy to capture causal relationships, and a Causal Graph Neural Network (CGNN) that uses the causality matrix as an adjacency matrix. The authors argue that conventional time series analysis methods often fail to capture these complex relationships, leading to incomplete or misleading conclusions.\n\nExperimental results demonstrate that SPACE outperforms eight state-of-the-art baseline models on nine real-world datasets. The model shows improved performance across various prediction lengths and datasets, including weather, electricity, and financial data. The authors emphasize that the integration of causal information is essential for improving forecasting performance in complex, real-world time series data. Additionally, they claim that SPACE enhances the interpretability of forecasts, especially on weather-related data, by capturing and visualizing causal relationships between different variables." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces SPACE (Series Prediction Augmented by Causality Estimation), a novel model for multivariate time series forecasting. SPACE addresses three key characteristics of multivariate time series: causal relationships rather than mere similarities, information across multiple independent factors, and inherent temporal dependencies. The model integrates several components, including a Sequence Enhancer using attention mechanisms, a Cross-TE module that computes transfer entropy to capture causal relationships, and a Causal Graph Neural Network (CGNN) that uses the causality matrix as an adjacency matrix. The authors argue that conventional time series analysis methods often fail to capture these complex relationships, leading to incomplete or misleading conclusions." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Oversimplification of causality: The paper appears to oversimplify the concept of causality and its role in prediction tasks, potentially conflating correlation with causation in some instances. In paragraph 1, the example does not clearly demonstrate direct cause-and-effect relationships. For instance, a sporting event doesn’t directly cause increased electricity usage - it’s correlated with increased usage due to more people using electrical devices to watch the event.\n\n2. Lack of precision in terminology: The use of terms like “dissimilar information” in Figure 1 is vague and doesn’t clearly explain what causal information might be captured that similarity-based approaches miss. In addition, the paper seems to imply that causal approaches are superior to similarity-based approaches in all cases, without providing sufficient evidence for this claim.\n\n3. Overgeneralization: The authors make broad claims in their first contribution about the applicability of their approach to “a large class of time series data” without specifying the types of data or providing adequate evidence.\n\n4. Absence of ablation studies: There’s no mention of ablation studies to demonstrate the impact of specific components, such as the residual connections (line 236-238), on the model’s performance and preservation of time step information.\n\n5. Oversimplification of complex systems: The paper seems to underestimate the complexity of systems like weather (Figure 3), suggesting that simple causal relationships can capture their full complexity. In this example: Some of the described relationships may be indirect. For example, increased cloud cover associated with precipitation could lead to decreased solar radiation and temperature, rather than the precipitation itself directly causing these changes. In addition, weather systems often involve feedback loops where variables influence each other in complex ways. For instance, increased humidity can lead to more precipitation, which in turn affects humidity levels.\n\n6. Insufficient rigor in establishing causality: The paper doesn’t describe a sufficiently rigorous approach to establishing true causality, which would require consideration of potential mechanisms, controlled experiments (where possible), and careful examination of temporal sequences and potential confounding factors." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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 dataset used in the experiments are rather toy small set. What are the experiment results of SPACE on larger dataset, such as NY taxi or climate datasets? \n\nCan you add more recent baselines, such as TSmixture? https://arxiv.org/abs/2303.06053" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "+ Causal structures is an important aspect for time series forecasting. Most exsiting work did not take it into consideration\n+ The authors considers scalability when designing the model. \n+ The proposed idea is reasonable and some encouraging experiment results." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents an end-to-end trainable Series Prediction model Augmented by Causality Estimation, namely SPACE, to incorporate temporal dependencies and causal relationships in time series forecasting. SPACE utilizes a temporal embedding and a transfer entropy module in the hope to capture the causal structures within multivariate time series for better forecasting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- While considering causal structures is a good idea, the way SPACE infers the causal structures is not fully convincing. The paper goes a long way to discuss issues with existing work, especially those utilizes granger causality, ad argues for transfer entropy. Since transfer entropy (TE) is difficult to calculate, the authors proposed to use pseudo TE which assumes that \"the time series follow the normal distribution\". This is an extremely strong assumption and not \"acceptable assumption for real-world time series\". \n\n- The improvement SPACE achieves over state-of-art methods are very marginal. It is not clear whether it is worthwhile to go with such a complicated model but very limited improvement." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 1 }, "primary_area": null, "questions": { "value": "1. Can the authors clarify the specific problem in time-series causality extraction that SPACE addresses, beyond combining existing modules?\n2. Could additional ablation studies be provided to isolate and validate the causal contributions of Cross-TE versus standard attention mechanisms?\n3. Are there any plans to release additional experiments or dataset evaluations, specifically with *Traffic* and *Electricity*, to strengthen model generalizability?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. **Innovative Approach**: Integrates causality estimation with multivariate time-series forecasting, which is a unique perspective.\n2. **Real-World Application**: Demonstrates effectiveness in real-world, multivariate forecasting tasks, which may offer practical benefits." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents the SPACE model, an approach for multivariate time-series forecasting that integrates causality estimation to capture complex interdependencies." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. **Dataset Coverage**: Commonly used datasets, such as *Traffic* and *Electricity*, are missing from the evaluation, limiting the scope of comparison.\n2. **Efficiency Claims**: While SPACE is described as highly efficient, there is no empirical comparison of runtime with established models like iTransformer and DLinear.\n3. **Unverified Claims in Figures**: Figure 1 implies that the attention mechanism only focuses on similar series, a claim not substantiated by experiments or theory.\n4. **Notation and Writing Inconsistencies**: The paper has numerous small errors in notation (e.g., non-italicized symbols) and inconsistent symbol usage, particularly in the methodology section (e.g., line 276 “N” should be italic, lines 274 and 284 should use “P_N,” and line 285 should italicize “T”). Such inconsistencies affect readability and technical accuracy.\n5. **Causal Adjacency Matrix Evaluation**: It is unclear in Figure 3 how the learned adjacency matrix by Cross-TE outperforms traditional attention in identifying causal relationships. Figure 1 claims attention ignores dissimilar information, yet Figure 3 does not convincingly demonstrate that Cross-TE resolves this.\n6. **Clarity in Equation 1**: Equation 1 lacks clarity in summation notation, as it’s unclear which part of the formula is summed over. Additionally, the notation for \\(i\\) as a time step index raises questions about why an additional superscript is needed to denote the previous time.\n7. **Inconsistencies in Variable Definition**: In the problem definition, \\(x\\) is defined as a time series, \\(X\\) as historical sequences, and \\(Y\\) as future sequences. However, line 263 redefines \\(y\\) as historical, which can be misleading and suggests potential label leakage. The overall presentation of the methodology is unclear and could benefit from consistent use of symbols and fonts.\n8. **Perceived Model Complexity**: Without clear experimental support for its causal relationship advantages, SPACE may appear as a combination of existing techniques (PTE, attention, GCN, and Time Mixers) without sufficient innovation in causality extraction." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "-\tStrong motivation: The paper presents an interesting method by providing a causative perspective for analyzing time series.\n-\tGood clarity: The writing is clear and well-structured, making the paper easy to follow and understand." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces SPACE to forecast MTS enhanced by causality estimation. \nSpecifically, this method captures causal relationships by a transfer entropy-based cross-TE module and a casual GNN." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Lack detailed comparisons with highly related works. \n - The proposed method relies heavily on patch mixers and time mixers (as described in Equation 6). However, state-of-the-art mixer-based methods, e.g., TimeMixer [1] and Timexer [2], are neither discussed in the related work nor included in the experimental comparisons.\n - Since one of the main contributions is about causal GNN, some GNN related works, e.g., CrossGNN [3], should be mentioned and compared.\n- Lack a comprehensive efficiency study.\n - The method introduces the fast-pTE algorithm and emphasize its efficiency advantage in contributions. It is recommended to provide detailed theoretical analysis or discussion and conduct corresponding ablation studies about variants with TE, pTE, and fast-pTE. \n - The efficiency study about the proposed work and SOTA works are also necessary, which can better demonstrate practical benefits of the proposed work.\n- Need more evaluations about the learned relationships. \n - The authors claim that Cross TE is “designed to dynamically learn and adapt to evolving causal dependencies while preserving memory of past relationships.” However, the current evaluations are insufficient to convincingly demonstrate this claim. Relying on a single sample of the learned adjacency matrix (as shown in Fig. 3) is not enough to explain the dynamic nature of causality. Additional evaluations focusing on dynamic causal dependencies should be provided.\n - It is also recommended to include examples from real-world datasets that showcase causal relationships. Presenting these examples in a visual format similar to Fig. 1 would help clarify the learned relationships.\n- Should enhance presentations.\n - In Fig. 2, subfigures (a) and (b) seem unnecessary. It is more valuable to provide a detailed illustration of the Entropy Graph construction, as it is closely related to the paper’s main contribution.\n - For Fig. 3, adding variable names instead of numbers on the axes can enhance clarity and understanding.\n\n[1] Wang, S., Wu, H., Shi, X., Hu, T., Luo, H., Ma, L., ... & ZHOU, J. 2024. TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting. In The Twelfth International Conference on Learning Representations.\n\n[2] Wang, Y., Wu, H., Dong, J., Liu, Y., Qiu, Y., Zhang, H., ... & Long, M. 2024. Timexer: Empowering transformers for time series forecasting with exogenous variables. arXiv preprint arXiv:2402.19072.\n\n[3] Huang, Q., Shen, L., Zhang, R., Ding, S., Wang, B., Zhou, Z., & Wang, Y. 2023. CrossGNN: Confronting noisy multivariate time series via cross interaction refinement. Advances in Neural Information Processing Systems, 36, 46885-46902." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024multivariate,\ntitle={Multivariate Time-series Forecasting with {SPACE}: Series Prediction Augmented by Causality Estimation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v5BouOktUP},\nnote={under review}\n}" }, "abstract": { "value": "The analysis of multivariate time series (MTS) presents a complex yet crucial task with substantial applications in areas such as weather forecasting, policy formulation, and stock market prediction. It is important to highlight three key characteristics of MTS that contribute to the challenging and multifaceted nature of their analysis: (i) their interrelationships are represented through causal relationships rather than mere similarities; (ii) they convey information across multiple independent factors; and (iii) their dynamics often arise from inherent temporal dependencies. While conventional time series analysis frameworks often fail to capture one or more of these aspects, resulting in incomplete or even misleading conclusions, we propose an end-to-end trainable $\\textbf{S}$eries $\\textbf{P}$rediction model $\\textbf{A}$ugmented by $\\textbf{C}$ausality $\\textbf{E}$stimation (SPACE) to address these limitations. This model effectively incorporates temporal dependencies and causal relationships, featuring a temporal embedding and a transfer entropy-based Cross-TE module designed to enhance predictions through causality-augmented mechanisms. Experiments demonstrate that SPACE achieves state-of-the-art results on challenging real-world time series prediction tasks, showing its effectiveness and versatility." }, "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", "Causal Learning", "Transfer Entropy", "Graph Based 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/fd62db3a4e937fb3bc598cbae54038fef716c955.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": "Multivariate Time-series Forecasting with SPACE: Series Prediction Augmented by Causality Estimation" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": null, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "Do you consider applying the high-pass filter over the Fourier domain?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The exposition is very clear, and the work is comfortable to read. \n2. The motivation of the approach is clear and reasonable. \n3. The overall framework design seems to be natural and reasonable following the basic idea.\n4. The idea is validated on two frameworks (DreamGaussian and DreamFusion), generating consistent performance improvement." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The submission proposes a pipeline to generate 3D Gaussian representation from a single image using a score distillation-based method by generating high-frequency details leveraging the Fourier domain and ensuring cross-view consistency leveraging the novel view synthesis ability of Zero-1-to-3. To the review's knowledge, the perspective is new, and the results seem promising, among score distillation-based methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The field of 3D generation has been shifting from score distillation approaches toward feed-forward methods. The manuscript would benefit from a more comprehensive discussion of this methodological evolution and its implications for the current work.\n2. While the paper mentions Wonder3D, it lacks direct experimental comparisons with this significant baseline. Additionally, Unique3D, another prominent method known for high-quality frontal appearance generation, should be included in the comparative analysis. Both Wonder3D and Unique3D have publicly available implementations, making such comparisons feasible." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "L522: I suggest using threestudio's implementation instead, which is a common practice and performs much faster and better than Zero123's results in the paper." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The method is well-motivated by insights from the frequency domain.\n\n2. It proposes a better alternative loss for SDS, which performs better when combined with Zero123 SDS." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method for single image to 3D named Fourier123. The method inherits from Magic123 but changes SDS loss to the frequency domain and only keeps the amplitude components. The adapted version is coined hy-FSD. It is applied on DreamFusion and DreamGaussian. It achieves better CLIP-Sim than the original SDS." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Low-quality texture: The appearance of 3D models looks noisy, even on colors that are supposed to be uniform. It does not look more preferable to InstantMesh, which is much faster.\n\n2. As an optimization-based method, it is much slower than inference-only methods like Stable Fast 3D, which only takes seconds and may have better geometry and appearances.\n\n3. Insufficient quantitative experiments: No 3D metrics are reported, such as CD, F-Score, Vol. IoU. It is only evaluated on 100 objects. Since most of the competitive baselines take no longer than a few minutes per shape, it would be proper to evaluate at least a few hundred objects on different datasets like GSO, Omni3D, 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": 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": "DFT amplitudes of a single image (or the $\\epsilon$ noise) is very noisy. I'm unsure if using such a noisy gradient could really improve the score distillation quality fundamentally.\n\nFor rebuttal, please address the weaknesses listed above." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 1 }, "strengths": { "value": "- Using DFT and defining a score distillation loss on the frequency amplitudes is a very original attempt at restoring texture details in the absence of accurate spatial (phase) alignment. Modeling frequency domain features is an interesting direction in generative models and I believe it's definitely worth exploring.\n\n- The writing and clarity of the manuscript is good in general." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This manuscript proposes an image-to-3D generation pipeline using pretrained diffusion models. The main idea is, while generic image diffusion models (SD) and finetuned novel view models (Zero-1-to-3) can be used jointly to optimize the 3D content via score distillation, SD provides better texture details while Zero-1-to-3 is 3D consistent but blurry. In order to better combine the strengths of these two diffusion models, the manuscript proposes Fourier score distillation (FSD), i.e., defining the score distillation loss on the frequency amplitudes. FSD is therefore applied to the SD model for preserving the texture details." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- One of the central assumptions, that novel-view models produce over-smooth results, is very outdated. This paper has only experimented with Zero-1-to-3, which was the first ever generalizable novel-view generative model. Nowadays we have Zero123++, CAT3D, and a lot of video models, some of them are open-source as well. These more recent models can already generate detailed novel views while also being more 3D consistent than Zero-1-to-3. \n\n- While the proposed Fourier score distillation method is claimed to be able to improve the visual details, I find the qualitative results very underwhelming. All images in figure 4 look blurry and cartoonish to me. In Fig. 5, Magic123 clear has better texture details, despite some failure cases that can be attributed to weak global 3D consistency.\n\n- Quantitatively, the evaluation metrics also cannot directly reflect the effect of the proposed FSD in improving texture details. CLIP-similarity is used for ablation studies but it is mostly reflecting the overall semantic alignment rather than texture details. The user studies also don't provide an aspect for texture details. In general, none of the experimental results can adequately support the central claim of the paper.\n\n- There seems to be a small error with Eq 9: if z is the denoising output as defined in L231, then there should also be an Jacobian term that backpropagates the gradient from the amplitudes to the spatial domain." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "Refer to the weakness." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The main strength is that the utilization of the amplitude of Fourier transformation in SDS loss is novel. This paper claims that Zero123 produces better structures while SD produces better details. Then, the proposed method supervises renderings using only the amplitude components of the predicted noises to extract structures of generation. This idea is overall novel and new for me." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a new SDS-based method to generate 3D models from single-view images. \nThe key idea is to incorporate both Zero123 and the original Stable Diffusion in the distillation process.\nZero123 will be in charge of generating structures in the distillation so the proposed method would use the amplitude of the predicted noise to optimize the rendering. Stable Diffusion will be in charge of generating fine details so a vanilla SDS is adopted.\nExperiments demonstrate that proposed distillation on amplitude improves the results and show some better performances than baseline methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The main weakness is that the experiments are not convincing enough to demonstrate the effectiveness of the idea.\n\n1. Lack of discussion on multiview generation methods. Distilling multiview diffusion models like MVDream often produce better results than those only using SD or Zero123, which is not discussed in the paper. Recent 3D generation methods are almost dominated by these multiview diffusion models and though the proposed method achieves better results, it does not show the potential to outperform the existing multiview diffusion methods.\n2. The claim that Zero123 always has better structures is not well demonstrated. Only one example in Fig. 2 is shown. Even though this is validated by more examples, this observation may be strongly limited to the specific trained Zero123 model instead of being a general and fundamental attribute of any other diffusion model. Thus, the observation will be too restrictive with limited impact.\n3. The results of InstantMesh and direct distillation Zero123 are not convincing. According to my experience distilling Zero123 using ThreeStudio, the results of Zero123 would be much better than the qualitative results shown in Fig. 5. I strongly suggest the authors double-check whether the implementation of the baseline method is correct or not." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Using both 2D and 3D generation priors to generate 3D from a single image with hybrid fourier score distillation" }, "_bibtex": { "value": "@misc{\nyang2024hybrid,\ntitle={Hybrid Fourier Score Distillation for Efficient One Image to 3D Object Generation},\nauthor={Shuzhou Yang and Yu Wang and Haijie LI and Jiarui Meng and Yanmin Wu and Xiandong MENG and Jian Zhang},\nyear={2024},\nurl={https://openreview.net/forum?id=v5JrYUdMxc}\n}" }, "abstract": { "value": "Single image-to-3D generation is pivotal for crafting controllable 3D assets. Given its under-constrained nature, we attempt to leverage 3D geometric priors from a novel view diffusion model and 2D appearance priors from an image generation model to guide the optimization process. We note that there is a disparity between the generation priors of these two diffusion models, leading to their different appearance outputs. Specifically, image generation models tend to deliver more detailed visuals, whereas novel view models produce consistent yet over-smooth results across different views. Directly combining them leads to suboptimal effects due to their appearance conflicts. Hence, we propose a 2D-3D **hy**brid **F**ourier **S**core **D**istillation objective function, **hy-FSD**. It optimizes 3D Gaussians using 3D priors in spatial domain to ensure geometric consistency, while exploiting 2D priors in the frequency domain through Fourier transform for better visual quality. hy-FSD can be integrated into existing 3D generation methods and produce significant performance gains. With this technique, we further develop an image-to-3D generation pipeline to create high-quality 3D objects within one minute, named **Fourier123**. Extensive experiments demonstrate that Fourier123 excels in efficient generation with rapid convergence speed and visually-friendly generation results." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Shuzhou_Yang1", "~Yu_Wang85", "~Haijie_LI2", "~Jiarui_Meng1", "~Yanmin_Wu1", "~Xiandong_MENG1", "~Jian_Zhang22" ] }, "authors": { "value": [ "Shuzhou Yang", "Yu Wang", "Haijie LI", "Jiarui Meng", "Yanmin Wu", "Xiandong MENG", "Jian Zhang" ] }, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "3D Generation", "One Image to 3D Generation" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "yang|hybrid_fourier_score_distillation_for_efficient_one_image_to_3d_object_generation" }, "pdf": { "value": "/pdf/14eedcb0a99030c137fb79d77db3c43ba637a5f0.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/b4ebc2df6ce7a0032b75a34082f67eb4334c6c0d.zip" }, "title": { "value": "Hybrid Fourier Score Distillation for Efficient One Image to 3D Object Generation" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Was the linear probing of CINP performed on concatenated MRI and FCN embeddings, or was it based on one of the modalities?
\n2. Regarding network prompting, what is meant by partitioning all samples into 5/10 subsets? What is the purpose of this partitioning? Additionally, could the authors clarify why they chose to use 10% or 50% of the data? Is this to assess retrieval performance in a low-data regime? It would also be helpful to explain why the results are better when using 10% of the training data compared to using 100%.Also, the number 29.33% seems inconsistent with the figures in 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 effectively describe the motivation behind this study and highlight the potential contributions of integrating contrastive pretraining with MRI and FCN data. The approach of cross-modal contrastive learning is well-founded and offers a solid framework for using information from both modalities to enhance diagnostic accuracy while also enabling retrieval/diagnosis when one of the modalities is missing.
\n2. The three objective functions are well-described, and the authors provide detailed ablation studies, which contribute to a better understanding of their impact on model performance.
\n3. The clear delineation between pretraining and evaluation sets facilitates relatively fair comparisons across out-of-domain datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors have proposed an interesting contrastive pretraining framework that utilizes 3D T1 MRI and functional connectivity networks (FCNs) derived from fMRI data to learn robust representations for mental disorder diagnosis. The model has been evaluated in both linear probing and retrieval settings, showcasing intriguing ideas. However, the organization of the paper and the presentation of results could be enhanced." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The writing could be improved to enhance the presentation of results. For instance, the experimental setup for network prompting is somewhat challenging to follow, as detailed in my questions below.
\n2. The authors hypothesize that the “CINP” model has potential for improvement through fine-tuning; it would be better to directly include corresponding results in the tables for a more comprehensive understanding.
\n3. The comparisons are primarily between CINP and single-modality models (sMRI or FCN). There is a lack of discussion and direct comparisons with existing multi-modal methods for mental health diagnosis, both in linear probing and fine-tuning contexts. At least, some consensus on FCN and SSP-based model predictions would allow for fairer comparisons.
\n4. Although several metrics are presented, the authors did not discuss in detail the differences, especially when two metrics offer contrasting results.
\n5. The authors did show the advantages of pretraining over simply fine-tuning a model directly on the evaluation dataset that utilizes both modalities as input. This is an important aspect to demonstrate the value of pretraining. Further improvements could include testing in a low-data regime to see if pretraining can reduce data requirements for subsequent fine-tuning.
\n6. While improvements are shown, the absolute values of metrics appear low for potential clinical applications. Providing context on results from the literature for the same task or similar datasets would help readers unfamiliar with this specific field better interpret the model's performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "- Please see weaknesses above.\n- How were the confidence intervals computed?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- The approach appears original in the way it allows the diagnosis of psychiatric disorders based on the patient's T1w MRI only. At the inference stage, two components are required: i) the T1w MR image of the patient to be diagnose and ii) fMRI data of patients with known diagnosis arranged in different diagnostic classes. The predicted diagnosis corresponds to one for which the similarity between the T1 and fMRI embedding was the highest.\n- This particular approach seems novel." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a contrastive learning approach to learn from both fMRI and anatomical T1w MRI with the aim to differentiate various psychiatric disorders." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- The paper is not always easy to follow, in particular the Quantitative results section. Careful proofreading would help.\n- It is not entirely clear how the ABIDE, ADHD and SRPBS data sets were split and used during the evaluation: i) were the splits stratified, and if so on what criteria, ii) were the results of the ablation study displayed in Table 4 obtained on the test set (and so was hyper-parameter selection based on the test set)?\n- The diagnostic classes are not balanced and no metric adapted to this scenario is used to assess the performance.\n- The references of the first paragraph of the introduction mostly do not seem appropriate:\n - 'Over recent years, there has been growing evidence that mental disorders arise from dysfunction of interconnected patterns of regions-of-interest (ROIs) in the whole brain (Krishna et al., 2023) […].' This paper is about glioblastoma, it has nothing to do with fMRI nor mental disorders.\n - '[…] fMRI-derived functional connectivity network (FCN) […] has received considerable attention in diagnosis of mental disorders (Yang et al., 2021; Bastos & Schoffelen, 2016) […].' The first paper is about diffusion MRI and the second one describes functional connectivity analysis in general and is not focused at all on mental disorders.\n- The organisation of the paper does not seem optimal. I do not see the point of having a Related Work section at the end of the paper knowing that the methods are only briefly describes and no conclusion is being drawn.\n- Several works with similar aims should be cited and discussed, e.g.\n - He, Zhibin, et al. \"F2TNet: FMRI to T1w MRI Knowledge Transfer Network for Brain Multi-phenotype Prediction.\" International Conference on Medical Image Computing and Computer-Assisted Intervention. Cham: Springer Nature Switzerland, 2024.\n - Fedorov, Alex, et al. \"Self-supervised multimodal learning for group inferences from MRI data: Discovering disorder-relevant brain regions and multimodal links.\" NeuroImage 285 (2024): 120485.\n - Li, Tongtong, et al. \"Automated diagnosis of major depressive disorder with multi-modal MRIs based on contrastive learning: a few-shot study.\" IEEE Transactions on Neural Systems and Rehabilitation Engineering (2024)." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "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": "My main concern is with the rationale behind using contrastive loss in CINP to model sMRI and fMRI representations. Based on my knowledge, there is a clear distinction between neuroimaging and image-text domains. In the image-text domain, it makes sense to maximize the differences between different categories while minimizing the differences within the same category. For instance, aligning a cat’s textual description with its image, while creating a contrast with dog-related text and images, is logical. However, in neuroimaging studies, why should we align sMRI and fMRI feature representations in this way? And why should we increase inter-subject differences?\n\nI also note that a healthy cohort was used for CINP’s pre-training. The differences within a healthy population do not fall into distinct categorical boundaries, so it is unclear what meaningful patterns are learned by maximizing inter-subject representation differences. Brain imaging features of healthy individuals are generally quite similar in terms of overall structure and function. By maximizing these differences within healthy individuals, the model may pick up biologically insignificant or irrelevant details, diverting its focus from core features. Such differences are more likely to be noise than meaningful information. In contrastive learning, distinct category boundaries are typically required to generate positive and negative samples, but healthy individuals do not naturally fall into clear categories, making the construction of inter-subject contrasts somewhat artificial. \n\nAdditionally, enforcing a “strict alignment” between sMRI and fMRI features may lack generalizability. For mental disorders, patients do not always exhibit abnormalities in both sMRI and fMRI simultaneously. For example, a patient might have a normal gray matter thickness in the prefrontal cortex as shown by sMRI, yet fMRI may reveal weakened functional connectivity between the prefrontal cortex and other regions (such as the parietal lobe or hippocampus). In such cases, a “strict alignment“ seems less appropriate. Similarly, even among healthy individuals, while there may be some correlation between gray matter thickness and functional connectivity, it is not necessarily consistent. Forcing alignment within a healthy cohort may lead the model to mistakenly assume a strong correlation between structural and functional features, thereby overlooking the natural variability and dynamic nature of functional connectivity. \n\nI look forward to seeing the authors' response and would consider raising my score if they can adequately address my concerns." }, "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 a significant problem in mental disorder diagnosis by leveraging both structural and functional MRI data, a topic relevant and valuable to the ICLR communities.\n2. The figures and descriptions in the paper are well-organized and clear, which aids in understanding the proposed approach and the CINP framework.\n3. The writing in the paper is clear, making the methodology and results easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the problem of mental disorder diagnosis using multimodal MRI data. Specifically, it proposes a framework called CINP (Contrastive Image-Network Pre-training), which applies contrastive learning between 3D T1-weighted (T1w) MRI and functional connectivity networks (FCNs) derived from fMRI. CINP aims to create a joint latent space integrating functional and structural information, enhancing diagnostic capabilities. During pre-training, the framework incorporates masked image modeling and network-image matching losses to improve modality alignment and representation quality. Moreover, the CINP contains a network prompting, enables the use of 3D T1w MRI from suspected patients and FCNs from confirmed cases to differentiate mental disorders. Extensive results on public datasets shows CINP has good performance." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper's technical contributions appear limited, as it mainly combines existing methods like contrastive learning and masked autoencoders (MAE) without significant innovation in methodology. The CINP framework may be seen as a straightforward integration of known techniques rather than a novel approach.\n2. There is still room for performance improvement. The performance of CINP on the ABIDE dataset is noticeably lower than that of the baselines, as indicated in Table 2. This suggests that the framework may not be fully optimized or may have limitations, and further improvements are needed to make it competitive across all datasets.\n3. The assumption that sMRI and fMRI features can be effectively aligned using contrastive learning lacks theoretical or empirical support from a neuroimaging or neuroscience perspective. This forced alignment may overlook important modality-specific differences, making the approach less effective for capturing unique structural-functional relationships in brain data. (See my questions below.)" }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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) Present findings related to diseases and provide analysis.\n2）Compared with the multi-model fusion methods." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1) A contrastive image-network pre-training method is proposed to use the multi-modal data for mental disorder diagnosis.\n2) The performance shows that the proposed method achieves better performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposed a contrastive image network pre-training method for mental disorder diagnosis. By using contrastive learning between the structure MRI and functional MRI, the proposed can use the useful information from both sMRI and fMRI for mental disorder diagnosis. The proposed method has been compared with several competing methods. The results show that the proposed method achieves better performance than the competing methods." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1) The significant advantage of brain networks lies in their interpretability, yet the paper lacks an interpretability analysis related to diseases.\n2）Many multimodal methods based on functional and structural MRI have been proposed, but this paper does not compare with these methods.\n3) The details of the methods are not clear enough." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024learning,\ntitle={Learning 3D Medical Image Models From Brain Functional Connectivity Network Supervision For Mental Disorder Diagnosis},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v5bK7cQch3},\nnote={under review}\n}" }, "abstract": { "value": "For mental disorder diagnosis, most previous works are task-specific and focus primarily on functional connectivity network (FCN) derived from functional MRI (fMRI) data. However, the high cost of fMRI acquisition limits its practicality in real-world clinical settings. Meanwhile, the more easily obtainable 3D T1-weighted (T1w) MRI, which captures brain anatomy, is ofen overlooked in standard diagnostic processes of mental disorders.\nTo address these two issues, we propose CINP (Contrastive Image-Network Pre-training), a framework that employs contrastive learning between 3D T1w MRI and FCNs. CINP aims to learn a joint latent semantic space that integrates complementary information from both functional and structural perspective. During pre-training, we incorporate masked image modeling loss and network-image matching loss to enhance visual representation learning and modality alignment.\nFurthermore, thanks to contrastive pre-training which facilitates knowledge transfer from FCN to T1w MRI, we introduce network prompting. This protocol leverages 3D T1w MRI from suspected patients and FCNs from confirmed patients for differential diagnosis of mental disorders. \nExtensive experiments across three mental disorder diagnosis tasks demonstrate the competitive performance of CINP, using both linear probing and network prompting, compared with FCN-based methods and self-supervised pre-training methods.\nThese results highlight the potential of CINP to enhance diagnostic processes with the aid of 3D T1w MRI in real-world clinical scenario." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "3D medical image", "functional connectivity network", "contrastive learning", "mental disease diagnosis" ] }, "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/983ae776d2e1fe0addea1af14f86675d2246fc77.pdf" }, "presentation": null, "primary_area": { "value": "applications to neuroscience & cognitive science" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Learning 3D Medical Image Models From Brain Functional Connectivity Network Supervision For Mental Disorder Diagnosis" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. I appreciate the authors discussing the runtime of POST v.s. Full PT in Table 3, but I have some questions regarding the costs of the distillation step. While the distillation costs of a large LLM can be amortized as we finetune more soft prompts with the distilled LLM, could the authors quantitatively compare the computational overhead of finetuning one soft prompt on the large LLM to that of distilling the large LLM and finetuning one soft prompt on the distilled LLM?\n2. The authors use a fixed set of distilled models throughout the paper. How does the quality of the distilled model influence the transferability?\n3. Could other model compression techniques like pruning and quantization be used in this framework instead of model distillation?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The setting of the paper is well-motivated. The introduction sections and the figures clearly explain the setting and the proposed methods." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "Prompt tuning, the process of optimizing the prefix tokens (hard) or their embeddings (soft) to improve the utility of a downstream task has become one important way of adapting Large Language Models (LLMs). Traditional soft prompt tuning paradigms require backpropagation through the full model, which is computationally expensive to perform on-premise, or handing the private data to a central LLM provider, which could lead to privacy concerns. To reduce the computational overhead of local soft prompt tuning while preserving privacy, the authors propose Privacy Of Soft-prompt Transfer (POST), which performs soft prompt tuning on a small local LLM distilled from a large LLM and transfers the soft prompt to the large model with th help of public data. In addition, the proposed framework can also be coupled with differentially private prompt tuning techniques to achieve privacy guarantees. To validate their proposed method, the authors evaluate the transferability of soft prompts finetuned on small models distilled from Roberta-base, GPT2-XL, and Llama2-7b for classification tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. While the paper presents an appealing middle ground by performing prompt finetuning on a distilled proxy model, the evaluations are performed with a fixed set of proxy models. As briefly mentioned by the authors, in the case of the centralized LLM provider, the utility of the distilled model should ideally be not strong enough to replace the original model but is still good enough to serve as a proxy for prompt finetuning. However, the authors do not further explore the trade-off between the quality of the proxy model and the effectiveness of the proposed pipeline.\n2. The evaluation of the proposed method mainly focuses on classification tasks, which have limited practicality given that modern autoregressive LLMs can perform open-ended generation tasks. While I recognize that prior works also mainly focus on classification tasks, I encourage the authors to discuss the applicability of the proposed method on open-ended tasks, or even better, demonstrate such ability through follow-up experiments." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. I wonder whether the framework needs the original input during evaluation. If so, the model input may contains privacy information.\n2. For the setting, could you please give further explanation about whether it is white or black box?\n3. Could you please provide more evidence (citations, or experiments) to confirm the claim“soft prompts are highly coupled to the LLM they were tuned on, making them difficult to transfer”?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The author proposed a novel framework for transferring soft prompts between SLM and LLM, and achieved completive results on the test dataset\n2. The paper brought focus on the privacy and efficiency issue on prompt tuning, and tried to reduce the privacy leakage and computational cost." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors propose a framework for Privacy Soft-Prompt transfer to reduce the computation cost and potential privacy leakage. The framework mainly contains three steps as deriving SLM using knowledge distillation, locally prompt tuning, and prompt transferring using public dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper writing needs to be improved. For example, the author does not clearly represent whether POST needs the original input x during the inference. If so, this may lead to privacy concerns. I mean, in tasks like classification, the LM input contains privacy information.\n2. The paper needs to be clear about the setting. For black-box LMs, they cannot provide the service as giving a SLM through KD. If the author means for white-box LMs, they author should further explain why there is privacy concerns.\n3. For the experiment, I think other tasks in spite of classification help improve the credibility. For tasks as classification, the LM input contains privacy information which violates the setting in this paper. For other tasks as Q&A, the soft prompt may contains privacy information and the model inputs may have less privacy concerns, which is consistent with the setting in this 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": { "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": "1.\tNo ablation study is found for the selection of different value of $\\alpha$ in Eq. (3).\n2.\tHow to select a public dataset used for prompt transfer? From Table 1, it seems like sometimes selecting a different public dataset affects the final test accuracy a lot.\n3.\tOther than next word prediction task, does the proposed method perform well on other language task? For example, question answering and reading comprehension?\n4.\tWhy there is no result for llama2-7b in Table 2?\n5.\tWhen comparing to the status of the art baselines, it seems like larger models have better zero-shot performance. In this case, in Table 4, can you include more results on larger models like llama2?\n6.\tFrom Fig. 5, it seems like the attack success rate is low even without DP. In this case, would that be the MIA algorithm used in the paper is not good enough? It seems like there is no motivation to add DP to protect private data. \n7.\tIf the original LLM is mixture of expert model, does the proposed framework apply to that case?" }, "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 a very interesting problem and proposed a novel framework to allow users to “perform soft prompt tuning” without revealing their private data and LLM providers not revealing their protected LLMs. The most novel part comes from privacy-preserving prompt transfer through public data, which largely protect data privacy and maintain inference performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper, the authors proposed a new framework that allows the users to perform private soft prompt tuning without accessing the original protected LLMs without leaking their private data. They proposed to first obtained a distilled model from the original LLM, a step needed from the LLM provider, and then the users will utilize this distilled model to perform soft prompt tuning and transfer the resulting soft prompt to prompt the original LLM. They conducted a few experiments to demonstrate the performance of their proposed framework and compare it with several baselines." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I listed a few weaknesses below:\n1. It is unclear how to select public dataset for prompt transfer, there is no guidance on how to select such a public dataset and how would the selection of public dataset affect the performance. From their experimental results in Table 1, it seems like different selection would make a large enough impact on the inference accuracy.\n2. No explanation on what would be the termination criterion for knowledge distillation step. The authors described that the distilled model should behave closely to the original model but not meeting the actual inference performance of the original model, which is too vague. I do not see either the ablation studies on how to select $\\alpha_{ce}$, $\\alpha_{lm}$ and $\\alpha_{cos}$.\n\nOverall, the main concern is that the proposed method contains many steps involving hyperparameters and public datasets need to be tuned and selected. More ablation studies and further theoretical analysis are needed to understand the general practical performance of the proposed method." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_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 don't understand the statement in this paper ``KD has been shown effective to compress LLMs during the pre-training phase while maintaining\ntheir performance (Sanh et al., 2019)\". The cited paper discusses BERT, which is not typically recognized as an LLM. Could you clarify this point? Additionally, can all types of LLMs (e.g., GPT, Llama, Mistral) be effectively compressed using knowledge distillation?\n\n\n\n2. This paper considers two baselines, namely DP-OPT (Hong et al., 2023) and Zero-Shot Transfer (Wu et al., 2023b). Hong et al. (2023) use the TREC, MPQA, and Disaster datasets, and Wu et al. (2023b) consider the LAMA dataset. It is unclear why these datasets are not included in the experiments of this paper. Could the authors also include these datasets in the experiments?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The related previous work and key concepts (e.g., prompt tuning, differential privacy, knowledge distillation, prompt transfer) are clearly explained. The proposed algorithm is described in detail." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies the problem of differentially private soft prompt tranfer. They propose the POST framework which consists of 3 steps: 1) The LLM provider compresses their model into a smaller model using techniques in knowledge distillation and then sends the distilled model to the user. 2) The user performs private prompt tuning using PromptDPSGD (Duan et al., 2024) on the distilled model. 3) The user transfers this prompt on the large LLM." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1, Motivation Unclear: This paper focuses on privacy-preserving soft prompt transfer, where the LLM provider is untrusted. Could you provide examples of closed-source LLM providers that support soft prompts? If, instead, the LLM provider is an open-source model that can be run locally (e.g., GPT-2 XL or Llama-2-7b, as mentioned in this paper), privacy concerns could potentially be mitigated by running the model locally. Could you clarify the motivation behind the need for privacy-preserving transfer in this context?\n\n\n\n\n2, Novelty Concerns: Given the existing work for knowledge distillation (Sanh et al., 2019) and PromptDPSGD (Duan et al., 2024), what is the technical novelty in this paper?\n\n\n\n\n3, Practical Concern: The proposed approach involves the LLM provider performing knowledge distillation to compress their LLM into a smaller model and then sending this distilled model to the user. Is this practical? It’s unclear if an LLM provider would be willing to share a distilled model, given proprietary concerns, and risks of potential attacks. \n\n4, Missing DP Guarantee: I recommend including a formal theorem that states the DP guarantee of the proposed method, and a proof to ensure the privacy guarantee is rigorously supported.\n\n\n\n5, Missing Experimental Details: The value of $\\delta$ used in this paper should be mentioned in the main paper. What is the standard deviation of the results in Table 1, 2 and 4? \n\n\n6, Some Typos:\n\n(1) Line 117: fist $\\rightarrow$ first.\n\n(2) Line 119: There’s an extra comma.\n\n\n\n(3) Line 132: from student to teacher $\\rightarrow$ from teacher to student.\n\n\n(4) Line 256: $\\mathcal{N}(0,\\sigma^2 c^2)\\rightarrow \\mathcal{N}(0,\\sigma^2 c^2 \\mathbf{I})$" }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose a method on how to transfer soft prompts tuned on a distilled small model to a larger model using public data." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024post,\ntitle={{POST}: A Framework for Privacy of Soft-prompt Transfer},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v6NNopExN4},\nnote={under review}\n}" }, "abstract": { "value": "Prompting has emerged as a dominant learning paradigm for adapting large language models (LLMs). While discrete (textual) prompts prepend tokens to the input for optimized outputs, soft (parameter) prompts are tuned in the embedding space via backpropagation, requiring less engineering effort. However, unlike semantically meaningful discrete prompts, soft prompts are tightly coupled to the LLM they were tuned on, hindering their generalization to other LLMs. This limitation is particularly problematic when efficiency and privacy are concerns, since (1) it requires tuning new prompts for each LLM which, due to the backpropagation, becomes increasingly computationally expensive as LLMs grow in size, and (2) when the LLM is centrally hosted, it requires sharing private data for soft prompt tuning with the LLM provider. To address these concerns, we propose a framework for Privacy Of Soft-prompt Transfer (POST), a novel method that enables private soft prompt tuning on a small language model and then transfers the prompt to the large LLM. Using knowledge distillation, we first derive the small language model directly from the LLM to facilitate prompt transferability. Then, we tune the soft prompt locally, if required with privacy guarantees, e.g., according to differential privacy. Finally, we use a small set of public data to transfer the prompt from the small model to the large LLM without additional privacy leakage. Our experimental results demonstrate that our method effectively transfers soft prompts, protecting local data privacy and reducing the computational complexity over soft prompt tuning on the large model." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "prompt transfer", "soft prompt", "privacy", "distillation", "confidentiality" ] }, "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/a0725135bbb852ab3f20b3caf6f16d76ebff91b9.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/357986e7541e58036b5abfb09e9a252bd8d0b114.zip" }, "title": { "value": "POST: A Framework for Privacy of Soft-prompt Transfer" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Can the authors calculate the full gradient when \"spurious alignment\" happens? Does that also align with the Hessian's top eigenspace? Or it has a very small similarity?\n2. Can the authors experiment with some kind of Gaussian noise for the toy example? And can you empirically show why they believe the stochastic noise has a larger bias in the top-eigenspace?\n3. For Appendix D.1, can the authors (1) report the learning rates (2) train with longer time s.t. EoS happens (3) add experiments with MSE loss?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This paper systematically investigates the phenomenon of gradient-Hessian alignment in various optimization algorithms, including SGD, GD in the Edge of Stability (EoS) regime, and the Sharpness-Aware Minimization (SAM) algorithm, with a particular emphasis on the analysis of SGD. Previous studies have primarily focused on understanding full-batch algorithms like GD or Adaptive GD without stochasticity. \n\nMoreover, this work takes an initial step toward understanding the 'spurious' alignment where the actual useful component of the gradient is the non-dominant part. They also claim that batch noise is the cause of this alignment by introducing a toy model to provide theoretical intuition. Though it is not very precise to say the mechanism is totally different between GD and SGD (see weakness), I believe the **\"ill-conditioned-valley\" view of loss landscape** captures why dom-GD doesn't decrease the loss but bulk-GD does." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates whether neural networks can be effectively trained by focusing only on the \"dominant subspace\", which is the top eigenspace of the Hessian of the loss, where the gradient seems to align with. The authors projects SGD updates onto this dominant subspace and find that cannot reduce the training loss further. In contrast, removing the dominant subspace from the updates proves equally effective for training, even though it projects out most of the original gradient component. The authors suggest that the alignment between the gradient and the dominant hessian subspace is \"spurious\". This \"spurious\" alignment appears consistently across various practical settings, such as the Edge of Stability, Sharpness-Aware Minimization, and when using momentum or adaptive optimizers. The authors also discuss the possible main causes of the phenomenon, and propose some toy models to understand it." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "First, the so-called 'spurious alignment' is long observed in EoS literature in my opinion. For example, Damian et al. [2] showed that in the EoS regime, (1) the gradient alignment happens (2) the loss decrement in the EoS regime depends on the constrained trajectory **by projecting out the top eigenspace**. It is exactly the finding listed in section 5.1 of this paper that bulk-GD is as effective as GD. I believe the authors should discuss those related works.\n\nThe author may argue that \"The mechanisms behind gradient alignment differ between GD with large learning rates and SGD with small learning rates.\" so their findings are novel. But that is also my primary concern of the paper's conclusion \"Stochastic noise is the cause of spurious alignment\". I agree that \"stochastic noise is part of the cause of spurious alignment\", but I don't think the alignment mechanism is intrinsically different between GD and SGD.\n\nIn section 4. The authors define the \"small LR regime\" and \"large LR regime\" of SGD and claim they are different by setting the threshold of $2/\\lambda_1$ of the learning rate. This threshold is for the GD descent lemma to hold. Then the authors claim in the small LR regime, the cause of the spurious alignment is the gradient noise. They 'verify' their findings by switching SGD to GD, and observe that the alignment disappears. However, [1] found that SGD begins to oscillate within the top eigenspace smaller than $2/\\lambda_1$. An intuitive explanation for this is **injected gradient noise increases the second order term of the Taylor expansion**, making the loss begin to oscillate **when the learning rate is below $2/\\lambda_1$** as the descent lemma predicted. In this case, GD is still stable and descent lemma holds, quickly converging back to the \"bottom of the valley\". That is why the alignment disappears after switching to GD, since after it's in the \"bottom of the valley\" the gradient component in the top eigenspace will be very small. \n\nTherefore, my argument is that stochastic noise indeed reduces the threshold of learning rate for oscillation and makes it easier to have spurious alignment. But the mechanism is the same: the self-stabilization effect induced by the gradient(+noise) within the top Hessian subspace. **To test this argument, the author should also calculate the full gradient when \"spurious alignment\" happens and see if the full gradient aligns with the top eigenspace. Also, the experiments should include various batch sizes to test different levels of gradient noise.** If my conjecture holds, I think this work can be seen as a generalized empirical validation of self-stabilization [2] for SGD, which somehow limits the novelty of this paper. I might also be wrong, so if the conjecture is not correct, I will raise my score.\n\nThe author also did some small learning rate GD in Appendix D.1 to corroborate their argument. I wonder what is the learning rate of the \"small learning rate\"? In the original EoS paper, most of the figures will exhibit EoS due to progressive sharpening when the model is trained for a long time. Also, the authors use the cross-entropy loss, where EoS may not happen due to some margin-maximization effect. What if the authors switch to square loss? I believe when you enter the EoS regime (when trained with enough time), the gradient alignment phenomenon will also appear.\n\nThe authors' toy examples constructed the two loss functions to introduce the gradient noise. However, I don't think this reflects the real-world gradient noise, since the 100xy term is artificially added to introduce gradient components in $x$-direction unless $y = 0$. But I think injecting some Gaussian noise biasing toward $x$-direction may also work for the toy case. It would be great to have some empirical evidence as to why the authors believe the gradient noise has a larger component in the top eigenspace.\n\n[1] Lee, Sungyoon, and Cheongjae Jang. \"A new characterization of the edge of stability based on a sharpness measure aware of batch gradient distribution.\" The Eleventh International Conference on Learning Representations. 2023.\n\n[2] Damian, Alex, Eshaan Nichani, and Jason D. Lee. \"Self-Stabilization: The Implicit Bias of Gradient Descent at the Edge of Stability.\"" }, "withdrawal_confirmation": null }, { "TLDR": null, "_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 experiments focus on the optimization comparsion between SGD, Bulk-SGD, and Dom-SGD. \nWhat are the generalization differences among these methods, particularly in the experiments conducted on MNIST and CIFAR, as shown in Figure 1?\n\n- Could the authors provide sufficient theoretical explanation for the main finding?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Many recent papers (Blanc et al., 2022; Li et al., 2022) highlight that the dynamics in bulk subspace (i.e., along flat directions) are crucial for SGD/SAM to move to flat minima, thereby improving generalization.\nIn contrast, this paper emphasizes the significant role of the dynamics in bulk subspace (i.e., along flat directions) in relation to optimization." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates that neural nets cannot be trained within the dominant subspace (i.e., along sharp directions)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "My primary concern are (i) the paper does not sufficiently explain the main finding, i.e., why only the dynamics in the bulk subspace are crucial for optimization, and (ii) the novelty of many contexts:\n\n- Section 4. This section focus on the alignment between the stochastic gradient and the sharp directions. However,\n\n - This section fails to adequately explain the main finding: why only the dynamics in the bulk subspace are crucial for optimization, expect for a very toy model.\n\n - Even regarding the alignment between stochatic gradient and the loss landscape, there are numerous previous works that address this point: $\\mathbb{E}[g_tg_t^T]\\approx 2 L(\\theta)\\nabla^2 L(\\theta_t)$ (Wojtowytsch, 2021; Mori et al., 2022; Wu et al., 2022; Wang and Wu., 2023), implying that stochastic gradient concentrate more along sharp directions of the landscape (i.e., the bump directions). It appears that the authors have overlooked these works.\n\n- Section 5. This section demonstrates that the gradient of GD aligns with the sharp directions in Edge of Stability (EoS), However, \n\n - The authors do not explain the main finding: why only the dynamics in the bulk subspace are crucial for optimization.\n\n - Regarding the alignment between the gradient and sharp directions, this has beed well studied in (Arora et al., 2022; Lyu et al; 2022; Damian et al., 2023) that this occurs due to approximate power iterations). However, the authors do not provide new 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": 3 }, "primary_area": null, "questions": { "value": "Why did the authors limit their experiments to a small subset of the datasets and can the authors share any insights for training with the full dataset?\n\nCould the authors provide any test accuracy results for Dom-SGD and Bulk-SGD to offer some insight into their impact on model generalization? When Bulk-SGD performs similarly to SGD in terms of training loss, does this trend hold for test accuracy as well?" }, "rating": { "value": 8 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "The paper addresses a question of significant interest in the ML community regarding SGD's effectiveness in low-dimensional subspaces. It helps address potential misconceptions arising from the well-cited work of Gur-Ari et al. (2018), which suggests that gradient descent would occur primarily within a tiny subspace.\n\nA convincing quadratic toy model effectively reinforces the authors’ interpretation of the empirical results, lending credibility to their main conclusions.\n\nThe paper is well-structured, with a logical flow that makes the argument easy to follow and the findings readily accessible to readers.\n\nGur-Ari et al. (2018): https://arxiv.org/abs/1812.04754" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper challenges existing assumptions about stochastic gradient descent (SGD) learning within a low-dimensional dominant subspace defined by the top eigenvectors of the Hessian of the loss. The authors find empirically that while gradients appear to align with this dominant subspace during training, projecting updates solely into this subspace (Dom-SGD) halts learning. Instead, projecting updates onto the orthogonal \"bulk\" subspace (Bulk-SGD) is as effective as standard SGD, suggesting that meaningful learning occurs outside the dominant subspace." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Although the paper includes experiments on three datasets, training is restricted to small subsets (e.g., 5,000 of 50,000 samples of CIFAR10) and primarily uses mean squared error loss instead of cross-entropy, despite focusing on classification tasks. This may limit the generalizability of the findings and should be communicated more clearly.\n\nThe paper exclusively examines the effects on training loss, with no analysis of test accuracy under Dom-SGD and Bulk-SGD. Including at least one plot of test performance would offer valuable insights into the practical implications for readers." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "Instead of computing the subspace based on the Hessian of the full loss, what if the subspace were computed based on the Hessian of the mini-batch loss. Would alignment still hold? Do the findings in the paper still hold?" }, "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 with a clear structure and compelling experimental results. The design of the experiments is sophisticated, covering both Dom-SGD and Bulk-SGD setups, along with more nuanced experiments like switching between SGD and GD during training. These additional experiments strongly support the paper’s claims, adding robustness to the argument. Additionally, the observations found in this paper is interesting and should contribute to a better understanding of the training dynamics of deep neural networks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper is built on recent studies that claim the gradients during neural network training live in a small subspace spanned by the top few eigenvectors of the Hessian. The authors decompose the gradients into the dominant component that lives in the dominant space and the bulk component that is orthogonal to the subspace. They find that dominant component of the gradient does not contribute to decreasing the training loss, while the bulk component, though only accounts for a small portion of the total gradient magnitude, is the main source driving the reduction in the loss. Through different experiments, the authors demonstrate that this observation holds for SGD, GD on the edge of stability, SAM, and adaptive optimizers. The authors also discuss potential explanations from a loss landscape perspective." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Major point:\n\n1.\tGeneralization: The authors acknowledge in the limitation that this paper does not consider the generalization part. However, in modern deep learning, generalization error is usually more important than training loss. And it is widely believed that SGD has an implicit bias that benefits generalization (arxiv: 1611.03530, 2006.08680, 2306.04251, etc.). The absence of results or discussion on generalization weakens the paper. Including an analysis of generalization effects could significantly enhance the paper's impact.\n\n2.\tBulk-SGD and noise: The results suggest that Bulk-SGD is less noisy than SGD (e. g. in Fig. 1) and Bulk-GD is less noisy than GD (Fig. 9). This reduction in noise may explain the speedup observed in Fig. 9 and 10. On the other hand, it is unclear how this will affect generalization. The authors have not adequately addressed this part in the paper, and I suggest including a more thorough discussion around that.\n\n3.\tDom-SGD and loss increase: It appears that with cross-entropy loss or using a standard architecture (VGG-11) (Fig. 14 and 17), the loss increases if training with Dom-SGD. This is a surprising finding and I think the authors should discuss this phenomenon more explicitly in the main text. Additionally, the absence of training curves for Bulk-SGD in these figures raises questions. Is there a reason for not including the curves for Bulk-SGD? I am also wondering whether this increase in the training loss relates to findings in prior works: arxiv: 2107.11774, 2306.04251.\n\n\nMinor point:\n\n1.\tFor results shown in table 1, the authors calculate the effective learning rates over the first 1000 steps. However, the authors also show that the alignment usually happens after some warm-up steps (e. g. Fig. 3). I think it makes more sense to plot the effective learning rates as a function over time (if noise is a concern, then perhaps run some kind of EMA). \n\n2.\tAlthough the toy quadratic model explains the observations well, the source of randomness seems a little contrived. I suggest incorporating the discussion of another model at least in the appendix, where the noise is Gaussian noise directly added on to the gradients, for a more natural comparison." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Deep neural networks cannot be trained within the dominant subspace, even though gradients align with this subspace along the training trajectory." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024does,\ntitle={Does {SGD} really happen in tiny subspaces?},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v6iLQBoIJw},\nnote={under review}\n}" }, "abstract": { "value": "Understanding the training dynamics of deep neural networks is challenging due to their high-dimensional nature and intricate loss landscapes. Recent studies have revealed that, along the training trajectory, the gradient approximately aligns with a low-rank top eigenspace of the training loss Hessian, referred to as the dominant subspace. Given this alignment, this paper explores whether neural networks can be trained within the dominant subspace, which, if feasible, could lead to more efficient training methods. Our primary observation is that when the SGD update is projected onto the dominant subspace, the training loss does not decrease further. This suggests that the observed alignment between the gradient and the dominant subspace is spurious. Surprisingly, projecting out the dominant subspace proves to be just as effective as the original update, despite removing the majority of the original update component. We observe similar behavior across practical setups, including the large learning rate regime (also known as Edge of Stability), Sharpness-Aware Minimization, momentum, and adaptive optimizers. We discuss the main causes and implications of this spurious alignment, shedding light on the dynamics of neural network training." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "optimization for deep networks", "training dynamics", "SGD", "Hessian", "low-rank subspace" ] }, "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/d6cf17930e00818bdbba62bd201f665a03858871.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/eaad4768ba931c00559a5c290bd97ba1a983dba7.zip" }, "title": { "value": "Does SGD really happen in tiny subspaces?" }, "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": "I submitted the wrong manuscript for the final version." }, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": null, "primary_area": null, "questions": null, "rating": null, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": null, "summary": null, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": null, "withdrawal_confirmation": { "value": "I have read and agree with the venue's withdrawal policy on behalf of myself and my co-authors." } }, { "TLDR": { "value": "A method to scale MoE models using structured modules" }, "_bibtex": { "value": "@misc{\nsarwar2024structmoe,\ntitle={StructMoE: Augmenting MoEs with Hierarchically Routed Low Rank Experts},\nauthor={Zain Sarwar and Ashwinee Panda and Benjamin Th{\\'e}rien and Stephen Rawls and Sambit Sahu and Supriyo Chakraborty},\nyear={2024},\nurl={https://openreview.net/forum?id=v71Nsh6R7m}\n}" }, "abstract": { "value": "The traditional approach to scaling Mixture of Experts for transformer models has been to increase the total number of experts. While performance improves with more experts, the gains are diminshing whereas\nmemory scales linearly with the number of experts. We introduce $\\textit{StructMoE}$, a scaling approach for Mixture of Experts which augments experts with additional dynamic capacity using routed structured matrices which we refer\nto as $\\textbf{L}$ow $\\textbf{R}$ank $\\textbf{E}$xprts ($\\textbf{$\\textit{LoRE}$}$). At a high-level, we introduce hierarchical MoEs where the first level of routing decides which expert each token\nshould be routed to and the second level of routing decides which $\\textit{LoRE}$ should each token be routed through. The outputs of the expert and the $\\textit{LoRE}$ are then entangled together to provide\nthe final output. This introduces more dynamism into the model which has empirically been demonstrated to improve model performance. We find this scaling approach to outperform a standard MoE baseline in terms of loss on a held out validation. Thus, we propose this to be an effective scaling technique for MoEs compared to the standard approach of adding more \nexperts to the model." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": { "value": [ "~Zain_Sarwar1", "~Ashwinee_Panda1", "~Benjamin_Thérien1", "~Stephen_Rawls3", "~Sambit_Sahu2", "~Supriyo_Chakraborty1" ] }, "authors": { "value": [ "Zain Sarwar", "Ashwinee Panda", "Benjamin Thérien", "Stephen Rawls", "Sambit Sahu", "Supriyo Chakraborty" ] }, "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": [ "moe", "mixture of experts", "LLM", "transformer" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": { "value": "sarwar|structmoe_augmenting_moes_with_hierarchically_routed_low_rank_experts" }, "pdf": { "value": "/pdf/12cb0452b8beb875e939d465e8c558127d8d0e97.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": "StructMoE: Augmenting MoEs with Hierarchically Routed Low Rank Experts" }, "venue": { "value": "ICLR 2025 Conference Withdrawn Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Withdrawn_Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 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 different training operations, such as collision handling and quantization, impact the computational time and resource requirements of the model?\n\n2. What methods or algorithms are employed to map these quantitative vectors back to specific items in the dataset?\n\n3. Why does the model not use multimodal information simultaneously to predict the next click item? What are the challenges or limitations that prevent this integration?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The approach to transform diverse item content from different domains and modalities into a unified quantitative language is highly innovative. This integration allows for a more robust and versatile recommendation system that can handle varied inputs effectively.\n\n2. The paper introduces a well-structured design of pre-training and fine-tuning tasks that includes not only generative prediction but also alignment tasks, enhancing the robustness of the method. This comprehensive approach allows the model to effectively leverage both generative capabilities and alignment strategies to improve overall recommendation accuracy.\n\n3. The proposed framework significantly outperforms existing models on several benchmark datasets, particularly in terms of the NDCG metric. This suggests that the method is not only theoretically sound but also practically effective." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper titled \"Multimodal Quantitative Language for Generative Recommendation\" introduces a novel approach to enhance generative recommendation systems by converting item content from multiple domains and modalities into a unified \"quantitative language\". This methodology seeks to bridge the gap between the generalized linguistic knowledge of pre-trained language models (PLMs) and the specialized needs of recommendation systems. The authors developed a new framework, MQL4GRec, which employs \"quantitative translators\" to convert textual and visual item data into a shared vocabulary. This shared language is then enriched with semantic information through various generation tasks to enable effective knowledge transfer from multimodal data to recommendation systems.\n\nThe paper's main contribution lies in its innovative method of integrating multimodal data to improve recommendation performance significantly, surpassing baseline methods by notable margins in terms of the NDCG metric across multiple datasets. Furthermore, the framework introduces a potential shift towards more universal recommendation systems that do not rely on traditional item IDs, thereby addressing common challenges like improving scalability and transferability across different domains." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. The paper does not discuss the impact of various training operations on the algorithm's time complexity, such as the handling of collisions. This oversight might leave questions about the scalability and efficiency of the proposed method in practical applications.\n\n2. The methodology section lacks a clear explanation of how the generated quantitative vectors are used to retrieve corresponding items during the next item generation task. This omission could lead to ambiguity regarding the operational specifics of the model.\n\n3. The model does not utilize multimodal information simultaneously to predict the next click item, which limits the paper's innovativeness and potential for expansion. Leveraging such multimodal data more effectively could enhance the model's predictive performance and applicability in more complex scenarios." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- User ID Collision: With only 2000 tokens representing a large user base, there is a potential for ID collisions, which may lead to inaccuracies in recommendation results in real-world applications.\n - Domain Adaptability: The model performs poorly on certain domain-specific datasets, such as the Games dataset, suggesting limitations in domain transferability." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. Innovation: The proposed MQL4GRec method translates content from different modalities into a unified “quantitative language,” enabling cross-domain and cross-modal recommendation knowledge transfer. This approach addresses limitations in handling multimodal data in existing generative recommendation models.\n\n\n2. Superior Performance: Experimental results on multiple public datasets demonstrate that MQL4GRec outperforms baseline methods on key metrics such as NDCG.\n\n\n\n3. Open-Source Availability: The paper provides a fully accessible code repository, facilitating reproducibility and further research within the community." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "MQL4GRec introduces a novel approach for generative recommendation by transforming multimodal content from different domains into a unified \"quantitative language,\" facilitating cross-domain knowledge transfer in recommendation tasks. The method uses quantitative translators for text and image content, building a shared vocabulary to encode semantic information across modalities. A series of language generation tasks further enriches this vocabulary, enhancing the model's capacity to represent multi-faceted user preferences. Experimental results demonstrate notable performance improvements over baseline models on key metrics across multiple datasets, showcasing MQL4GRec's scalability and potential in multimodal recommendation.\n\n\n\nHowever, its innovation appears limited, closely resembling existing methods like GenRet and MMGRec in both tokenization approach and generative structure." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Similarity to Existing Tokenization Approaches\n- The unified vocabulary for multimodal information closely resembles the generative retrieval tokenization approach found in \"Learning to Tokenize for Generative Retrieval,\" [1] particularly the \"GenRet\" model, which uses discrete auto-encoding for compact identifiers. This \"multimodal codebook\" seems to be an adaptation of single-modality tokenization, relying on established techniques like RQ-VAE and offering only incremental improvements without substantial architectural or performance innovation.\n- The motivational structure and initial figures in MQL4GRec are closely aligned with those in GenRet, which may diminish the perceived originality of the proposed approach.\n\n\n\n2. Lack of Comparative Analysis\n- MQL4GRec’s generative approach appears heavily inspired by MMGRec [2], which also employs Graph RQ-VAE for multimodal representation through user-item interactions, raising concerns about the uniqueness of MQL4GRec’s contributions. \n- The paper does not clearly distinguish MQL4GRec's advancements over MMGRec, especially in terms of multimodal token-based representations. A more thorough comparison is needed to establish any unique contributions beyond MMGRec’s existing framework.\n\n\n[1] Sun, W., Yan, L., Chen, Z., Wang, S., Zhu, H., Ren, P., ... & Ren, Z. (2024). Learning to tokenize for generative retrieval. Advances in Neural Information Processing Systems, 36.\n\n[2 ] Liu, H., Wei, Y., Song, X., Guan, W., Li, Y. F., & Nie, L. (2024). MMGRec: Multimodal Generative Recommendation with Transformer Model. arXiv preprint arXiv:2404.16555.\n\nThe paper does not clearly distinguish MQL4GRec's advancements over MMGRec, especially in terms of multimodal token-based representations. A more thorough comparison is needed to establish any unique contributions beyond MMGRec’s existing framework." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "* Figure 3 shows that the performance on the Games dataset slightly declines as the amount of pre-training data increases. Does this suggest that the pre-training strategy has limitations when applied to domains with substantial cross-domain differences?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "* High Innovation: This work is the first to propose using a unified quantitative language to address knowledge transfer in multimodal recommendation, which is highly valuable for enhancing the generalization ability of recommendation systems.\n\n* Thorough Experimental Validation: The paper conducts extensive experiments across three datasets, showcasing not only the overall performance of the method but also analyzing the role of individual components through ablation studies, indicating rigorous experimental design." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a novel multimodal generative recommendation method, MQL4GRec, which facilitates the effective transfer of recommendation knowledge by converting item content from different domains and modalities into a unified quantitative language. Specifically, MQL4GRec introduces a unified quantitative language representation to handle multimodal content, including text and images. Additionally, a series of quantitative language generation tasks are designed to enrich the semantic representation. Extensive experiments across three datasets show that this method significantly outperforms baseline approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The model's high complexity poses significant computational and storage demands, which could lead to considerable costs in real-world deployment." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "1.As an efficient framework, I am interested in understanding the training costs of MQL4GRec compared to other baselines. Specifically, how does it perform in terms of training time, VRAM usage, and inference time?\n\n2.In the field of recommender systems, there is a lack of widely recognized pre-trained models. Could MQL4GRec potentially serve as a foundation model for other downstream tasks?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. The proposed concept of a multimodal quantitative language, together with the design of quantitative language generation tasks, represents a novel and innovative advancement in the field of generative recommendation\n\n2. The architecture design is elegant, presenting the idea in a straightforward way, yet experiments demonstrate its strong effectiveness. I personally appreciate this type of work and believe it can make a meaningful impact in the field of generative recommendation.\n\n3. The availability of the code significantly enhances reproducibility." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a novel approach, MQL4GRec, designed to convert item content from diverse domains and modalities into a unified quantitative language." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1.The proposed framework is rooted in the generative recommendation paradigm and aligns with a preprint in a similar research direction [1]. However, it still represents a valuable contribution in my view, even if not groundbreaking.\n\n2.The authors should consider testing the statistical significance of MQL4GRec results.\n\n3.There remains a limitation in zero-shot capability, which is a known challenge in the field of recommendation.\n\n4.To enhance the comprehensiveness of the \"Multi-modal Recommendation\" section in the related work, the authors could consider including more recent state-of-the-art multimodal recommender system papers, such as [2,3]. The field of multimodal codebooks from other communities should also be included in the related work section to clarify the distinctions between the proposed MQL approach and existing methods, such as those in [4, 5]\n\n\n[1] Liu, Han, et al. \"MMGRec: Multimodal Generative Recommendation with Transformer Model.\" arXiv preprint arXiv:2404.16555 (2024).\n\n[2] Fu, Junchen, et al. \"IISAN: Efficiently adapting multimodal representation for sequential recommendation with decoupled PEFT.\" *Proceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval*. 2024.\n\n[3] Liu, Han, et al. \"MMGRec: Multimodal Generative Recommendation with Transformer Model.\" *arXiv preprint arXiv:2404.16555* (2024).\n\n[4] Lan, Zhibin, et al. \"Exploring better text image translation with multimodal codebook.\" arXiv preprint arXiv:2305.17415 (2023).\n\n[5] Duan, Jiali, et al. \"Multi-modal alignment using representation codebook.\" Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024multimodal,\ntitle={Multimodal Quantitative Language for Generative Recommendation},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v7YrIjpkTF},\nnote={under review}\n}" }, "abstract": { "value": "Generative recommendation has emerged as a promising paradigm aiming at directly generating the identifiers of the target candidates.\nMost existing methods attempt to leverage prior knowledge embedded in Pre-trained Language Models (PLMs) to improve the recommendation performance. However, they often fail to accommodate the differences between the general linguistic knowledge of PLMs and the specific needs of recommendation systems. Moreover, they rarely consider the complementary knowledge between the multimodal information of items, which represents the multi-faceted preferences of users. To facilitate efficient recommendation knowledge transfer, we propose a novel approach called Multimodal Quantitative Language for Generative Recommendation (MQL4GRec). Our key idea is to transform items from different domains and modalities into a unified language, which can serve as a bridge for transferring recommendation knowledge. Specifically, we first introduce quantitative translators to convert the text and image content of items from various domains into a new and concise language, known as quantitative language, with all items sharing the same vocabulary. Then, we design a series of quantitative language generation tasks to enrich quantitative language with semantic information and prior knowledge. Finally, we achieve the transfer of recommendation knowledge from different domains and modalities to the recommendation task through pre-training and fine-tuning. We evaluate the effectiveness of MQL4GRec through extensive experiments and comparisons with existing methods, achieving improvements over the baseline by 11.18\\%, 14.82\\%, and 7.95\\% on the NDCG metric across three different datasets, respectively. Our implementation is available at: \\href{https://anonymous.4open.science/r/QL4GRec-ED65/}{\\textcolor{blue}{https://anonymous.4open.science/r/MQL4GRec-ED65/}.}" }, "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": [ "Recommendation System", "Generative Recommendation" ] }, "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/8fe206daf248f9d554b8f72822516d89245de282.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": "Multimodal Quantitative Language for Generative Recommendation" }, "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": "- **History length**: How does the choice of history length L in the action-level dependency affect the performance of SWIRL? Did you experiment with values beyond 1 and 2, and if so, what were the findings? What trade-offs did you observe between increased history length and computational complexity? What guidelines would you suggest for choosing L in practice?\n- **Computational complexity**: Can you provide more details on the computational complexity of SWIRL compared to baseline models? How does it scale with the size of the dataset and the length of the history dependency? Could you provide specific runtime comparisons on the datasets used in the paper? Additionally, can you provide insights into the convergence properties of the EM algorithm?\n- **Hyperparameter selection**: How were hyperparameters, such as the temperature parameter α in the soft-Q iteration, selected? Was any hyperparameter tuning performed, and if so, what criteria were used? Did you employ any cross-validation procedures to ensure the robustness of the results? How sensitive is the model to the choice of initial parameters?\n- **Limitations**: Could you elaborate on any limitations of SWIRL in modelling certain types of animal behaviours? Are there situations where history dependency might not adequately capture the decision-making processes? How might SWIRL perform on behaviours with very long-term dependencies that extend beyond the history length L? Also, does this framework work in both discrete and continuous state/action spaces?\n- **Robustness to noisy data**: How does this framework handle noisy or incomplete data, which are common in real-world animal behaviour datasets? Did you assess the robustness of the model under such conditions?\n- **Minor:** line 83: In the experiment \"section?\" - or \"In the Results section,\"\n- **Discussion suggestion:** Could this framework be used to provide evidence whether models of intrinsic reward (e.g. expected free energy or empowerment) are indeed able to capture animal behaviour?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- **Originality/innovation**: The paper presents a novel extension to IRL by incorporating history-dependent reward functions, addressing a significant gap in modelling complex, naturalistic animal behaviours. This is the first IRL model to integrate both decision-level and action-level history dependency.\n- **Quality/empirical validation**: The authors provide a thorough mathematical formulation of SWIRL, including detailed explanations of how history dependency is incorporated at different levels, and a clear demonstration of improvements over baseline methods. The choice of the authors to use both simulated and real-world datasets strengthens the validation of their approach.\n- **Clarity**: The paper is generally well-written, with clear explanations of the concepts and methods. The connection drawn between SWIRL and traditional autoregressive dynamics models helps to contextualize the work within existing literature.\n- **Significance**: The SWIRL framework offers a more accurate model of animal decision-making. Hence, I believe it has the potential to advance our understanding in neuroscience and behavioural science, opening up new ways for analysing long-term, complex behaviours driven by intrinsic motivations. Finally, the presented experiments are (in theory) reproducible with public datasets and publicly available code." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper introduces SWIRL (SWItching IRL), a novel framework that extends inverse reinforcement learning (IRL) by incorporating time-varying, history-dependent reward functions to model complex animal behaviour. Traditional IRL methods often assume a static reward function, limiting their ability to capture the shifting motivations and history-dependent decision-making observed in animals. SWIRL addresses this limitation by modelling long behavioural sequences as transitions between short-term decision-making processes, each governed by a unique reward function. It incorporates biologically plausible history dependency at both the decision level (transitions between decision-making processes depend on previous decisions and environmental context) and the action level (actions depend on the history of states within a decision-making process). The authors apply SWIRL to simulated data and two real-world animal behaviour datasets, demonstrating that it outperforms existing models lacking history dependency, both quantitatively and qualitatively. They also highlight connections between SWIRL and traditional autoregressive dynamics models, arguing that SWIRL offers a more generalized and principled approach to characterizing animal behaviour.\n\nI think this is a very interesting and well-written paper. I gave a score of 6 but I am willing to reconsider this score if the authors can adequately address my concerns, particularly regarding the methodological details, hyperparameter selection, and theoretical analysis." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Methodology**: Some aspects of the implementation, such as hyperparameter selection and the specifics of the inference algorithm, are not fully detailed. Providing more information on these would enhance reproducibility and allow for better assessment of the method. There is limited analysis on hyperparameter sensitivity or discussion of how to choose the history length L in practice. In addition, the impact of number of hidden modes is not thoroughly explored. Finally, there is a lack of a theoretical analysis that would strengthen the paper, such as providing convergence guarantees, a discussion on optimality conditions.\n- **Scalability**: The computational complexity of SWIRL, especially with history dependency and the EM algorithm, may pose challenges for large datasets. The paper would benefit from a discussion on scalability to larger state/action spaces and potential optimization strategies. Could also benefit from runtime comparisons with baseline methods.\n- **Biological plausibility**: While the model is said to incorporate biologically plausible history dependency, the paper could provide more evidence or discussion on the biological validity of the specific mechanisms used." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "* Why is the problem modelled as a hidden-mode MDP rather than a POMDP or Hierarchical RL setting?\n\nOverall I think this paper has potential, and if the issues with the experimental validation and related work discussed above are corrected I would be happy to increase my score." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "* The method seems to work well and produces nicely interpretable result on the real world dataset.\n * The paper is well written and reads quite nicely." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper addresses the problem of inverse RL in a hidden-mode MDP, i.e. an MDP with an additional hidden-mode parameter that affects the reward. The authors propose an EM-style algorithm that learns both the reward function/policy and hidden mode in the given expert trajectories.\nThe authors then validate their approach on a synthetic gridworld task and go on to use it to model animal behavior in a rat maze, where the hidden mode represents the rat's current objective (i.e. get water, explore, go home)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "* The literature review is limited. For example, a different interpretation for this problem setting would be a POMDP, for which previous IRL literature exists, e.g. [1]. Yet another interpretation of the mouse experiments would be a hierarchical RL setting, for example an Option setting in which options might be \"get water\" \"explore\" \"go home\". For this setting previous IRL method exist as well, e.g. [2]. I'm not sure if they are directly applicable to this paper's problem setting, but I think they might be applicable?\n * The method seems to be limited to relatively small discrete state and action sets, limiting it's general applicability.\n\n**The presentation of the experiments is misleading**\n * The most competitive baseline was labeled as \"I-1\" in plots, while the poorly performing baselines are labeled with their names. This might lead to it being confused as the author's contribution and is thus highly misleading and should definitelyl be corrected.\n * It is not clear how often each experiment was run, or how the box plots were created. How were outliers selected? Figure 3E eliminates the best result for baseline I-1.\n * It is also not clear what shaded areas in Figure 4B represent.\n * The MaxEnt baseline is missing in Figure 4? Why?\n\nMinor points\n * L453 refers to an appendix which seems to be missing?\n * L202 $\\xi$ is never defined in the paper?\n\nReferences: \n[1] Choi et al \"Inverse Reinforcement Learning in Partially Observable Environments\", JMLR, 2011\n[2] Chen et al. \"Option-Aware Adversarial Inverse Reinforcement Learning for Robotic Control\", ICRA, 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": "**Complexity and scalability of the algorithm**\n\nOne major issue with variants of MaxEnt IRL methods (even with fixed reward function) is that the computing complexity would be very high since within each loop one needs to solve a forward problem to obtain the policy under the current reward function estimation. This problem would be further scaled up when there are multiple reward functions that needed to be evaluated via solving a RL problem (e.g., in equation (6) of this paper). So my question would be:\n\n1. Is SWIRL guaranteed to converge under some large environments? For example, if the size of the simulated gridworld environment in section 4.1 is not $5 \\times 5$ but, say, $50 \\times 50$?\n2. Following the last question, what is the relationship between computing time of SWIRL and the cardinality of the state, action, and the hidden mode space? This would be helpful to understand what is the maximum capacity of this method, i.e., to which scale of the environment that SWIRL can still handle. To address these questions, it would be helpful if the authors can provide some empirical results on the computing time of SWIRL under different setups.\n3. I believe this question would be more suitable for a future work, but just out of curiosity, would SWIRL be possible to integrate some function approximations (of e.g., the state, action, or hidden mode space, since it is now already using gradient method for optimization) so that it can still be applied in high-dimensional tasks?\n\n**Experiments**\n\n1. How did the author choose the number of hidden modes? In the three experiment discussed in this paper, the total number of hidden modes varies across experiment (2 for the first experiment, 3 for the second, and 5 for the last), so I would be curious to know what is the criterion of the authors' to select this hyperparameter. Or did I missed some part such that this number is learnt by the algorithm automatically?\n2. The discussion about the reward maps in section 4.3 (lines 492--505) is very hard to follow, as well as the related figures (Figure 4A). In general, I may propose that the results shown in Figures 4A and 4C and corresponding text do not really help for supporting the major claim about SWIRL. This could be due to the lack of space so that lots of experiment details are omitted, and this part would be more suitable for a scientific journal but not a conference. I may recommend the author to change the way of presenting the results.\n\n**Math writing**\n\nSome modifications that would be helpful to increase the readability of the paper are:\n\n1. Numbering of equations in the paper is sloppy, e.g., all equations in the paper are labeled, though some of them are neither referenced nor needed for further discussion. I would recommend labeling only those equations that are referenced later. For example, in lines 660--671, four lines of a single equation is labeled \"(8)\", but then in lines 685--698, all lines within the same expression are given a different label from \"(10)\" to \"(15)\". Or are there any special meanings that I missed?\n2. Some display-mode equations are missing punctuation.\n3. Line 136, \"$R$ corresponds to the reward function $r \\in \\mathbb{R}$\": What is the difference between \"$R$\" and \"$r$\" in this notation? According to the rest of the paper (where the notation \"$R$\" never occurrs again), I would assume they both represent the reward function, then why introducing two different symbol for the same meaning? Besides, the notation \"$r \\in \\mathbb{R}$\" defines the reward function \"$r$\" as a real number, but subsequent reference of \"$r$\" indicates that it is actually a function on the cartesian product of the state space and action space into the real line. I would recommend make sure the definition of \"$r$\" is consistent through the paper for a better clarification.\n\n4. Line 170: The symbol \"$\\mathcal{S}^L$\" is used without definition. Does that mean the cartesian product of $L$ state spaces?\n5. Line 231, \"$\\forall s \\in \\mathcal{S}, z \\in Z$\": Did you mean \"$z \\in \\mathcal{Z}$\"?\n\n6. In the appendix, I only see section A.1.1 but no further sections. Is there are missing section after that? If not, I would suggest remove the subsubsection label for this part.\n\n7. In the derivations in the appendix, the range of some summation is not consistent and not clear. For example, in lines 661--671, the summation is given by \"$\\sum_{n}$\", but in subsequent lines (e.g., line 677) the same summation (I assume) is given by \"$\\sum_{n = 1}^N$\". Is \"$\\sum_{n}$\" just a slang for \"$\\sum_{n = 1}^N$\", or the range in the two sums are actually different?\n\n8. Line 701--703: The notation \"$Q(\\theta, \\theta^k)$\" is used without definition. According to the context, I assume it is a typo and the authors would have really meant \"$G(\\theta, \\theta^k)$\".\n\n\n**Minor**\n\nIn Figure 3, the color for mode \"water\" and \"explore\" is very hard to distinguish if the reader only has access to the printed paper (especially in Figure 3F). Similar issue exists in Figure 4, but not so severe. Consider revising." }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "This work is well motivated by recent advances in applying IRL methods for characterizing animal complex behavior. By incorporating history-dependent policies and rewards, SWIRL outperformed the state-of-the-art in understanding animal decision-making processes." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces a framework called SWIRL that extends traditional IRL to capture the complex, long-term decision-making processes of animals in natural environments. SWIRL incorporates biologically plausible history dependency at both the decision-level and action-level, allowing it to better account for how past decisions and environmental contexts shape current behavior. The algorithm is evaluated on both synthetic and real-world datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "•\tNovelty: Insufficient comparison with Nguyen et al.'s 2015 work, which leaves SWIRL's unique contributions unclear.\n\nThe difference between SWIRL and the previous work by `Nguyen, Quoc Phong, Bryan Kian Hsiang Low, and Patrick Jaillet. \"Inverse reinforcement learning with locally consistent reward functions.\" Advances in neural information processing systems, 28 (2015)` needs further discussion to clarify the novelty of this work.\n\nUnder the similar IRL framework with multiple locally consistent reward functions, Nguyen's algorithm proposed that the transition kernel of reward functions can be dominated by some external inputs. In this case, could SWIRL then be considered as a special case of Nguyen's algorithm with the external input defined as $(s_1, \\ldots, s_L)$? Although the RL inner-loop in the two algorithms are slightly different, but I suppose it is not the major difference between SWIRL and Nguyen's work.\n\nIt would be helpful if the authors could provide a more detailed comparison between these two algorithms, as well as some additional experiments to compare the performance between them.\n\n•\tUnclear scalability of SWIRL, particularly in large environments or with high-dimensional tasks.\n\n•\tLack of transparency in choosing the number of hidden modes across experiments.\n\n•\tThe math writing of this paper is a bit difficult to follow (see below)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We develop an novel inverse reinforcement learning framework that can model the history-dependent switching reward functions in complex animal behaviors" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024inverse,\ntitle={Inverse Reinforcement Learning with Switching Rewards and History Dependency for Characterizing Animal Behaviors},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v7a4KET0Md},\nnote={under review}\n}" }, "abstract": { "value": "Traditional approaches to studying decision-making in neuroscience focus on simplified behavioral tasks where animals perform repetitive, stereotyped actions to receive explicit rewards. While informative, these methods constrain our understanding of decision-making to short timescale behaviors driven by explicit goals. In natural environments, animals exhibit more complex, long-term behaviors driven by intrinsic motivations that are often unobservable. Recent works in time-varying inverse reinforcement learning (IRL) aim to capture shifting motivations in long-term, freely moving behaviors. However, a crucial challenge remains: animals make decisions based on their history, not just their current state. To address this, we introduce SWIRL (SWItching IRL), a novel framework that extends traditional IRL by incorporating time-varying, history-dependent reward functions. SWIRL models long behavioral sequences as transitions between short-term decision-making processes, each governed by a unique reward function. SWIRL incorporates biologically plausible history dependency to capture how past decisions and environmental contexts shape behavior, offering a more accurate description of animal decision-making. We apply SWIRL to simulated and real-world animal behavior datasets and show that it outperforms models lacking history dependency, both quantitatively and qualitatively. This work presents the first IRL model to incorporate history-dependent policies and rewards to advance our understanding of complex, naturalistic decision-making in animals." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "neuroscience", "decision-making", "inverse 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/4d53fc3b184ea24ae9c79b753b8f2e0059aed139.pdf" }, "presentation": null, "primary_area": { "value": "applications to neuroscience & cognitive science" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "Inverse Reinforcement Learning with Switching Rewards and History Dependency for Characterizing Animal Behaviors" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 4 }, "primary_area": null, "questions": { "value": "This is a list of minor comments, questions and suggestions:\n\n- On line 351: \"an experienced researcher adjudicating the annotation by selecting the better analysis and making further refinements\".\nWhat do these \"annotations\" refer to? I assume that these doesn't refer to gene annotations \n(correct me if wrong). Without this information is really hard to evaluate what \"Inter-Annotator Agreement (IAA)\" refers to.\n\n- On line 418, \"complexity of clinical data extraction and the need for nuanced knowledge inference\". Could you please elaborate more on what does this mean?\n\n- On line 424, \" allowing several LLMs or agent-based models to achieve decent performance\".\nIt seems to me that in the previous tasks, different models could also perform differently. Please, correct me if wrong.\n\n- What does \"merged data\" refer to in Table 4?\n\n- In Table 5, what is really the difference between GenoAgent (based on GPT-4o) and GPT-4o itself?\nI understand that the former doesn't have agents involved, is this correct?\n\n- It would have been great to comment on the \"Statistical analysis\" results reported.\nWhat were the expected results? Are the results reported positive?\n\n- Is \"Dataset selection and filtering\" with GenoAgent performed entirely with metadata\nfrom the datasets only?\n\n- On line 395: \"measuring their performance in gene identification from raw input data\"\nWhat does \"raw input data\" refer to here?\n\n- I am assuming that the goal of \"end-to-end performance\" in Section 5.1\nis to measure the performance in identifying significant genes related to traits.\nThen I assume that this is a multi-label classification problem. Is this correct?\nIf so, how many classes are considered for the results? How imbalanced is the dataset?\n\n- From Table 5, is there any advantage here to using GenoAgent against the other models?\n\n- On line 417: \"However, preprocessing of trait data was significantly\nweaker, with a CSC score of 32.28%, due to the complexity of clinical data extraction and the need for nuanced knowledge inference\"\nCould you please elaborate more on this?\n\n- What is the total number of datasets in the results of Table 2?\n\n- Table 2 says that it reports \"F1 and Accuracy\" for \"DF\" and \"DS\", which one of the two is reported?\n\n- About the results reported in Table 4, I assume that the results are indeed good,\nbut without any context or reference it is hard to evaluate whether CSC of 79.71 is \na good or a bad result. Would it be possible to provide some context on this?\n\n- This is not an issue, but in Table 5, results with \"Llama 3 (8B)\" are reported. While it is great to have a comparison of\ndifferent models, Llama 3 (8B) is known to be comparably worse than e.g. GPT-4. The results reported with this model are not really very informative. For more significant results, a more powerful model of the Llama series, e.g. Llama 3 (70B), or Llama 3.1 (70B), that is claimed to be on par with close-models, should have been used.\n\n- On line 52 \"However, these studies have mostly focused on simplified synthetic datasets\"\nCould the authors please provide some references in this regard?\n\n- On line 150: \"The agent selects the optimal tool by minimizing both time and error\"\nCould the authors please explain why (and how) is time and error minimized?\n\n- It seems to me that the reference \"BPC (2023)\" does not belong to a serious\nscientific work. I am sure that other references could be used here.\n\n- The reference used for Llama 3 seems to be incorrect, and is incorrectly spelled \nas \"Lamma\" instead of \"Llama\". The technical report where these models \nwere reported can be found at https://arxiv.org/pdf/2407.21783.\n\n- The presented work seems to have potential ethical issues that may have\nnot been addressed or at least mentioned in the paper. For instance, since\nthe authors propose a potential method for predictive medicine, it is \nimportant to note that a method with poor accuracy could lead to wrong\ndiagnosis and treatment. Besides this, is there any potential risk\nin LLMs leading to biases in genomic analysis?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 4 }, "strengths": { "value": "The paper is generally well-written, and in good English. The motivation\nof the problem is clearly stated and the related work is well covered.\nThe method proposed is really innovative and promising. Additionally,\nthe authors define carefully the tasks undergone to curate the dataset\nand the steps reported seem reasonable. Generally, the results are positive\nand I believe that the paper has the potential to be a great contribution\nto the field." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work introduces GenoAgent and LLM agent-based framework for gene expression\nanalysis tasks. This framework mainly consists of a project manager agent,\ndomain expert, data engineer, code reviewer and statistician agents.\nGenerally, the method is able to solve trait-related questions from raw data\nby leveraging statistical (code-based) tools. To evaluate the proposed\nmethod, the authors curate the GenoTEX benchmark, a benchmark of unconditional\nand conditional pairs of traits genes, which is also manually adjusted by experts. \nThe results generally show great performance of the proposed method in the\ntasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "I am mainly concerned about two major issues in the paper:\n\n- The first one is regarding the availability and clarity of the reported\nmethods. The authors claim that the GenoTEX benchmark will be a great \ncontribution to the field. Does this mean that the benchmark will be\npublicly available? If so, it would have been great for the authors\nto provide (if possible) an anonymized link to the benchmark.\nRegarding source code, the authors do not mention if the source code\nwill be available and no link/supplementary material is provided.\nAgain, it would be very positive for the authors to release\nthe source code for the proposed method, as this would allow for\nthe reproducibility of the results. Otherwise, there are parts of the \nwork that might seem obscure to the reader, e.g. what are the statistical\ntools used by the agents and how? How are the agents and their communication implemented?, etc.\n\n- The second issue is regarding the clarity of the results. The authors\nprovide a good description of the curation process of the benchmark,\nhowever, the results provided are not very detailed, which makes it \nhard for the reader to understand wether the authors achieved the\ngoals of the paper and how well the proposed method performs.\nI would suggest the authors to provide a more clear, detailed and\na comprehensive description of the results, including a more detailed \ndescriptions of the tasks at hand, the significance of the metrics used\nand to provide insightful comments that could serve the community\nor the reader of the paper. Please see the questions below for a few examples\nof the issues regarding the clarity 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": 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. Have you tried your system with different backbones? You have claimed that you use GPT-4o as the backbone LLM. I’d like to know the performance of Llama3-8B or other open-source models. \n2. How could you ensure no data leakage on the gene identification problems? Could you show the comparison experiments between your agent and other LLMs? You have claimed that MetaGPT cannot generate runnable code for the preprocessing of gene data. How could you make sure that your model could generate runnable code? Only by iterations and reviews? In addition, since the code reviewer is still based on LLMs, this agent is also based on the ability of LLMs. It doesn’t contain any fine-tuning stages, so how could you make sure your agent could generate runnable code?\n3. How about the success rate for generating code? \n4. In addition, the authors have claimed a set of function library L. Is this process automatic? Or just by prompting LLMs?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "1. Proposed an agent system to explore gene expression datasets.\n2. Proposed a benchmark dataset, GenoTEX." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "They introduce GenoAgent, a team of LLM-based agents designed with context-aware planning, iterative correction, and domain expert consultation to explore gene datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The experiments are not enough." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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. The paper mentions that feedback from the Code Reviewer is sometimes inconsistent. How frequent are these inconsistencies, and how do they typically impact GenoAgent’s task outcomes? Could the authors provide data on the frequency of conflicting or erroneous feedback in the code review process that leads to downstream errors in the analysis?\n\n2. A comparative table or discussion on how GenoAgent specifically improves upon or extends prior work would help emphasize its novelty.\n\n3. A discussion of runtime, memory usage, or potential computational optimizations (such as modularizing tasks or limiting interactions between agents) would aid in understanding the feasibility of GenoAgent for widespread use. Since the analysis pipeline can be resource-intensive, identifying strategies to minimize costs would be crucial for scaling GenoAgent in practical environments.\n\n4. While the paper includes error analysis, it could benefit from a deeper examination of errors in statistical analysis and preprocessing steps. Are there specific types of errors (e.g., issues with gene symbol normalization or confounding factor adjustments) that are more prevalent, and what are the proposed fixes?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1.\tThis paper is among the first to apply an LLM-based multi-agent system to gene expression data analysis, simulating collaborative workflows typical in bioinformatics teams.\n\n2.\tDataset Contribution: The GenoTEX benchmark provides a valuable resource for future research in AI-driven genomics data analysis, offering a standardized framework for evaluating model performance." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents GenoAgent, a novel framework leveraging a team of LLM-based agents to automate the exploration and analysis of gene expression data, addressing challenges in scalability and expertise demands in genomics. Each agent in GenoAgent has a specialized role—such as data engineering, statistical analysis, code review, and domain expertise—mimicking a collaborative bioinformatics workflow for tasks like dataset selection, preprocessing, and statistical analysis. Additionally, the authors introduce GenoTEX, a benchmark dataset designed for evaluating and advancing automated methods in gene analysis. Experiments demonstrate that GenoAgent achieves promising accuracy in gene identification, with iterative error correction and expert consultation mechanisms enhancing its overall performance, while GenoTEX provides a resource for consistent, real-world evaluation of automated genomics tools." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. GenoAgent has been tested on benchmark datasets, but applying it to novel, unseen datasets in real-world genomics research to assess its robustness under different conditions would enhance the effectiveness of this work.\n \n2. While GenoAgent introduces a structured, team-based approach to LLM-driven gene expression analysis, the novelty of its methodology could benefit from deeper differentiation from existing frameworks in multi-agent LLM systems and automated bioinformatics workflows." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "Q1: How reliable is prompt engineering since it does not really eliminate the possibility of hallucinating?\nQ2: Can improving the preprocessing of clinical data improve the performance? If yes, what strategies have the authors considered?" }, "rating": { "value": 3 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "S1: The authors use different agents that are responsible for different roles in the GenoAgent pipeline. These agents are LLM-based and the authors have clearly defined the responsibilities each “agent” would undertake such as developing guidelines or reviewing historical actions and current context.\nS2: There are specialized roles in the pipeline as well – “Project Manager”, “Data Engineer”, “Statistician”, “Code Reviewer” and “Domain Expert”. These roles bring about modularity in GenoAgent which in turn would help with issues like troubleshooting.\nS3: The creation of the GenoTex benchmark is novel in its contribution to the genomic data analysis since it comprises of gene identification problems, data from open gene expression databases, manual analysis data, quality control and assessment." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The authors in this paper present GenoAgent which comprises of LLM-based agents that are responsible for different roles. They also present a benchmark, GenoTEX for automatic exploration of gene expression data and use this to assess different tasks such as automation in genomics. They report F1, Precision, Recall, Accuracy and Jaccard similarity between gene sets (trait and condition)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "W1: The performance metrics, especially using F1 and accuracy are not very expressive in understanding just how good the pipeline is. F1 can sometimes not fully capture data filtering requirements and this is very crucial in genomics.\nW2: There are several ambiguous statements, ex: “the process of gene expression data analysis with good overall Accuracy” or “The agents show decent performance, likely” which do not correctly reflect the accuracy or performance.\nW3: The feedback mechanism does not seem entirely reliable in the Code Reviewer agent, especially when it shows diminishing performance after the first feedback round. This issue is quite critical when it comes down to understanding the robustness of GenoAgent.\nW4: Data normalization can be further refined to reduce variability for traits. \nW5: While GenoTEX benchmark is novel, the authors do not provide an exhaustive comparison with any existing domain-specific methods. \nW6: What are the computational costs for using computationally expensive models like GPT-4o and other LLMs? To add to this, all the agents are LLM-based. Have the authors kept this in check?\t\n\nTypos:\n1. Line 57, why are “D” in “Data”, “Auto” in “Automatic” and “E” in “Exploration” made bold when it is not used in the abbreviation “GenoAgent”?\n\nReferences:\n1. Line 501 (part of the url) is spilling outside the margin. \n2. Incorrect citations for several references (ex: lines 496, 521, 540, 575, 579, 591, 658, 694 ). I have listed some of the corrected references at the bottom (Harvard format) to give the authors an example.\n3. Citation format inconsistent (and possibly incorrect) in line 593.\n4. There is inconsistency in the format for the references in general. \n5. Medium articles are often unchecked, unverified facts wherein the scientific rigor can be easily questioned. Authors have cited some, like “BPC (2023)” on line 38/39.\n\nLine 496: Besta, M., Blach, N., Kubicek, A., Gerstenberger, R., Podstawski, M., Gianinazzi, L., Gajda, J., Lehmann, T., Niewiadomski, H., Nyczyk, P. and Hoefler, T., 2024, March. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 16, pp. 17682-17690).\nLine 521: Dong, Y., Jiang, X., Jin, Z. and Li, G., 2024. Self-collaboration code generation via chatgpt. ACM Transactions on Software Engineering and Methodology, 33(7), pp.1-38.\nLine 540: Guo, T., Nan, B., Liang, Z., Guo, Z., Chawla, N., Wiest, O. and Zhang, X., 2023. What can large language models do in chemistry? a comprehensive benchmark on eight tasks. Advances in Neural Information Processing Systems, 36, pp.59662-59688.\nLine 575: Ma, P., Ding, R., Wang, S., Han, S. and Zhang, D., 2023, December. InsightPilot: An LLM-empowered automated data exploration system. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing: System Demonstrations (pp. 346-352)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "A Baseline method for LLM-Based Exploration of Gene Expression Data in Alignment with Bioinformaticians" }, "_bibtex": { "value": "@inproceedings{\nanonymous2024genoagent,\ntitle={GenoAgent: A Baseline method for {LLM}-Based Exploration of Gene Expression Data in Alignment with Bioinformaticians},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v7aeTmfGOu},\nnote={under review}\n}" }, "abstract": { "value": "Recent advancements in machine learning have significantly improved the identification of disease-associated genes from gene expression datasets. However, these processes often require extensive expertise and manual effort, limiting their scalability. Large Language Model (LLM)-based agents have shown promise in automating these tasks due to their increasing problem-solving abilities. To leverage the potential of agentic system, we introduce GenoAgent, a team of LLM-based agents designed with context-aware planning, iterative correction, and domain expert consultation to collaboratively explore gene datasets. GenoAgent provides generalized approach for addressing a wide range of gene identification problems, in a completely automated analysis pipeline that follows the standard of computational genomics. Our experiments with GenoAgent demonstrate the potential of LLM-based approaches in genomics data analysis, while error analysis highlights the challenges and areas for future improvement. We also propose GenoTEX, a benchmark dataset for automatic exploration of gene expression data, and also a promising resource for evaluating and enhancing AI-driven methods for genomics data analysis." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Multi-agent", "Bioinformatics" ] }, "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/417ff9c0b9dd312f1b37fe16d885b0084c3f841e.pdf" }, "presentation": null, "primary_area": { "value": "applications to physical sciences (physics, chemistry, biology, etc.)" }, "questions": null, "rating": null, "reciprocal_reviewing": { "value": "I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6." }, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": null, "strengths": null, "student_author": null, "submission_guidelines": { "value": "I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide." }, "summary": null, "supplementary_material": null, "title": { "value": "GenoAgent: A Baseline method for LLM-Based Exploration of Gene Expression Data in Alignment with Bioinformaticians" }, "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": "- I kindly suggest resolve the confusions mentioned in the Weakness part for the \noverall clarity of the work.\n- While Figure 7 is placed in the appendix, it can effectively offer insights into\nthe evolution of the embedding distribution. I kindly suggest to put it to the\nmost visible point in the paper and compare the evolution process of the ITM and\nother existing transferability estimation approaches to demonstrate the\nsuperiority of the proposed framework visually.\n- It might be better to move the Table 4 into the main content to better support\nthe effectiveness of the proposed framework.\n- Since LEAD is cited and can outperform the evaluated PED and SFDA, I am interested\nin whether IMT could outperform it.\n- What would happen if n is not fixed to 1? Would this lead to better results at\nthe cost of higher computational efficiency? I kindly suggest to carry out\nexperiments to discuss this in more detail, as this iterative optimization is the\ncore part of the DEA.\n- The PCO strategy uses a strong inductive bias that the static distribution of \neach class should be well-separated. Since this might not be always valid, I am\nvery interested in the applicability of IMT to other computer vision tasks like\nimage segmentation. Would there be any modifications to the framework to facilitate\napplying it to other downstream tasks?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- The proposed Implicit Transferability Modeling (ITM) framework introduces a fundamentally new approach to transferability estimation, distinguishing itself from existing models such as PED, SFDA, and LEAD. ITM not only improves model effectiveness but also addresses computational efficiency through the Divide-and-Conquer Adaptation (DCA) and Pseudo-Clustering Optimization (PCO) strategies, which optimize processing time and resource usage.\n- The work further evaluates the applicability of transferability estimation on advanced architectures like Vision Transformers (ViT) and cutting-edge pre-training paradigms (e.g., MAE, SimMIM). This alignment with recent advancements in computer vision enhances the relevance and applicability of ITM within the current research landscape.\n- The work is strengthened by rigorous mathematical derivations, which provide theoretical underpinnings that enhance clarity and make the methodological flow accessible and easy to follow." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper presents a novel implicit transferability modeling approach to address the challenges in estimating transferability of pre-trained models due to the varied architectures and training strategies. By incorporating newly proposed modeling and optimization strategies, the resultant framework demonstrates superior performance than existing methods across ten datasets with various model architectures and pre-training paradigms without the need of extensive training and computational resources." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Figures are not well-used in the paper.\n - Too many figures (i.e., Figure 1 and 2) are used to describe the well-known problems in the field of transferability estimation.\n - Figure 1 is hard to understand at the first glance, particularly when it includes many abbreviations that are explained in the later literature.\n - Too much contents are included in the Figure 3, making it difficult to read\n at the first glance.\n- There are several confusions in the paper:\n - If Equation (7) is re-written from Equation (6) by denoting C, then a $W^n$ term\n is missing in the formula. I doubt the mathematical correctness of Equation (7)\n and the related discussion on the benefits of this recursive form.\n - In Section 3.4, it has been mentioned that $\\lambda$ is a hyper-parameter and\n is controlled by C. However, the paper doesn't specify how this is achieved.\n - In Section 3.4, it has been mentioned that the embedding pre-standardization\n is applied for various benefits. However, it doesn't specify how this is achieved\n as well.\n - In Section 3.4, it has been mentioned that the iteration number in DCA is fixed\n to 1. However, it doesn't specify the purpose of this setting. Also, if there is\n only one iteration required, what would be the benefits of transforming the update\n of the embedding space into a recursive formulation as in Equation (7)?\n - In the Conclusion, it has been claimed that \"We conduct experiments on recent \n models and a diverse set of downstream tasks,...\". However, experiments are only\n conducted on the image 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": 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 are the main contributions or innovations of the embedding space division and dynamic equation-based approach compared to existing methods?\n- For different generation methods in Pseudo-cluster center constraints, could the authors provide theoretical analysis alongside the empirical results?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Easy to follow\n- Tested on multiple datasets" }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces an Implicit Transferability Modeling (ITM) paradigm to improve the accuracy of embedding space modeling, with experimental validation conducted across multiple datasets. Overall, the paper looks interesting." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In Lines 240-243 and 268-269, the authors state that 'the division and the recursive formulation reduce overall complexity.' It is recommended to substantiate this claim with both theoretical analysis and quantitative experiments.\n2. What are the main contributions or innovations of the embedding space division and dynamic equation-based approach compared to existing methods?\n3. For different generation methods in Pseudo-cluster center constraints, could the authors provide theoretical analysis alongside the empirical results?\n4. It is recommended to include more comparison methods from recent 2024 publications.\n5. A discussion on the limitations of the proposed approach would be valuable.\n6. Thorough proofreading and refinement of illustrations would enhance the clarity and quality of the paper." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. How is the independence of the K subspaces ensured? If the method involves dimensional grouping, does this assumption of independence hold true in practical scenarios?\n\n2. The random generation of pseudo-cluster centers, as mentioned in Weakness 2, seems counterintuitive. A more thorough explanation would enhance understanding\n\n3. I find it noteworthy that the predicted state in Equation (7) exhibits a similar format to the predictions made in LEAD[3]. Both methodologies utilize dynamic equations to model the evolution process, yielding initial and final state interpolations. It would be beneficial to provide a more detailed comparison between their method and LEAD, highlighting the key differences and potential advantages of their approach. This would help clarify the novelty and contribution of this work.\n\n4. The experimental results indicate that previous dynamic evolution-based methods perform poorly on the proposed benchmark. A detailed explanation, accompanied by visualizations (e.g., Figure 7), illustrating the advantages of the proposed method in modeling the evolution process compared to earlier approaches would strengthen the discussion.\n\n5. The absence of evaluation on common benchmarks (e.g., SFDA, Logme) raises concerns regarding the generalizability of the method. I recommend supplementing the results with additional evaluations on these established benchmarks to demonstrate the method's applicability.\n\n[3] LEAD: Exploring Logit Space Evolution for Model Selection." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The study addresses a significant research area: model selection. The ability to choose an appropriate pre-trained model for specific downstream tasks is crucial and has the potential to enhance performance outcomes.\n\n2. Great visualization of the proposed method, providing clear insights into the framework's structure and functionality." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents an Implicit Transferability Modeling (ITM) paradigm that employs an implicit modeling strategy to capture the intrinsic properties of pre-trained models. The ITM framework integrates a Divide-and-Conquer Adaptation (DCA) process for efficient modeling of independent subspaces, along with a Pseudo-Clustering-based Optimization (PCO) strategy to enable effective estimation without extensive fine-tuning." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. In line 225, the methodology mentions \"K independent subspaces sampled from E\"; however, the process for obtaining these divisions is not adequately described. The criteria for determining K and the methodology used to derive these independent subspaces require further clarification. It would be helpful if the authors could provide specific examples of how K is determined and the practical sampling process for the subspaces.\n\n2. In line 307, the generation of pseudo-cluster centers is noted to involve various methods, including the use of random vectors from high-dimensional space. This approach appears to conflict with conventional methods (centers are determined by some clustering methods) but lacks sufficient explanation within the text.\n\n3. The evaluation of the proposed method is conducted solely on a single benchmark, omitting more widely recognized benchmarks (used in SFDA[1], Logme[2]), which limits the robustness of the findings.\n\n[1] Not All Models Are Equal: Predicting Model Transferability in a Self-challenging Fisher Space.\n\n[2] LogME: Practical Assessment of Pre-trained Models for Transfer 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": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "N/A" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "The questions are mostly related to the major weakness mentioned:\n\n1. Please run a larger set of ranking experiments, using subsampling from a (larger) model-zoo and share these results, preferably in some scatter plot (eg LogMe vs ITM) on all 10 target datasets.\n2. Please specify the final transferability scores s of the ITM method for model i and a target task.\n3. Please explain what Ê (E-hat) is in Eq 7 and how it is obtained (without full fine-tuning).\n4. Please provide the loss functions used for Lpc and Lobj and how they are used to compute s. \n5. Please explain how the validation data is used in the ITM method.\n6. Please explain for how many epochs ITM is run over the target data (and how that compares to related work)." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The paper starts with a reasonably strong analysis of transferability estimation, that these yield inconsistent estimates between different architectures or pre-training methods. This is interesting to see, and yields already some questions. For example, from Fig 1: I mainly observe that LogMe and ETran have a preference of a single network architecture for any target dataset, so their transferability scores seems to be a function of the pre-trained network, more than for the given target dataset. Does that hold for other TE methods as well? Do the target tasks prefer different models, or is there just a single best model for all target tasks? And from Fig 2: this plot makes me wonder what the performance of a supervised ViT model would be, it seems that it is not in the mix of pre-trained models?\n\nThe individual elements of the proposed method (estimating target embedding, using some pseudo clustering) seems to make sense." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "In this paper transferability estimation is studied, that is, rank a set of pre-trained networks for a specific target task, using a computational efficient method (the underpinning assumption is that fine tuning all pre-trained networks is too time consuming). In this work, the mapping from the original embedding space (E) of a model from the zoo, to the fine-tuned embedding space (E’) on the target data is estimated via a combination of three loss functions, and this is used to estimate the transferability estimation score. The method is validated on a set of 10 classification tasks datasets, with a model-zoo of 10 models." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "# Major weaknesses\n\n1. The biggest weakness is that the method is not clearly presented, while conceptually the elements make sense, the interconnection and the final TE method (or scoring) is not directly clear from the manuscript. To make this weakness more concrete:\n - (major) The final transferability estimation score (s_i) for model i (for a particular target task) is not explicitly defined. It is related to the ideal mapping gamma, but not provided.\n - (major) The loss (eq 7) uses Ê (E-hat), which is defined above Eq 7 as the final embedding state after fine-tuning. Defining a transferability metric which works after fine-tuning is probably not the goal, so please clarify.\n - (major) How is the lambda parameter (Eq 8), controlled by the learning rate (eta) in Eq 7 as stated in L331. How is Eq 8 in general related to Eq 7? Eq 8 defines a joint loss, Eq 7 only an update of the parameters.\n - (major) What is the objective driven loss? How is it defined?\n - (major) In the implementation details it is stated that some of the target dataset is used for validation purposes in ITM. What parts are validated, and when in the pipeline, and how is that used?\n - The difference between mapping Gamma (L222) and mapping Phi (L224) is not clear to me. And what is the difference between psi(.,.)(L224) and psi(.) (L245)?\n - What is z, it seems only implicitly used in the (approximation) of ^Z (Z-hat)? Moreover, the decomposition of q_phi(z|E), is purely based on the independence assumption of E, I don’t think a double bayes rule is necessary to write that down. Finally, how is q used? Is the loss in Eq 4, a sum over all j in K (of E 3)?\n - What is the difference between the subsets A (indexed by j) and the subset of K (also indexed by j)? \n - What does it mean that the process is “treated as an interaction between the model’s transferability and downstream tasks, leading to more effective and adaptable estimation across a broader range of scenarios.” What interaction is used, is the model learned once for multiple downstream tasks? Does it then generalize to others?\n\n2. The second biggest weakness of the current study is the experimental evaluation, although this is a more general problem within transferability estimation literature, but the current experimental evaluation is too weak to draw any conclusions. The main problem is that a single ranking experiment is evaluated, while the outcome of this ranking really does depend on the selected individual models in the model-zoo. This means that adding or removing a single model to/from the zoo all results are likely to be significantly different. This has been studied in [AgostinelliECCV22]. Therefore, the method should be evaluated on a larger set of ranking experiments. This is not too complicated nor too computational expensive: one could draw samples from a larger model-zoo. For example one could draw all sets of 10 models out of a model-zoo of 14 models, this provides (approx) 1000 ranking experiments (14 choose 10). This only requires computing once the scores (s) and ground truths (r) for 4 additional models, while providing 999 more ranking tests. Without such an experiment, I think it is impossible to draw any conclusion on the success of any method.\n\n# Minor weaknesses\n- Figure 1: I’m unsure what I see (and should see) in this plot. It is remarkable that the MAE-B16/SimMIM-B16 models have different GT radar plots between (a) and (b). \n- Figure 2: The y-scales are different from plot (a) and plot (b). \n- (nitpick) In the related work the `cite` command is used, where it should be the `citep`, for example: NCE Tran et al. (2019) → NCE (Tran et al., 2019).\n- (nitpick) The sub-index i is used for the number of elements in the dataset (L186) and for the number of models in the zoo (L187)\n- It is unclear how many epochs over the target task training data are performed by each method. Most transferability estimation methods assume (eg) 1 epoch (to get feature embeddings and target labels).\n- In the abstract and introduction ‘generalizability’ of a transfer estimation method is mentioned, what do you mean with that?\n\n# References\n- [**AgostinelliECCV22**] Agostinelli et al, How stable are Transferability Metrics evaluations?, ECCV 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.", "Yes, Other reasons (please specify below)" ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "See the weakness." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "[1] The method tackles the problem of transferability estimation. The author conducts exhaustive methods to validate the effectiveness of the proposed methods.\n\n[2] The method seems to be reasonable." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper tackles the problem of transferability estimation, which aims at selecting the best pretrained model for specific downstream tasks. The authors propose the implicit transferability method to efficiently model the transfer process, reducing both learning complexity and computational costs. Specifically, they propose a divide and conquer adaptation process to model the transfer process. They also introduce a Pseudo-Clustering-based Optimization (PCO) strategy with static and dynamic constraints to estimate transferability without intensive retraining. Their approach outperforms current methods, showing significant improvements across ten benchmarks, making it highly effective and generalizable for model selection." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "[1] The motivation is unclear. In the introduction (L081-085), the author states two challenges, i.e., the strategy of implicit transferability modeling remains largely unexplored and the implicit modeling process requires the final embedding states of the model after fine-tuning. However, it can not fully understand the where the challenge is. The author more specifically explain the challenges.\n\n[2] The description of the method is not clear. For example. In L221-L222, the embedding space division: The author is suggested to briefly introduce the motivation of division, what to divide and how to divide. Currently, I can not fully understand the motivation and implementations.\n\n[3] The pseudo-clustering accuracy is essential to transferability estimation. The authors are suggested to evaluate the sensitivity of the clustering performance." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": { "value": "@inproceedings{\nanonymous2024generalizable,\ntitle={Generalizable Transferability Estimation of Foundation Vision Models via Implicit Learning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v8GuB74YRA},\nnote={under review}\n}" }, "abstract": { "value": "Transferability estimation aims to identify the most suitable model from a collection of pre-trained models for specific downstream tasks, playing a crucial role in the success of the pre-training and fine-tuning paradigm. However, the recent proliferation of pre-trained models with diverse architectures and training strategies poses significant challenges for transferability estimation due to discrepancies in intrinsic model characteristics, making it difficult for existing methods to accurately simulate embedding space evolution within feasible computational limits. To address these challenges, we propose an Implicit Transferability Modeling (ITM) paradigm that incorporates an implicit modeling strategy for the intrinsic properties of pre-trained models, enabling more accurate transferability estimation. ITM employs a Divide-and-Conquer Adaptation (DCA) process to efficiently model the transfer process, reducing both learning complexity and computational cost. Additionally, we introduce a Pseudo-Clustering-based Optimization (PCO) strategy that eliminates the need for extensive fine-tuning, enabling effective estimation without intensive retraining. Our method significantly outperforms state-of-the-art approaches, achieving notable improvements across ten widely used benchmarks and demonstrating its effectiveness and generalizability in enabling accurate and efficient model selection for downstream tasks." }, "anonymous_url": { "value": "I certify that there is no URL (e.g., github page) that could be used to find authors’ identity." }, "authorids": null, "authors": null, "code_of_conduct": null, "code_of_ethics": { "value": "I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics." }, "comment": null, "confidence": null, "contribution": null, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": null, "keywords": { "value": [ "Transferability Estimation", "Transfer 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/d826f7789e6f17d19d2af3ad326de87fba994f62.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": "Generalizable Transferability Estimation of Foundation Vision Models via Implicit 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": 2 }, "primary_area": null, "questions": { "value": "My main questions are in the weakness section. Below is a question regarding the proof: In the proof of Theorem 2 (line 712), I am confused since it is first stated that $Y_i - f_{\\text{hi}}(\\hat Y_i^{b^{--}}) <0$ and then \n$Y_i > f_{\\text{hi}}(\\hat Y_i^{b^{--}})$." }, "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 question and makes an interesting \nobservation. Some theoretical attempts have also been made to understand the \nobservation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper studies the efficiency of conformal prediction when the predicted value has a systematic bias, aiming to understand the effect of symmetric/asymmetric quantile adjustment on the corresponding \nconformal prediction length. Under a stylized model, theories are \ndeveloped to understand the behavior of symmetric/asymmetric quantile adjustment.\nThe theory is then evaluated on synthetic and real data." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The paper's main theories (Theorem 2 and 3) are under highly stylized models and the statement lacks rigor (please correct me \nif I misunderstood anything). For example, Theorem 2 is stated for prediction intervals in the generic form of Eq. (2), but it appears that the proof only focuses on the case of residual errors and CQR. For another example, the proof of Theorem 2 seems to use the fact that \n$f_{\\text{hi}}(\\hat Y_i^{b--}) = f_{\\text{hi}}(\\hat Y_i^{0})+b^{--}$, which is not stated as an assumption. The theorem also appears to assume that the bias is sufficiently large in magnitude.\n\n\nI am also not following the form of CQR used in this paper; in my understanding, \nCQR fits conditional quantiles of $Y$ given $X$ by minimizing the pin-ball loss, \nand the resulting fitted conditional quantile function operates $X_i$ instead of the point prediction of $\\hat Y_i$ \n(unless the quantile is obtained in a specific form centered around a point prediction).\nI am also confused by the description of CQR in lines 194-196 --- what is $n_s$, what is the set of samples $\\hat Y^b_i$ \nand how are they generated. If the prediction is deterministic, does this reduce to the residual error case? Please let me know if I misunderstood anything." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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": "- Is Romano et al. 2029 the only work where non-symmetric intervals are used and tested?\n- What happens if an adaptive scheme is used to compute the interval? E.g. if the symmetric constant intervals are replaced by reweighted intervals as in Papadopoulous 2008? In particular, how would Theorem 2 change? \n- Would it be possible to compute an upper bound of the bias using the gap between symmetric and asymmetric intervals?\n- Can the proposed approach help study the non-exchangeable situation where the bias is only present at test time? In other words, can one obtain an approximated version of Lemma 1 and Theorem 2?\n- Have you run any experiments on Bias estimation?" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "- As CP applies to any given model, the results may help establish a good trade-off between model flexibility and efficiency in applications where data are scarce.\n- The algorithm for estimating the bias from the obtained interval is interesting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The proposed method aims to improve uncertainty quantification in the presence of model bias. The authors show that non-symmetric prediction intervals may be more robust to model bias than their symmetric counterparts. The claims are supported by theoretical and empirical results." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Non-symmetric CP is not new. The paper's contribution would be better introduced by adding an intuitive explanation of why *asymmetric adjustments have been theorized to yield longer interval lengths as a consequence of stronger guarantees (Romano et al.,2019)* but *has also been empirically observed that asymmetric adjustments yield tighter intervals than symmetric ones.*\n- The authors focus on the setup where comparable bias is present in the calibration and test samples. They should comment on whether it would be more efficient to 1) improve the underlying point-prediction model or 2) consider more flexible conformity scores (e.g. adding and training a free shifting term to the conformity scores).\n- Asymmetric intervals can be defined by splitting the calibration samples according to the residual signs and then evaluating two sample quantiles. It would be helpful to see a theoretical comparison of the gap between the proposed method and such a naive approach as a function of the size of the calibration set." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 5 }, "contribution": { "value": 1 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- How do the paper results get used in practice? Since any learning algorithm perfectly mitigates the bias that you present. So, in your case, the bias is not present in the training set but is in the calibration set (so that it can account for this). Can you provide a use case where this occurs?\n- You need to correct the table reference; you see Tab. 4.2.1, while in the caption, it is Table 1. The same is true for the references to the figures, where you use Fig. X, and the caption states Figure X. Change either the captions or the references." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- **Clear and Comprehensive:** The CP mechanisms, the derivation of the bias effects, and the comparison between symmetric and asymmetric intervals are clearly explained.\n- **Formalizing Good Message:** Recently, in the CP field/community, asymmetric adjustments or conformal prediction intervals with symmetric (the same) density are used beneath the lower bound and above the upper bound, which is appropriate for the end-user from an interpretation standpoint. Additionally, there is occasional observation of performance gains using asymmetric adjustment, as the paper rightfully mentions. Formalizing and theorizing these observations is undoubtedly beneficial for the field and practitioners." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper studies how a systematic homoscedastic bias is present in the point prediction. It studies how it influences the efficiency of prediction intervals produced by inductive conformal prediction (ICP) and conformalized quantile regression (CQR), explicitly comparing the effect of symmetric and asymmetric adjustments. For ICP, the impact of using the absolute residual (symmetric adjustment) and the residual (asymmetric adjustment, controlling the coverage on the left and right) is compared. For CQR, similarly. \n\nThey show that in the case of simple homoscedastic bias, $b$: 1) the upper bound for the size of the intervals of the biased predictions, which of the size of the intervals without prejudice and $2|b|$; 2) the asymmetric adjusted prediction interval is of the same size as there was no bias, and 3) they show under which condition the asymmetric or symmetric adjustment will result in smaller prediction intervals (however, this requires knowledge of the bias)." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- **Elementary Theoretical Findings**: While the theoretical contributions are sound, the results may be considered elementary. The core findings, such as the impact of bias on symmetric intervals and the lack of bias influence on asymmetric intervals, are straightforward and follow from the basic principles of CP. Though helpful, these insights do not introduce particularly advanced or novel theoretical complexity, which may limit the paper's appeal to a more specialized audience seeking more profound theoretical innovations.\n- **Simplified Bias Assumptions**: The assumption of bias as a constant noise term following the same global distribution across all predictions may not adequately capture more complex forms of bias present in real-world data, such as feature-specific or covariate-dependent biases. The paper also mentions on lines 089-094 different reasons for these biases. However, they result in way more complex biases.\n- **Limited Discussion on More Complex Non-Conformity Scores**: While the paper acknowledges that more complex non-conformity scores may require different approaches, it does not explore these in-depth, potentially limiting the generality of the findings for more advanced CP methods.\n- **Limited Synthetic Experiment Settings:** The paper only evaluates experiments where $n$ is large, and the data is generated from symmetrical distributions. However, this is problematic because the strong point of the symmetric adjustment is that it can better leverage asymmetries (skewness) of the aleatoric uncertainty distribution (noise distribution on the label).\n- **Wrong Claims:** Your statement on line 266 is incorrect. When the number of samples is large and no bias is present, the length of prediction intervals generated by a symmetric adjustment is approximately equal to the ones generated from asymmetric adjustments. See the above point.\n- **Section 3.1 is too bloated:** Given your bias, you could just take the mean of the calibration set's error, to retrieve it." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 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": "Please have look at the weakness section." }, "rating": { "value": 1 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Practical Consideration of CP Efficiency: The authors address the critical issue of efficiency in conformal prediction (CP) sets, specifically focusing on the size of prediction intervals, which is relevant in real-world applications. By exploring the impact of bias on CP intervals, the authors attempt to improve the accuracy and reliability of predictions, which is essential for high-stakes tasks.\n\nExperiments: The paper demonstrates multiple real-world experiments, such as sparse-view CT reconstruction and weather forecasting." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper investigates how bias, defined as the systematic deviation from ground truth, affects Conformal Prediction (CP) intervals in regression tasks. CP is used for uncertainty quantification in machine learning models. The authors focus on two adjustment methods for CP intervals: symmetric (where both sides of the interval are adjusted equally) and asymmetric (where adjustments can be unequal). Through theoretical and empirical analyses, they show that:\n1. Symmetrically adjusted intervals increase in length by 2|b|, where b is the bias.\n2. Asymmetrically adjusted intervals are unaffected by bias.\n3. Under certain conditions, asymmetric intervals are tighter than symmetric ones.\nTheir findings suggest that asymmetric intervals maintain their \"tightness\" even under biased predictions, unlike symmetric intervals that inflate in length. These conclusions are validated with real-world tasks in CT reconstruction and weather forecasting, highlighting the potential for more bias-robust machine learning systems." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "Inconsistent Notations and presentation: The paper suffers from unclear and inconsistent notations. For instance, in equation (1), it's not explicit which bias is being addressed --- it is not clear how the randomness in the ground truth is being dealt with. Additionally, \\hat{Y_i^b} is ambiguously defined; it is not clear whether these quantities are scalars or they are allowed to be a set of values. For instance, in the first paragraph of Section 3 the Y_i is treated as scalars and in the second paragraph, they use Y_i = {Y_{ij}} as a vector. Besides, the score definition proposed by Romano et al.[1] is presented inaccurately, which raises concerns about the foundation of the theoretical analyses. The overall readability of the paper is not good. \n\nLimited Insight from Corollary: While Theorems 1 and 2 seems a bit straightforward and is a direct consequence of the linear assumptions. The utility of Corollary 3.1 is not clear, the RHS of eq. (6) is independent on b, and the LHS is dependent on b? It is not clear how the authors are setting L_{asymb} \\leq L_{sym} in the proof. \n\nMissing Theoretical Justification for the Algorithm: The authors do not provide a theorem guaranteeing that the proposed algorithm retains the validity of the coverage guarantee for CP intervals. This omission is significant as it leaves a gap in understanding whether the method meets one of CP's fundamental requirements.\n\nLack of Coverage Probability in Simulations: The simulations fail to report the coverage probability, which is a critical metric for evaluating the reliability of proposed algorithm. This weakens the experimental validation of the proposed methods.\n\nComparison with related work: There are some related work which also aims to reduce the length of the conformal intervals [2,3]. It would be nice to compare the contribution of this work with these two related works. \n\n[1.] Romano, Yaniv, Evan Patterson, and Emmanuel Candes. \"Conformalized quantile regression.\" Advances in neural information processing systems 32 (2019).\n\n[2.] Xie, Ran, Rina Foygel Barber, and Emmanuel J. Candès. \"Boosted Conformal Prediction Intervals.\" arXiv preprint arXiv:2406.07449 (2024).\n\n[3.] Liang, Ruiting, Wanrong Zhu, and Rina Foygel Barber. \"Conformal prediction after efficiency-oriented model selection.\" arXiv preprint arXiv:2408.07066 (2024)." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "Conformal prediction intervals computed using asymmetric adjustments remain tight and valid when predictions are biased, while conventional symmetric adjustments inflate with increasing bias." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024regression,\ntitle={Regression Conformal Prediction under Bias},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v8RDgaEtE2},\nnote={under review}\n}" }, "abstract": { "value": "Uncertainty quantification is crucial to account for the imperfect predictions of machine learning algorithms for high-impact applications. Conformal prediction (CP) is a powerful framework for uncertainty quantification that generates calibrated prediction intervals with valid coverage. \nIn this work, we study how CP intervals are affected by \\emph{bias} -- the systematic deviation of a prediction from ground truth values -- a phenomenon prevalent in many real-world applications.\nWe investigate the influence of bias on interval lengths of two different types of adjustments -- symmetric adjustments, the conventional method where both sides of the interval are adjusted equally, and asymmetric adjustments, a more flexible method where the interval can be adjusted unequally in positive or negative directions.\nWe present theoretical and empirical analyses characterizing how symmetric and asymmetric adjustments impact the \"tightness\" of CP intervals for regression tasks. \nSpecifically for absolute residual and quantile-based non-conformity scores, we prove: 1) the upper bound of symmetrically adjusted interval lengths increases by $2|b|$ where $b$ is a globally applied scalar value representing bias, 2) asymmetrically adjusted interval lengths are not affected by bias, and 3) conditions when asymmetrically adjusted interval lengths are guaranteed to be smaller than symmetric ones.\nOur analyses suggest that even if predictions exhibit significant drift from ground truth values, asymmetrically adjusted intervals are still able to maintain the same tightness and validity of intervals as if the drift had never happened, while symmetric ones significantly inflate the lengths. \nWe demonstrate our theoretical results with two real-world prediction tasks: sparse-view computed tomography (CT) reconstruction and time-series weather forecasting. Our work paves the way for more bias-robust machine learning 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": [ "Conformal Prediction", "Bias", "Uncertainty Quantification" ] }, "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/71f4af246309326189286577b71558fb73aca9da.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": "Regression Conformal Prediction under Bias" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. In line 776, “counterfactual editing” is mentioned. According to the description in lines 776-781, modifying certain facts constitutes counterfactual editing. Is this definition accurate?\n2. The case study in Section 5.3 could include more analysis. Figure 5 and Figure 6 only present a relatively simple entity editing example (person-related information) and do not showcase complex cases of visual semantic editing or user-specific knowledge editing. It’s recommended to select representative and challenging cases or specific types of cases in visual semantic editing or user-specific knowledge editing. \n3. The superiority of IKE has been demonstrated in many previous works. What new perspective does your benchmark provide? Section 5.3 only states that IKE performs better than FT-LLM without explaining why IKE has stronger generalization and transferability. It would be helpful to explain the reasons for the performance differences across methods based on the proposed benchmark. It is suggested to provide a detailed analysis of how IKE's performance on this new benchmark differs from its performance on previous benchmarks, and what these differences reveal about the nature of multimodal knowledge editing tasks.\n4. This benchmark is designed to uncover issues that other benchmarks cannot detect. Could you provide some examples or data to demonstrate that your benchmark is indeed more challenging and more valuable for improving models compared to other benchmarks?" }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. The benchmark proposed in this paper uses knowledge represented in free-form natural language, making it more applicable to real-world scenarios. In addition to traditional visual entity editing, the benchmark incorporates visual semantic editing and user-specific editing, allowing for a more comprehensive evaluation of model editing capabilities.\n2. The paper provides a detailed description of the dataset construction process, offering valuable insights and methodologies for data collection and structuring.\n3. The experiments conducted on the benchmark cover a wide range of model editing methods across different types of knowledge editing tasks, resulting in an extensive and insightful evaluation." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper presents a benchmark for knowledge editing in multimodal large models, specifically targeting knowledge represented in free-form natural language. The benchmark focuses on tasks such as visual entity editing, visual semantic editing, and user-specific editing. The paper provides a detailed description of the data collection and construction process for this benchmark. Various model editing methods were evaluated on this benchmark using models like BLIP-2, MiniGPT-4, and LLaVA 1.5, revealing limitations in existing approaches." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. I’m not sure if the workload is sufficient. If this were solely a Dataset & Benchmark Track submission, the workload would be appropriate. However, as a long paper submission to ICLR, it may require additional technical contributions. For instance, adding theoretical analysis to explain why existing methods perform poorly in multimodal knowledge editing could provide new perspectives for improving this area. Therefore, it’s recommended to include an in-depth analysis on why current methods underperform in certain aspects relevant to the proposed benchmark. It is suggested to provide related theoretical analysis or technical contributions that you believe would enhance the paper's suitability for ICLR. \n\n2. I didn’t find evidence in the experimental or case study sections demonstrating the necessity of this benchmark. This is not to question the value of your work, but providing examples would strengthen the case. Additionally, clarifying which scenarios would particularly benefit from using this benchmark for evaluation would be helpful. It’s recommended to add more evidence highlighting the ‘differences’ and ‘necessity’ of this benchmark. I suggest you provide comparative analyses with existing benchmarks or specific real-world scenarios where this benchmark would be particularly valuable." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": 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) Would it be more beneficial to construct knowledge editing samples using incremental, real-world updates? Although challenging, creating samples based on real-world updates, such as player transfers or recent match results sourced from news articles, could enhance dataset relevance. Using real, evolving information would help models learn to handle dynamic knowledge updates and might improve their applicability in practical knowledge editing scenarios.\n\n(2) What impact did human verification have on the LLM-generated descriptions? An analysis or examples of descriptions before and after human verification would help demonstrate how much human intervention improved the dataset's consistency and accuracy.\n\n(3) Is there a comparison of Visual Entity Editing data quality between MMKE-Bench and prior datasets (e.g., MC-MKE, VLKEB)? While Table 1 highlights MMKE-Bench’s increased task diversity with the addition of Visual Semantic Editing and User-Specific Editing, it would be useful to understand how MMKE-Bench compares to previous datasets in terms of data quality for Visual Entity Editing. Such a comparison could clarify whether MMKE-Bench offers improvements in data quality alongside task diversity.\n\n(4) For additional questions, please refer to the Weaknesses section." }, "rating": { "value": 6 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "(1) Diverse Task Setup: MMKE-Bench covers three distinct knowledge editing tasks, from entity-level editing to more complex user-specific knowledge editing. This provides a comprehensive tool for evaluating multimodal models' ability to update knowledge and handle personalized information.\n\n(2) Free-Form Natural Language Descriptions: Unlike traditional triple-based representations, this benchmark uses natural language descriptions to represent knowledge items, enabling models to engage in editing tasks in a more realistic scenario. The free-form descriptions combined with image data make the tasks closer to the complexity of real-world semantics.\n\n(3) Extensive Evaluation Using State-of-the-Art Multimodal Models: The paper evaluates MMKE-Bench on prominent multimodal models such as BLIP-2, MiniGPT-4, and LLaVA-1.5. These models represent the cutting edge in the multimodal field, making the experimental results widely relevant and reflective of current model capabilities in knowledge editing tasks.\n\n(4) Systematic Experimental Analysis: The paper provides a thorough evaluation of various existing knowledge editing methods, including FT-LLM, KE, MEND, SERAC, and IKE, offering a comprehensive performance baseline for each task type. This analysis provides valuable insights into how different methods perform across the tasks presented in MMKE-Bench." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper introduces MMKE-Bench, a multimodal knowledge editing benchmark dataset aimed at addressing the current gap in resources for evaluating large multimodal models (LMMs) in knowledge editing tasks. MMKE-Bench is largely constructed using counterfactual samples to enhance the robustness of model evaluation and to examine their ability to perform knowledge editing across varied and challenging scenarios. The dataset includes three types of tasks: visual entity editing, visual semantic editing, and user-specific knowledge editing. Each task is represented through free-form natural language descriptions combined with images, generated by large language models (LLMs) and verified through human annotation to ensure consistency. The paper evaluates several state-of-the-art multimodal models, including BLIP-2, MiniGPT-4, and LLaVA-1.5, providing insights into the strengths and limitations of current models across different knowledge editing tasks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "(1) Figures 1 and 2 Could Benefit from Improved Clarity and Accessibility: Figures 1 and 2 could be refined to enhance clarity and accessibility for a broader audience. In Figure 1, the example used is soccer-related, which may not be immediately understandable to readers unfamiliar with the sport. A more universally recognizable example, such as common objects or activities, could make the data construction process clearer. For Figure 2, the visual design could better distinguish the four dataset construction steps; currently, it’s difficult to determine which modules belong to each step. Using distinct colors, numbering, or borders to separate steps would help readers follow the process more intuitively.\n\n(2) Limitations of Counterfactual Data for Real-World Applications: The counterfactual approach used to construct MMKE-Bench may lead to a distribution that differs from real-world data. While counterfactual samples aid in testing robustness, they represent hypothetical rather than naturally occurring scenarios. As a result, models fine-tuned on this dataset might learn patterns specific to these counterfactual cases, potentially reducing their effectiveness in real-world knowledge editing tasks. A comparison of model performance on both counterfactual and real-world data would provide valuable insights.\n\n(3) Lack of Empirical Analysis on the Impact of Human Verification: The paper mentions that human verification was conducted to improve the consistency of LLM-generated descriptions, but it does not provide examples or quantitative comparisons to illustrate the impact of this verification process. Including a before-and-after analysis would strengthen the case for the added value of human verification." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "As mentioned in the weaknesses, the similarity of editing visual entity and semantic editing, the user-specific editing scenario and T-Loc test warrant further exploration.\n\nGiven the modest amount of training data, is the dataset volume adequate for methods that require training, like serac or mend? Additionally, how is the IKE method adapted to LMM editing? Which examples are used in context? Is target knowledge incorporated within the context, and do you include image-based examples in this context?\n\nIn Section 5.2.1, there is a claim that “5) Modern LMMs excel in producing and applying edited knowledge. For reliability, generalization, and portability evaluations, LLaVA-1.5 outperforms BLIP-2 and MiniGPT-4. This improved performance can be attributed to its larger model size and better instruction-following capability.” This raises two questions: First, is LLaVA-1.5 indeed larger than MiniGPT-4, as both use a 7B LLM, and MiniGPT-4's vision encoder appears larger? Second, this statement is not directly related to the benchmark’s core focus, which is to compare editing methods rather than models.\n\nIn Section 5.2.2, further clarification is needed regarding the meaning of “user number” and “allowed maximum items.” Additionally, what is the precise gap between editing and testing in user-specific editing, and why is the gap in the two visual editing tasks similar (1, 3, 6, 10) without larger gap?\n\nTypo:\nFigure 2, bottom-right: \"G-Rel.\"\nTable 2: \"Entity Entity Editing\"\nTable 6: SERAC GAP 3 T-Rel, missing decimal point" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "The use of free-form natural language as input for knowledge editing tasks is a notable strength, enhancing flexibility and making the approach adaptable. The clarity of the writing aids comprehension, and the experimental setup is well-documented. Additionally, the benchmark spans diverse data sources and entity types, allowing for broad applicability across different tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This work focuses on knowledge editing in large multimodal models and introduces a new benchmark. It presents three distinct types of editing tasks: visual entity editing, visual semantic editing, and user-specific editing, using free-form natural language as input data for these edits. The benchmark also diverges from previous versions by removing the T-Gen test and adding a T-Rel test. The evaluation includes five editing methods across three large multimodal models (LMMs), effectively validating the benchmark's dataset." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "There is some overlap between visual entity editing and visual semantic editing, as both tasks involve understanding image content, which could blur the distinction between these editing types. Additionally, the user-specific editing scenario may lack practicality. In real-world applications, database or memory-based search might be more effective than training user-specific information for each user to achieve personalization in LLMs or LMMs.\nRegarding the T-Loc test, there’s room for improvement. The results are near 100, suggest that the randomly sampled T-Loc questions are relatively easy for methods such as SERAC, FT-Alignment, and MEND. Introducing more challenging cases could enhance the evaluation’s robustness. For instance, collect similar but harder test cases, either through web crawling or using LLM-generated content (e.g., from GPT) and verify them. This could improve the test’s effectiveness." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": { "value": "No Ethics Concerns" }, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "As mentioned in the weaknesses.\n\nFrom figure 1, there is no T-Gen test in MMKE-Bench, and you add a T-Rel test. Why do you make this change?\n\nAdditionally, how do you adapt each method to the lmm editing?\n\nFigure 4: why I-Gen not evaluated for MMEdit?\n\nSection 5.2.2 requires additional clarification on the terms “user number” and “allowed maximum items.” Further, what about the mend method in your sequential editing setting? And are the gaps substantial enough for the test? In LLM knowledge editing studies and the referenced vlkeb work, this gap can be larger, reaching 100 or more.\n\nThe analysis in line 425-426 only explains possible reason for serac, leaving out mend.\n\nLine 450, “visual knowledge” can be ambiguous, and the reliability seems to be not lower as stated.\n\nLine 466, how do you draw the conclusion “parameter-based methods are better at applying edited knowledge to new contexts”?\n\n\nWriting issue:\n\ndouble quotes in line 324-327 and elsewhere: please pay attention to how to type correct left quotation mark in latex.\n\nFigure 1: what is the difference between original and editing knowledge of user-specific editing? They are same that lead to confusion, which can be improved for btter clarity.\n\nFigure 2: Should the \"G-Rel.\" change to “T-Rel”? And the two question in the last box (bottom right) are same. And placing “hint” there can be confusing, as if the hint itself is part of the input.\n\nLine 355: you use “MLLMs” here, but in other places, you use LMMs\n\nTable 2: \"Entity Entity Editing\"\n\nTable 6: missing decimal point in SERAC results, GAP 3 T-Rel column" }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "Utilizing free-form natural language as input for knowledge editing tasks stands out as a key advantage, offering greater flexibility and adaptability to various contexts. The data collection process is robust and comprehensive, and the benchmark supports both individual and sequential editing tasks." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This research investigates knowledge editing within large multimodal models, introducing a novel benchmark for evaluation. It defines three specific editing tasks: modifying visual entities, altering visual semantics, and implementing user-specific adjustments, all using natural language as the input format. The benchmark is validated through an assessment of five different editing techniques across three large multimodal models, effectively confirming the dataset's applicability." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "The training dataset size seems to be small, which might not be enough for mend/ serac that need training. This could potentially lead to lower performance of such methods.\n\nThere is potential to enhance the T-Loc test. The current results, which are close to 100, indicate that the randomly selected T-Loc questions may be relatively simple for existing methods. Strengthening the evaluation could involve introducing more challenging cases. For instance, given that GPT was used in data collection, generating analogous but more difficult test cases and validating them could increase the test's effectiveness and robustness.\n\nThe division of the three tasks lacks clarity. Both visual entity editing and visual semantic editing require image comprehension and retention of the target knowledge provided in the text. While they may involve different data sources, they essentially fall under the same task, visual question answering. Similarly, user-specific editing could be seen as an application scenario, but it still aligns with the core task of VQA, making no real difference in “task type.” \n\nAdditionally, I question about if this user-specific editing approach is practical. In real-world applications, creating personalized databases for each user might be more efficient and effective than modifying the model or embedded knowledge directly.\n\nAnd the task generalization might not be persuasive using only one case. If you want to prove this, more comprehensive experiments and evaluations should be conducted." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose MMKE-Bench, a challenging benchmark for evaluating diverse semantic editing in real-world scenarios." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024mmkebench,\ntitle={{MMKE}-Bench: A Multimodal Editing Benchmark for Diverse Visual Knowledge},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v8qABSeeKO},\nnote={under review}\n}" }, "abstract": { "value": "Knowledge editing techniques have emerged as essential tools for updating the factual knowledge of large language models (LLMs) and multimodal models (LMMs), allowing them to correct outdated or inaccurate information without retraining from scratch. However, existing benchmarks for multimodal knowledge editing primarily focus on entity-level knowledge represented as simple triplets, which fail to capture the complexity of real-world multimodal information. To address this issue, we introduce MMKE-Bench, a comprehensive **M**ulti**M**odal **K**nowledge **E**diting Benchmark, designed to evaluate the ability of LMMs to edit diverse visual knowledge in real-world scenarios. MMKE-Bench addresses these limitations by incorporating three types of editing tasks: visual entity editing, visual semantic editing, and user-specific editing. Besides, MMKE-Bench uses free-form natural language to represent and edit knowledge, offering a more flexible and effective format. The benchmark consists of 2,940 pieces of knowledge and 7,229 images across 110 fine-grained types, with evaluation questions automatically generated and human-verified. We assess five state-of-the-art knowledge editing methods on three prominent LMMs, revealing that no method excels across all criteria, and that visual and user-specific edits are particularly challenging. MMKE-Bench sets a new standard for evaluating the robustness of multimodal knowledge editing techniques, driving progress in this rapidly evolving field." }, "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 knowledge editing; Large multimodal model; 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/ff6859b76de179de9c130119f93d6f3dc9f2f10c.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": "MMKE-Bench: A Multimodal Editing Benchmark for Diverse Visual Knowledge" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]

[ { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "- For stronger VLMs (e.g., GPT-4o), I don't quite get the motivation for authors to adopt the \"instantial experience\" approach instead of the \"empirical experience\" approach. Intuitively, allowing VLMs to reason and explicitly output the clear instructions with ambiguities removed should almost always improve their performance, regardless of the ability of the VLMs themselves. I also didn't find an ablation study in the paper to support why we need the \"instantial experience\" approach for stronger VLMs.\n- Additionally, per my comments in the review, I'd encourage authors design better prompts such that the generated instructions in their \"empirical experience\" approach do not defeat the purpose of an actual \"instruction\"." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 2 }, "strengths": { "value": "- Context and instruction ambiguity is very prevalent in language and vision-language applications. Handling these cases is an important research problem. The proposed approach presents a promising way to improve the accuracy of VLM's \"best guess reasoning\" under ambiguous instructions.\n- The paper is overall well-written, with clear methodology and extensive experiments." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Visual-O1, an approach that improves VLM's reasoning performance given ambiguous language instructions. For VLMs with stronger reasoning ability, the paper proposes an \"instantial experience\" approach where multi-turn multi-modal CoTs are generated to reason about the instruction, and the model subsequently directly outputs final answer conditioned on such reasonings. For VLMs with weaker reasoning ability, the paper proposes an \"empirical experience\" approach where after multi-modal CoTs are generated, explicit and clear instructions are generated before the model outputs the final answer. The authors demonstrate the effectiveness of their approach in referring image segmentation and visual question answering benchmarks." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "- Table 6 indicates that the effectiveness of the proposed approach and the results of the paper may be vulnerable to noise, and the improvements could lack statistical significance. Specifically, as the budget limit increases from 1 to 5, the success rates are \"55.87, 54.16, 57.38, 55.27, 55.56,\" showing no clear trend of performance improvement with respect to budget limit. For instance, the fact that \"budget limit = 3\" outperforms \"budget limit = 2\" by 3% and \"budget limit = 4\" by 2% could simply reflect evaluation noise.\n- The generated \"clear instruction\" in Figure 3 (which correspond to authors' \"empirical experience\" approach) appears to directly answer the initial question, rather than serving as an actual \"instruction\" that enables the LLM to explore various objects, select the relevant one, and then provide the final answer. Thus, authors' approach seems to undermine the purpose of an \"instruction.\" For instance, for the ambiguous prompt `white bear turned slightly`, a more suitable instruction with ambiguity removed would be, `identify the bear that is predominantly white with subtle color variations from white`." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 3 }, "contribution": { "value": 2 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 2 }, "primary_area": null, "questions": { "value": "1. Could you provide some further information to explain the difficulty of the ambiguous dataset? For example, how humans (might) perform on this dataset?\n2. How are the Chain-of-Though and FuDD baselines implemented?\n3. An example of $\\mathcal{A}_\\text{ins}$ is provided in the appendix. If I understand the method correctly, it should be composed of the alternating appearance of reasoning and reflection. Are the sentences in gray given by reflection? If so, is there an example of the full trajectory?\n4. Analysis of Visual-O1's failure modes will help a lot, as discussed in the weaknesses section." }, "rating": { "value": 5 }, "reciprocal_reviewing": null, "resubmission": null, "revert_desk_rejection_confirmation": null, "revert_withdrawal_confirmation": null, "soundness": { "value": 3 }, "strengths": { "value": "1. This paper proposed a prompting method that seamlessly applies to both high-intelligence and general-intelligence models within the same framework. This adaptability indicates that Visual-O1 is not limited to a specific model type and can scale across different levels of model intelligence, which enhances its utility for a wider range of applications and users.\n2. The authors presented extensive experimental results that include a variety of ablation studies and model comparisons. These results strengthen the credibility of their claims and provide thorough evidence of the framework's effectiveness. Notably, Visual-O1 not only improves performance on ambiguous instructions datasets but also enhances results on general datasets." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "This paper proposes Visual-O1, a multimodal multi-turn Chain-of-Thought reasoning framework to deal with ambiguous natural language instructions in two typical vision-language tasks: referring image segmentation and visual question answering. Visual-O1 builds\ninstance-specific experience for high-intelligence models or creates general experience for any ambiguous instructions for general-intelligence models. Experiments show that Visual-O1 improves the performance on both the ambiguous datasets and the general datasets." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. Although Visual-O1 shows clear improvements over baselines, the absolute success rates and IoU scores (on the ambiguity dataset) are still not entirely satisfactory. This gap suggests that there are underlying limitations to the model’s ability to handle certain types of ambiguous instructions. The paper could be strengthened if the authors provided a more detailed examination of failure cases (e.g., identifying specific patterns in instructions that remain challenging to disambiguate). Without this analysis, Visual-O1's readiness for real-world applications remains in question.\n2. The paper tends to overemphasize its successes by frequently using terms like \"significant,\" even when sometimes the improvements are marginal (e.g., success rate gains of less than 3%). This could be misleading, especially considering that the ambiguous instructions dataset is not large enough to robustly support such claims. Furthermore, the results do not report averages over different random seeds, which raises questions about the stability and generalizability of the improvements. \n3. Minor issues or typos: The term $x_\\text{rsn}$ should be $x_\\text{rfl}$ in Eq. (6). The word \"botlle\" should be \"bottle\" in Figure 3." }, "withdrawal_confirmation": null }, { "TLDR": null, "_bibtex": null, "abstract": null, "anonymous_url": null, "authorids": null, "authors": null, "code_of_conduct": { "value": "Yes" }, "code_of_ethics": null, "comment": null, "confidence": { "value": 4 }, "contribution": { "value": 3 }, "desk_reject_comments": null, "details_of_ethics_concerns": null, "flag_for_ethics_review": { "value": [ "No ethics review needed." ] }, "keywords": null, "large_language_models": null, "no_acknowledgement_section": null, "other_comments_on_LLMs": null, "paperhash": null, "pdf": null, "presentation": { "value": 3 }, "primary_area": null, "questions": { "value": "1. Can you write the exact GPU settings for Table M?\n2. How does the author interpret that the performance saturates quickly with the budget?" }, "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's proposed direction of Visual-O1 towards instructional QA is an interesting direction.\n2. By integrating the chain of thought into the pipeline, the model can improve the segmentation result of referring image segmentation, and the visual QA result. The performance gain on visual QA is larger than RIS, and the performance gain on ambiguous cases is higher than general cases, which indicates that the multi-round and multimodal inference needs more on the harder case.\n3. The additional overhead of the visual chain of thought is limited for both training and evaluation stages." }, "student_author": null, "submission_guidelines": null, "summary": { "value": "The paper proposes a method for visual o-1 that is similar to language o1 that increases the computation budget on test-time. The work is proposed to highlight the potential of artificial intelligence to work like humans in the real world, especially when the instruction is ambiguous for the instructional model. This is achieved by multi-turn reasoning with chain of thought. \n\nThey have separated the multi-round reasoning process for high-intelligence model and general intelligent model, where it tends to have more round interaction when the model has a higher intelligence level.\n\nThey have evaluated their result on visual-QA, and RIS in comparison with the state-of-the-art model and shows reasonable improved on both segemtantion and visual-QA." }, "supplementary_material": null, "title": null, "venue": null, "venueid": null, "weaknesses": { "value": "1. To be completely honest, although the performance gain shows statistically significant improvement, however, the improvement is marginal, which indicates that the chain-of-thought works.\n2. The author lacks examples when we actually need visual CoT (As the cipher example in o1).\n3. As shown in Table.6, I actually think the budget curve is confusing, the performance fluctuates for Acc and BLEU, and saturate w/ 3 budget, this indicates that the designed algorithm is not sophisticated enough for scaling up.\n4. They propose the empirical solution for less-intelligent model." }, "withdrawal_confirmation": null }, { "TLDR": { "value": "We propose Visual-O1, a multi-modal multi-turn reasoning framework that enhances high-intelligent and general-intelligent models' understanding of ambiguous instructions in multi-modal tasks by simulating human reasoning." }, "_bibtex": { "value": "@inproceedings{\nanonymous2024visualo,\ntitle={Visual-O1: Understanding Ambiguous Instructions via Multi-modal Multi-turn Chain-of-thoughts Reasoning},\nauthor={Anonymous},\nbooktitle={Submitted to The Thirteenth International Conference on Learning Representations},\nyear={2024},\nurl={https://openreview.net/forum?id=v9CDpLpjiE},\nnote={under review}\n}" }, "abstract": { "value": "As large-scale models evolve, language instructions are increasingly utilized in multi-modal tasks. Due to human language habits, these instructions often contain ambiguities in real-world scenarios, necessitating the integration of visual context or common sense for accurate interpretation. However, even highly intelligent large models exhibit significant performance limitations on ambiguous instructions, where weak reasoning abilities of disambiguation can lead to catastrophic errors. To address this issue, this paper proposes Visual-O1, a multi-modal multi-turn chain-of-thought reasoning framework. It simulates human multi-modal multi-turn reasoning, providing instantial experience for highly intelligent models or empirical experience for generally intelligent models to understand ambiguous instructions. Unlike traditional methods that require models to possess high intelligence to understand long texts or perform lengthy complex reasoning, our framework does not significantly increase computational overhead and is more general and effective, even for generally intelligent models. Experiments show that our method not only significantly enhances the performance of models of different intelligence levels on ambiguous instructions but also improves their performance on general datasets. Our work highlights the potential of artificial intelligence to work like humans in real-world scenarios with uncertainty and ambiguity. We will release our data and code." }, "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": [ "Understanding ambiguous instructions", "large multimodal model", "chain-of-thoughts", "multimodal" ] }, "large_language_models": null, "no_acknowledgement_section": { "value": "I certify that there is no acknowledgement section in this submission for double blind review." }, "other_comments_on_LLMs": null, "paperhash": null, "pdf": { "value": "/pdf/2c9e0c18094ed8c13dc367d37d60e112911e9aa5.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": "Visual-O1: Understanding Ambiguous Instructions via Multi-modal Multi-turn Chain-of-thoughts Reasoning" }, "venue": { "value": "ICLR 2025 Conference Submission" }, "venueid": { "value": "ICLR.cc/2025/Conference/Submission" }, "weaknesses": null, "withdrawal_confirmation": null } ]