Add files using upload-large-folder tool

human_reviews/0Ag8FQ5Rr3.md

@@ -0,0 +1,214 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces the concept of "super weights" in Large Language Models (LLMs), identifying a small number of individual weight parameters (as few as one) that have a disproportionately large impact on model performance.  Pruning these super weights drastically reduces the quality of generated text, while pruning thousands of other larger-magnitude outliers has a negligible effect.  The paper proposes a data-free method for identifying super weights based on their connection to "super activations," exceptionally large activation outliers previously observed in LLMs.  Finally, the paper demonstrates that preserving super weights and activations during quantization significantly improves compression quality, achieving competitive results methods like SmoothQuant.
+
+### Strengths
+- The identification of "super weights" and their connection to super activations represents a novel and potentially significant finding in understanding the inner workings of LLMs.
+- Connection of "super weights" to quantization accuracy is quite interesting and has practical implications.
+- The paper provides a clear methodology for identifying super weights and evaluating their impact, along with an index of super weight coordinates for common LLMs, facilitating further research.
+
+### Weaknesses
+# Major
+
+- Connection to Adversarial Examples: The literature extensively documents how small changes in the input domain can drastically alter output probabilities. Consequently, significantly harming the network by removing weights, as demonstrated, is somewhat expected.  A discussion addressing the connection between super weight removal and adversarial examples would strengthen the paper.
+
+- Magnitude Pruning Baseline: In Table 1, the comparison of super weight pruning with global magnitude pruning may not be the most informative.  A stronger baseline would involve pruning only within the layer where super activations occur. This would better isolate the impact of the super weight itself.
+
+- Quantization Baseline: The "Naive W8A8" quantization baseline should incorporate clipping. The current presentation makes it unclear whether the observed improvements stem from outlier removal or clipping, especially since super weight handling affects only a single layer during quantization, while clipping is applied to every layer.  Furthermore, it should be noted that the clipping threshold is determined using Wikitext-2, which is also included in the evaluation of quantized models.
+
+# Minor
+
+- Terminology:  The term "extreme" might be more descriptive and informative than "super" when referring to these weights.
+
+- Weight Distribution Visualization: Including a histogram visualizing the position of the super weight within the overall weight distribution would enhance understanding of its magnitude relative to other weights.
+
+### Questions
+- Section 3.2, "Prune SW+SA":  The description of the "Prune SW+SA" condition in Section 3.2 is unclear.  Specifically, how does this condition differ from the original model?  I understand that super activations typically precede super weights in the model.  Therefore, I am unsure what modification is being made in "Prune SW+SA" and how it distinguishes itself from the original, unpruned model.  Could you please elaborate on this procedure?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper reveals that Large Language Models (LLMs) contain a very small subset of weights
+(super weights) that are extremely important, where removing them severely degrades model
+performance. The researchers developed an efficient, data-free method to identify these super
+weights using only a single forward pass. They further investigated how these super weights
+influence network behavior by analyzing their relationship with activation outliers. Building on
+these insights, they proposed a quantization approach that carefully preserves these super
+weights while effectively compressing other weights, resulting in the maintenance of model
+quality after compression.
+
+### Strengths
+The discovery is interesting and the proposed quantization method is easy to implement, which
+can maintain better performance compared to Round to nearest quantization with the same
+block size.
+
+### Weaknesses
+The authors failed to show how the proposed methods can improve the SOTA.
+1. Although the method is data-free, its performance does not exceed SOTA methods like
+SmoothQuant, given incorporating a small calibration dataset would not increase the
+quantization complexity much.
+2. The author mentions that this method is hardware-friendly, but no experiments to show
+its effectiveness in improving latency, throughput, memory usage, etc.
+
+### Questions
+1. For equation 1, the median is used to replace super activation. Is getting the median
+time-consuming since GPU is not good at sorting? (Although there are GPU-version
+sorting algorithms)
+2. The authors mentioned that SmoothQuant does not report on some models this paper
+evaluates, they compare our results with naive W8A8 quantization (line 407 - line 409).
+Can the authors run SmoothQuant on these methods since it is open-source? The naive
+W8A8 is a too-weak baseline.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper investigates the sensitivity of a subset of outliers in LLMs, referring to them as "super weights." The authors conducted experiments to examine the impact of these super weights on model performance.
+
+### Strengths
+The authors conducted experimental explorations on the so-called "super weights."
+
+### Weaknesses
+1. The necessity of "super weights" is unclear, as outliers are already identified based on the threshold. Increasing the threshold will naturally reduce the number of outliers with very large weights. Given the known importance of outliers in LLMs, emphasizing "super weights" (outliers at a higher threshold) does not appear novel.
+
+2. Figure 1 is misleading. According to the author's definition, "super weights" are a subset of outliers. However, the figure suggests -1.9 is a typical outlier with nearby values being quite small (.1 and .2), implying that zeroing out outliers produces nonsensical text—a widely acknowledged fact. To better demonstrate the significance of super weights, it would be beneficial to explore whether zeroing out all outliers results in poor performance, and similarly, whether zeroing out just a small subset (e.g., 20-30) leads to comparably severe degradation.
+
+3. Table 1 raises critical concerns. First, the criterion for selecting outliers needs specification. Second, the "Prune SW, +SA" setting in Lines 146-152 is confusing, as it suggests pruning super weights while partially restoring super activations enhances quality. However, the authors did not prune activations, leading to confusion about this claim.
+
+4. Table 2 appears redundant and fails to convey meaningful information. Replacing it with visual representations of "super weights" distributions would be more informative, as the current table occupies considerable space without offering clear insights.
+
+5. Figure 2 is difficult to interpret. The depiction of super weights and their impact, such as generating nonsensical text, is not clear. The use of the same color block in both the network and the output is puzzling. Are the model's dynamics linear? How do the output and weights share the same significance? Clarification is needed on whether this figure is based on assumptions or empirical data.
+
+6. In Lines 189-190, the term "super activations" is introduced but lacks clarity on whether it is threshold-based or aligns with corresponding weights, which could be time-consuming. The authors should clarify this terminology.
+
+7. The paper contains several unprofessional notations. For example, "Yij" should be corrected to "Y_{ij}" in Line 204, and similarly, "Xik" and "Wjk" should be "X_{ik}" and "W_{jk}" in Line 205. The inconsistency in notation and dimensions between "d" and "D" in Line 204 suggests a lack of careful writing and review, raising concerns about the overall professionalism of the paper.
+
+8. Lines 198-210, which discuss the identification of super weights, are crucial yet unclear. The selection criteria for super weights remain ambiguous and need a precise mathematical description. Readers should understand the definition of outliers and the criteria for their selection explicitly.
+
+9. The paper lacks consistency in terminology. "Super weights" sometimes refer to both activations and weights, and at other times only to weights, adding confusion. In Line 306, the term "super outliers" is introduced, suggesting that the paper should maintain consistent terminology from the start, including in the title, if both weights and activations are discussed.
+
+After several careful readings, there are numerous additional concerns throughout the paper. The issues are substantial and critical, making it unlikely to meet the standards of ICLR. I recommend a strong reject based on the quality of this paper and will not change my rate.
+
+### Questions
+Please refer to the weaknesses section above.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+1
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper is about the discovery of super weights in LLMs that are disproportionately important, pruning these hurts model quality quite a bit. The authors have provided a way to identify these super weights using a forward pass. Super weights are activations are sensitive to quantization effects and hence authors propose a super weight aware quantization method enabling effective quantization.
+
+### Strengths
+Novel discovery about the importance of a few handful of neurons: The identification and analysis of super weights and super activations as critical outliers and their positive influence on model's performance is noteworthy and interesting. 
+
+Quantization proposals: Authors went one step further to propose a super weight-aware quantization method to make the best use of these super weights/activations. Data free quantization proposal with on par performance compared to SmoothQuant is also a worthy contribution.
+
+### Weaknesses
+Though the discovery is quite interesting, the improvements of proposed methods with existing baselines are quite marginal. In general, such kind of super weights might be a natural phenomenon in any machine learning model. How can one say this is relevant only to LLM's?
+
+The work seems to be very much based on empirical observations (which is not my concern) but more discussions/intuitions/explanations around how/why these super weights are formed will be useful.
+
+### Questions
+The paper mostly focuses on post training model weight/activation analysis and identifies certain handful of importance weights/activations. The authors also say that irrespective of the input prompt the super weights are always the same and they mostly occur in the early layer's down projection with some reasoning via skip connections diagram. 
+
+Though these insights are helpful, but it would be good if authors can follow up with what happens during the training process that such super weights are formed in the first place. Does the training methodology in terms of quantization during training/layernorm, gradient scaling, etc play any role in the forming of these super weights?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+This paper focuses on the impact of outlier weights in large language models (LLMs), specifically larger weights, which the authors term superweights and superactivations. First, the authors analyze how much these weights and activations affect LLM performance. They then use this as motivation to discuss quantization methods designed to account for superweights and superactivations. Throughout the paper, the authors also discuss the impact of superweight scaling and provide experimental results showing how their quantization method improves upon standard rounding, especially when using larger block sizes within the network.
+
+### Strengths
+The paper is well-written and effectively illustrates the importance of superweights and superactivations. I appreciate the discussion on the percolation of superactivations across the network and the identification of superweights across layers (Figure 3). Additionally, I find the potential implications of superweight upscaling presented in Figure 6 quite interesting.
+
+### Weaknesses
+While I appreciate the analysis presented in this paper, I am struggling to see the novelty of this work. I may be misunderstanding, but from what I gather, superweights and superactivations have already been discussed in prior analyses of LLMs. Additionally, it seems that methods like AWQ and SqueezeLLM inherently focus on superactivations. Furthermore, compared to other weight quantization techniques, the proposed method does not appear to offer significant improvements.
+
+### Questions
+1. Could the authors provide clarification on the points I raised in the weaknesses section, especially if I may have misunderstood some of the contributions?
+
+2. In Figure 6, do the authors have any insights into the concave behavior of the scaling factor? Are there specific explanations or potential methods for identifying this optimal scaling factor?
+
+3. Regarding the stop word shift in distribution, is it generally accepted that a higher probability of stop words negatively impacts LLM performance?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/0iAZYF9hrl.md

@@ -0,0 +1,177 @@
+## Human Reviewer 1
+
+### Summary
+The paper presents a study of disentangled representation learning on three microscopy image datasets. The representation learning strategy starts by training an Ada-GVAE model using a Textures-dSprites dataset introduced in this work. The dataset is supposed to reflect simple textures that could help interpret information in microscopy images. After training this model in a weakly supervised way, it is used to encode images of another domain, with optional unsupervised finetuning using a beta-VAE. The resulting features are low dimensional, and interpretable, and are used to train classifiers.
+
+The ideas and the study are generally interesting, but the paper lacks technical novelty, is limited to a small-scale empirical evaluation only, and the experiments are incomplete to fully understand the value of the proposed strategy.
+
+### Strengths
+* The paper evaluates the recent ideas of disentangled representation learning using weak supervision in a more realistic application.
+* The paper also presents an alternative to learning the disentangled representation from RGB images based on models pretrained at large scale.
+* The paper proposes a new sprites dataset to facilitate the interpretation of microscopy images.
+
+### Weaknesses
+* The technical contribution is limited. Beyond the sprites dataset and the use of pretrained features, many of the ideas have been presented in previous works.
+* The experimental evaluation is limited to quantifying the impact of classifier types (GBT vs MLP) and input type (RGB vs DINO features). Many questions remain open regarding how much classification accuracy could be obtained without the proposed disentanglement procedure. Can the authors compare results of training a classifier directly with RGB images and another classifier with DINO features without any modifications? These results would help understand how difficult the tasks are and what is the trade-off between using disentanglement vs not using it.
+* It is possible that DINO features are already disentangled and all what the proposed strategy is doing is assigning names to some of the factors of variation that DINO can detect. Therefore, the disentanglement is not really happening in the VAEs but rather obtained from a model pretrained at large scale. What type of experiment can the authors design to test this hypothesis?
+* If the hypothesis above is not rejected, the value of proposed methods is limited to only annotating factors of variation rather than identifying them in a weakly supervised manner to then being transferred.
+
+### Questions
+Can the authors clarify the questions above? Specifically, the extent to which DINO already offers certain degree of disentanglement and how the factors of variation of interest could be identified directly from these representations.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper addresses the interpretability challenge in microscopy image analysis using deep learning approaches. The authors propose a novel methodology based on Disentangled Representation Learning (DRL) to enhance model interpretability while maintaining classification performance. The approach leverages transfer learning from synthetic features and is validated across three diverse microscopy domains: plankton, yeast vacuoles, and human cells. The growing volume of microscopy images due to technological advances has necessitated automated analysis methods, yet interpretability remains crucial for practical applications in fields such as diagnosis and environmental monitoring. The authors demonstrate that their DRL framework successfully balances the trade-off between model accuracy and interpretability in microscopy image classification tasks.
+
+### Strengths
+1. The manuscript is well-written and easy to follow, with clear organization and logical flow.
+2. The application of weakly-supervised DRL to real-world image analysis represents a promising and valuable research direction.
+
+### Weaknesses
+1.The scope of this work appears too narrow, focusing solely on microscopy images.  The proposed approach might be more convincing if demonstrated on natural images as well.
+2.The authors fail to adequately justify why DRL should be specifically applied to microscopy image analysis.  Furthermore, they do not clearly articulate whether this specific application domain poses new challenges or requirements for DRL that could lead to innovative solutions.  The authors' insights into these aspects are not well presented.
+3.Given the lack of compelling insights, this work appears to be primarily an application of existing DRL methods without significant methodological or theoretical innovation. This level of contribution may not align with ICLR's focus on novel methodological and theoretical advances in machine learning.
+4.The paper appears to lack comparative experiments. While the disentanglement scores might be novel evaluation metrics, the absence of comparisons for classification performance is particularly concerning and unreasonable.
+
+### Questions
+Referring to the weaknesses noted above, I find the claimed contributions of this paper not sufficiently convincing.  Could the authors provide a more compelling explanation of their main contributions, particularly addressing:
+1. Why DRL is specifically suited for microscopy image analysis.
+2. What novel challenges or requirements this domain brings to DRL.
+3. How their approach advances the theoretical or methodological aspects of DRL beyond simple application.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper proposes a Disentangled Representation Learning (DRL) approach to improve interpretability in microscopy image classification. By pre-training on synthetic data (Texture-dSprite) to capture factors of variation, the authors apply these learned representations to real-world microscopy datasets (Plankton Lensless, Plankton WHOI15, Budding Yeast Vacuoles, and Sipakmed Human Cells). Their method aims to support model interpretability while achieving high classification performance with gradient-boosted trees and MLPs for downstream analysis.
+
+### Strengths
+The paper explores the application of an existing DRL framework to the specific domain of microscopy images. This idea is interesting as it shows a potential pathway for combining DRL with microscopy image analysis.
+
+### Weaknesses
+A significant weakness as it seems, is the absence of a comparison with other similar methods. The paper presents only one framework and does not discuss or evaluate alternative approaches, which weakens the case for this framework’s efficacy or advantage over existing methods.
+
+The contributions of the paper in terms of novelty are unclear. The study applies an existing DRL approach to a new domain but does not appear to introduce any fundamentally new concepts, techniques, or substantial modifications to existing methods. The only apparent novelty - the application of DRL to microscopy imaging does not suffice. This limits the potential impact and originality of the work.
+
+The paper’s presentation suffers from numerous issues that impede readability and clarity:
+1. There are instances of informal languages, such as the use of “thanks.”
+2. The text contains multiple errors at the word, sentence, and structural levels, which disrupts the reading experience. Sections like Section 2.2 (“Disentanglement Evaluation”) resemble output generated by ChatGPT and lack rigorous academic polish.
+3. Figures appear low-resolution, with inadequate explanations in captions. Captions should be comprehensive and self-contained, but here, they lack essential details, e.g., explanations of metrics like OMES and balanced accuracy.
+4. The use of multiple highlight types (underscoring, bold, italics) is excessive and distractive. Minimal highlighting would improve readability and make essential points more accessible.
+5. Important metrics are either not explained in the text or lack adequate definitions in the captions, leaving readers uncertain of their meaning. This omission impacts the study’s reproducibility and overall clarity.
+
+### Questions
+Most of my questions are related to major weaknesses.
+
+What specific contributions does this paper make beyond applying DRL to microscopy images? It would be helpful if the authors could clarify what is novel in their approach and how it advances the state-of-the-art in microscopy image analysis beyond existing techniques.
+
+What are alternative approaches the authors could have used for comparison? 
+
+Metric explanations (e.g., OMES, MIG, DCI and balanced accuracy) are mostly missing. Could the authors clarify these metrics, ideally using mathematical notation and provide justification for using them?
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+1
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+In this paper, the authors propose to use a disentangled representation learning framework to enhence model interpretability for microscopy image classifications. The method is based on fine-tuning a model trained on synthetic images, the proposed framework is tested on some microscopy images datasets.
+
+### Strengths
+- The paper addresses a significant challenge in representation learning: disentanglement, which plays a pivotal role in improving the interpretability of classifiers, particularly in the context of biological images.
+
+### Weaknesses
+- The proposed approach is not well explained. Indeed, the method proposed by the authors learns a disentangled model with weak-supervision using Ada-GVAE on a synthetic dataset and then fine-tune it on
+ a microscopy datasets. However, it is unclear why Ada-GVAE is choosed and how is the model fine-tuned.
+
+- The difference between the proposed method and Dapueto et al is unclear.
+
+- The authors claim that the disentanglment learned from a synthetic images can be transferred to microscopy images, such claim should be theoretically and empirically evidenced.
+
+- The paper is not well organized, for instance a "Related Work" section should be added. Two different sections (2.2 and 3.5) have the same title "DISENTANGLEMENT EVALUATION".
+
+### Questions
+- In Fig, 1, what is exactely fine-tuned and how ?
+
+- How is an RGB image directly fed to the classifier (GBT and MLP)?
+
+- In line 322, the authors state "We can observe that after finetuning, it may change, nicely adapting to the specificity of the dataset, where scale and texture are more relevant.", It is unclear for me why scale and texture are more relevent then "scale and shape", as it is the case before fine-tuning.
+
+- The proposed evaluation metrics (ex:OMES) are unclear.
+
+- The authors do not compare their method to any other work, having a solid baseline is important.
+
+- The used classifiers (GBT and MLP)  are very simple, more sophisticated ones should be used (CNNs based for example).
+
+- Inputting an RGB image to the classifer is unclear as it is well-established that deep features (in this cas the features extracted by DINO) have more important patterns.
+  
+- To assess the quality of the representation, the authors realied on classification. While a good representation can lead to a better accuracy. A good representation does not necessary mean a disentangled one.
+
+- Using the accuracy only to measure the quality classification performance is not enough.
+
+- The figures are small and the captions are not clear enough.
+
+- In Figure 6, the OMES indicates that the proposed method does not lead to better disentanglement.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+4

human_reviews/1yJP5TVWih.md

@@ -0,0 +1,145 @@
+## Human Reviewer 1
+
+### Summary
+This paper addresses rank collapse, a phenomenon where embedding vectors in deep learning models converge to a uniform state. Building on previous studies that focused on transformers, this paper extends the analysis to State Space Models (SSMs). The study employs theoretical and empirical analysis to demonstrate how lambda-skip connections, LayerNorm, and gating mechanisms contribute to both the stability and expressivity of transformers and SSMs.
+
+### Strengths
+S1. The paper tackles the problem of rank collapse, extending its analysis from transformers to SSMs.
+S2. Through theoretical proofs, the paper demonstrates that lambda-skip connections prevent rank collapse, preserving model expressivity in both transformers and SSMs.
+S3. Experimental results show that lambda-skip connections and other components enhance expressivity and stability across different model architectures.
+
+### Weaknesses
+W1. The definition of the residual term between Eq.(3) and Eq.(6) is inconsistent, with ambiguity around whether X or V serves as the residual term. This inconsistency impacts the theoretical derivations that follow and should be clarified to ensure precise interpretations. Additionally, certain symbols, such as D, are used in both the SSM and LayerNorm contexts but represent different meanings. Distinct notation would improve readability and reduce potential confusion.
+W2. While the experiments generally align with the theoretical predictions, some disparities remain unaddressed. For example, the theoretical threshold for λ appears more conservative than the empirical results suggest, and additional clarification would help. Further, the appendix notes rank stability even without skip connections, which might challenge the presented theory.
+W3. The paper primarily focuses on rank collapse within the model’s architecture but does not connect this phenomenon to downstream task performance. Adding experimental results that measure downstream task performance in relation to model depth and skip connection strength could provide a more comprehensive assessment.
+
+### Questions
+Q1. Between Eq. (3) and Eq. (6), there is ambiguity regarding the residual term, specifically whether X or V serves as the residual component. This inconsistency could impact the theoretical derivations that follow. Could the authors clarify this definition? Additionally, using the same symbol D for both SSM and LayerNorm contexts creates potential confusion. Distinct notations would enhance clarity.
+Q2. The theoretical conditions for λ appear to be conservative compared to empirical findings. Could the authors explain this discrepancy? Furthermore, the appendix notes cases of rank stability without skip connections, which might challenge the theory. An analysis of these cases would be valuable.
+Q3. Could the authors provide additional experiments showing the model’s downstream performance as a function of layer depth and skip strength? Also, would the inclusion of alternative metrics, such as effective rank, offer a more comprehensive assessment of rank collapse?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper examines the phenomenon of rank collapse in general sequence model architectures, including transformers and state space models. To mitigate this issue, the paper proposes a parameterized version of the skip connection that multiplies the residual stream by a constant factor. Theoretical analysis identifies the conditions on the parameter sufficient to prevent rank collapse, and an analytical example demonstrates that neither the absence of skip connections nor the standard implementation prevents rank collapse. Finally, empirical evaluations support the findings of theoretical analysis.
+
+### Strengths
+This paper addresses the significant issue of rank collapse in sequence model architectures. It offers both theoretical analysis and empirical evaluation to support the proposed architectural component aimed at resolving this problem. I like the remark that provides the parameters corresponding to the practical architectural settings.
+
+Additionally, the theoretical development and overall presentation of the paper are commendably clear and well-structured.
+
+### Weaknesses
+The theory investigates the sufficient conditions for preventing rank collapse in the worst-case scenario. This could imply that the required conditions are overly stringent.
+
+### Questions
+The rank collapse metric is not normalized in the definition. Would it be enough to lower bound the rank collapse metric, when the norm itself evolves across layers?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper analyzes the rank collapse of SSM due to identical $\lambda$ skip connections. The authors provide a rigorous convergence rate for the rank collapse and offer sufficient guarantees to prevent it. Experimental results demonstrate the effectiveness of their analysis.
+
+### Strengths
+1. The lower boundary of the rank collapse of $\lambda$ skip connections is analytically derived. The results agree well with empirical analysis.
+2. The paper presents the convergence rate in the absence of skip connections, contributing valuable insights.
+
+### Weaknesses
+1. The authors analyze the $\lambda$ skip connections. However, the skip strength $\lambda_k$ may vary on different layers.The paper should discuss how the findings hold up under these varying conditions. Additionally, many models implement skip connections selectively across layers rather than uniformly. A discussion on the generalizability of the results would enhance the paper.
+2. Theorem 4.1 paves the way to choose suitable $\lambda$. However, in Figure 2, it appears that when $\lambda$ is sufficiently large, the rank collapse index shows little variation. Clarification on how to determine the optimal value of $\lambda$ would be beneficial.
+3. Based on theorem 4.1, could the authors explore adding constraints to the parameters to optimize $C_M$, $S$ and $c$ for improved neural network performance?
+
+### Questions
+See above.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+Dao and Gu [https://arxiv.org/pdf/2405.21060] established a form of equivalence between transformers and continuous-time state-space models.  In a different development, Dong et al. [https://arxiv.org/abs/2103.03404] showed that self-attention layers without skip connections or MLPs suffer from "rank collapse" ― with increasing layers, the output matrix tends to rank-1, i.e., all token positions tend to the same representation.
+
+The present submissions puts these together to show that rank collapse is a problem also for state-space models.  It shows that the skip connection provides vital protection against rank collapse, but that a weighted addition (with weight $\lambda$ which may be regarded as a hyperparameter, or perhaps trainable) with the skip connection is more flexible.
+
+### Strengths
+Reasons to accept:
+* Identifies rank collapse problem in state-space models like MAMBA, similar to earlier discovery of this problem in transformer-type networks.
+* Identifies skip strength parameter $\lambda$ as an important knob to limit the damage of rank collapse.
+
+### Weaknesses
+Reasons to reject:
+* Given the two papers on which this paper builds, it might be argued that the present work is relatively incremental.  (That being said, I appreciate the candor while setting up the contributions of this paper, and I learnt something from it.)
+
+### Questions
+I have a reasonable estimate of creativity and technical depth, but it is difficult for me to assess impact.  I am not familiar with the area and my assessment has limited confidence. I would not, for example, know if rank collapse is widely appreciated within even the transformer "community" (if there is any such thing).  I have not seen MAMBA become that visible or widely used compared to standard transformer-based LLMs, but cannot speculate if rank collapse played a role.  $\lambda$-tuning for robustness seems quite useful, but again I do not know the area well enough to know if, in practice, $\lambda=1$ is frequently dangerous.  If the authors point to specific places in the paper where the above issues are discussed, or add some more motivating support, that would be helpful.
+
+A few writing style and notation nits:
+
+L156-L159 set up $X^{(k)}$ as layer input and $Y^{(k)}$ as layer output.  However, equation (1) introduces $O^{(k)}$ without explaining it will provide skipping ability in equation (2).
+
+L174-L178 There seem to be some inconsistent subscripts and superscripts. On one side we see $A^{(k)}_t, B^{(k)}_t$ etc. But just after the displayed equation, for LTI systems we see the superscript $(k)$ disappear, without an explanation if this is because the LTI system is assumed to have one layer.
+
+L888-L903 While setting up expressions and bounds with so many variables, it helps to afterward highlight the most import 1-2 variables, and give qualitative connections between their typical values in practice and the implications on the bounds.  E.g., how easy or difficult would be to choose an acceptable $\lambda$ in a typical LLM?  Also, some of the definitions like $S$ are very far from the proofs in the appendix.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+2

human_reviews/2MqyCIxLSi.md

@@ -0,0 +1,180 @@
+## Human Reviewer 1
+
+### Summary
+The authors tackle the challenge of systematically defining new Topological Deep Learning (TDL) architectures and to enlarge the accessibility of the latter to the broader community. The way they approach this endeavour is by (i) proposing a new class of TDL architectures that generalises previously proposed ones, and by (ii) implementing a software module that encapsulates architectural search over this class.
+
+As for (i), the authors build upon the concepts of “strictly augmented Hasse Graphs” and “Per-rank neighborhoods”. The former ones are employed to model the structure of a combinatorial complex via an ensemble of augmented Hasse graphs, one for each neighbourhood. The latter ones prescribe defining a specific set of neighbourhoods for each rank. The authors propose GCCN as architectures which process ensembles of strictly augmented Hasse graphs with per-rank neighbourhoods with specific neural models and “synchronisation” components.
+
+As for (ii), the module is called TopoTune and is a configuration-oriented component integrated with other TDL frameworks.
+
+Experiments are conducted on graph datasets, lifted to either simplicial or cellular complexes. Results show that GCCNs can outperform standard architectures with a smaller number of parameters or lower computational cost.
+
+### Strengths
+- The submission tackles an interesting research topic in a timely manner.
+- The implemented TopoTune module can be helpful to practitioners and researchers outside of the specific field of TDL.
+
+### Weaknesses
+- From the perspective of the framework generality, it is not clear how GCCNs would unlock new interesting operations or computational patterns.
+    - Eq. 3 and 8 look particularly alike, and it is not evident what kind of advantage the latter brings. In particular, in Eq. 3, the message function $\psi$ can be specific to a particular neighbourhood (and rank), similarly to the neighbourhood message function $\omega$ in Eq. 8 — which, incidentally, is not rank specific.
+    - Specific information about ranks and neighbourhoods could be specified by features akin to ”marks” over nodes and edges of an augmented Hasse graph, and a general enough neural architecture could then make use of these for neighbourhood and rank specific updates.
+- Proposition 3 appears to be quite trivial given Proposition 1. What is it telling us in addition to that?
+- It is not clear how the proposed contributions would help “democratising” TDL, as the authors claim. The proposed approach appears to significantly enlarge the hyper-parameter space by considering a plethora of possible architectural designs arising from the combination of neighbourhood and rank specific neural modules. Although TopoTune lowers the practical effort of searching over these spaces, these large parameter searches may still require large computational capabilities to be satisfactorily performed in a reasonable time frame.
+- The value and/or interest of some experimental questions and emerging results is not clear.
+    - “GCCNs outperform CCNNs”: It is not clear what the outperformance is due to when comparing to “standard” CCNNs, which could have, potentially, neighbourhood and rank-specific message functions. What is the take-home message for readers?
+    - “GCCNs are smaller than CCNNs”: the authors do not explain why this is the case, and it is seemingly the first time this concept emerges in the manuscript
+    - “GCCNs improve over existing CCNNs”: the results seem to be merely a matter of additional hyper-parameter search?
+    - “Performance-cost tradeoff”: The authors highlight the reduced number of parameters of GCCN models, but they do not expand into how this actually translates into lower computational cost (e.g. because run-time experiments are not discussed in this section).
+- Generally speaking, the manuscript would benefit from a clearer and more punctual presentation in regards to the motivations behind the proposed contribution and how these precisely address the research questions put forwards by the authors.
+
+### Questions
+- Can the authors expand on whether CCNNs can capture GCCNs? Are there functions expressed by GCCNs that cannot be expressed by CCNNs? If the two classes are equivalent, can the authors discuss more in detail what is the effective advantage of considering their proposed GCCN class?
+- Can the authors better explain what were the research questions addressed in their experimental section and how their results contribute to answer them?
+- Can the authors better discuss how TopoTune goes beyond merely a hyper-parameter search tool?
+
+Please also see weaknesses.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+In this work, the authors propose a generalization of Combinatorial Complex Neural Networks (CCNNs) called GCCNs and an accompanying software library called TopoTune, to generalize works on CCNNs into one computational framework and streamline the training and tuning of TDL architectures. Both theoretical and empirical results indicate that the proposed framework is indeed a useful generalization of previous efforts in TDL.
+
+### Strengths
+- The main strength of this work is that the authors are able to subsume TDL architectures under a single framework.
+- The empirical results indicate to me that the framework matches existing works, thus validating the claim that the framework is indeed general.
+- The framework allows the use of GNNs, which should bring the two fields closer together and have TDL research benefit from progress in GNNs.
+
+### Weaknesses
+- L458: The authors state that “GCCNs outperform CCNNs”. Out of the 8 presented datasets, I can only find two instances (NCI1, ZINC) where GCCNs actually perform better than the best CCNN baseline (accounting for one standard deviation). I could be convinced that the benefit of TopoTune is that one must only sweep over the GNN sub-modules to obtain an (at least) on-par model. However, this would still require some effort to find the best sub-module; see question 3 for more on this.
+- In L468 and Figure 5, the authors discuss performance vs. number of parameters. However, I don not find this comparison convincing as a smaller number of parameters may not necessarily be more cost-efficient. Instead, I would like to see a comparison in terms of runtime and memory usage of the different models.
+- Since the authors argue their approach to be superior to works on higher-order GNNs, a comparison of GCCNs and higher-order GNNs would be very useful. For example, PPGN++ (https://arxiv.org/abs/1905.11136), a higher-order GNN, performs much more on par with the best GCCN on ZINC than most CCNN baselines presented in the paper.
+
+### Questions
+- In the introduction you say “However, constrained by the pairwise nature of graphs, GNNs are limited in their ability to capture and model higher-order interactions […]”. I would expect that higher-order GNNs (https://arxiv.org/abs/1905.11136, https://arxiv.org/abs/1810.02244, https://arxiv.org/abs/1905.11136) are able to capture higher-order interactions. Could you elaborate on how TDL differs from higher-order GNNs?
+- Related to the first question, in L88-L93 you mention the work of Jogl et al. (https://openreview.net/forum?id=HKUxAE-J6lq) on Cell Encodings, which is equivalent to using the standard Weisfeiler-Leman test on a transformed graph, but your argument for the shortcomings of this approach is not clear to me. In particular, you state that “However, although these architectures over the resulting graph-expanded representations are as expressive as their TDL counterparts […] the former are neither formally equivalent to nor a generalization of the latter”. What is “the former”? What is “the latter”? Assuming the former are Cell Encodings and the latter topological GNNs, why is it important that they are formally equivalent or one being a generalization of the other? Are they different in their runtime or memory requirements? Do we expect better learning behavior from TDL methods?
+- As outlined in the weaknesses, in Table 1, only on two datasets GCCNs outperform the best CCNN from TopoBenchmarkX. Can you further elaborate on the benefits of TopoTune in this context?
+- Related to the third question, can the authors provide an overview over the runtime and memory complexity of the compared CCNNs, as well as GCCNs, possibly in relation to the complexity the underlying GNN submodules?
+- Am I correctly assuming that the ZINC dataset used in this work is the full ZINC dataset with 250K graphs, rather than the ZINC (12K) version frequently benchmarked in graph learning?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The authors propose a general topological deep learning (TDL) architecture called Generalized Combinatorial Complex Network (GCCN). It aims to unify prior work on TDL under a common mathematical framework.
+Additionally, the authors provide the TopoTune library, a reusable software implementation of the proposed GCCN method.
+The experiments show that the flexibility of the GCCN framework allows it to match or outperform previously proposed TDL methods while, oftentimes, requiring fewer model parameters to do so.
+
+### Strengths
+First, the proposed GCCN architecture (while fairly straight-forward) provides a useful framework for describing a large variety of TDL methods and it enlarges the design space for such methods.
+The experiments illustrate how this simplifies the optimization of TDL models and improving upon the state-of-the-art.
+Additionally, the authors show that GCCN can match or even outperform previously proposed approaches while requiring fewer parameters to do so.
+
+Second, the provided TopoTune implementation of GCCN integrates with existing GNN and TDL libraries.
+This simplifies the exploration of novel TDL architectures and, as stated by the authors, could help accelerate research on TDL.
+However, since I am not deeply familiar with the current literature on TDL and open problems, I can not confidently assess the relevance of this contribution.
+
+Last, I want to highlight the presentation. The paper is well structured and written. The figures are of high quality and helpful.
+
+### Weaknesses
+In Section 4 the authors show a number of theoretical properties of their proposed GCCN framework.
+While certainly desirable, the value of those properties is limited. 
+As stated by the authors themselves in the proofs in the supplement, those properties are, for the most part, fairly straight-forward.
+As far as I can tell, the GCCN framework is an intuitive generalization of prior work which only provides relatively small theoretical insights.
+The overall value of the contribution therefore seems to depend on the relevance of the previously described strengths of the paper, in particular, on the relevance of the provided TopoTune implementation.
+However, as mentioned, I cannot fully assess this aspect.
+Thus, one potential general concern might be the overall relevance of the paper.
+
+Apart from this point I have only minor suggestions for improvement:
+1. I would have found a (brief) explanation of the evaluated types of combinatorial complexes (cellular vs simplicial) to be helpful.
+2. There seem to be two small errors in the formal definitions in Section 2:
+	- p. 3 (127): At $\mathcal{P}(S) \setminus \{\emptyset\}$ it should probably read $\mathcal{V}$ instead of $S$.
+	- p. 3, eq. 2 (146): $\mathrm{rk}(\tau)$ after $\exists\ \delta$ should be probably $\mathrm{rk}(\delta)$.
+
+### Questions
+1. In Figure 5, it did not become entirely clear to me why the parameter size is reduced by changing the neighborhoods. I would expect that the total number of parameters of the GNN modules are independent of the specific types of neighborhood used. However, as shown in Figure 5 this does not appear to be the case. Can you elaborate on what exactly you mean by parameter size and how it relates the the choice of neighborhoods?
+2. It is not clear to me how exactly the GCCN models are parameterized in the different experiments. In particular, which intra- and inter-neighborhood aggregators were used for the different experiments?
+3. In the conclusion, you state that you hope that TopoTune might help "bridge the gap with other machine learning fields". Apart from the connection GNNs (and possibly Transformer models), are there any specific fields you envision that might profit from such a connection?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+2
+
+### Rating
+8
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper focuses on the topological deep learning (TDL) models in particular CCNNs and proposes a new powerful graph-based methodology for new TDL architectures, named GCCNs. The paper proves that GCCNs generalize and subsume CCNNs. The paper conducts extensive experiments and shows that the GCCN architectures achieve comparable performance with CCNNs. An efficient toolkit, TopoTune, is also introduced to accelerate the development of TDL models.
+
+### Strengths
+1. The paper proposes a new method to generalize any neural network to TDL architectures. 
+2. The proposed GCCNs formally generalize CCNNs and have the same expressiveness as CCNNs. 
+3. A new toolkit, TopoTune, has been developed to make it easy to design and implement GCCNs.
+
+### Weaknesses
+1. For node-level tasks, the paper only considers three very small datasets, which might limit the application of the method. 
+2. The complexity analysis of the method is missing and the paper does not report any training time in the experiment. 
+3. The experiment of "performance versus size" is not well analyzed especially for the graph-level datasets (i.e., PROTEINS, ZINC).
+
+### Questions
+1. Could the authors use larger node-level datasets for experiments?
+2. What is the time complexity of the proposed GCCNs compared with CCNNs?
+3. The GNN models perform very different results in Figure 5. More analysis is needed.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/2p03KljxE9.md

@@ -0,0 +1,172 @@
+## Human Reviewer 1
+
+### Summary
+On the basis of existing anomaly detection methods based on visual language alignment, this paper proposes using task related languages for task oriented feature information screening and transformation to improve the model's anomaly detection capability. The experiment was conducted on multiple datasets and demonstrated better performance compared with existing methods.
+
+### Strengths
+1. This paper is with clear motivation.
+2. This paper is well-organized and easy to follow.
+
+### Weaknesses
+1. The criteria for selecting text prompts are ambiguous. Some datasets utilize the category names of the samples, while others employ diverse descriptions. These approaches rest on the critical assumption that anomalies are distinctly defined, exemplified by MNIST, where anomalies arise from differences in numerals rather than variations in handwriting styles or colors. Should the actual anomalies diverge from these presuppositions, might the proposed model's performance diminish relative to methods devoid of textual guidance? In other words, could the model forfeit its capacity to detect all possible anomalies?
+
+2. In the MVTec dataset experiment, the author opted not to employ the concise anomaly descriptions provided by the dataset itself for text prompts, instead relying solely on item categories, mirroring the approach of WinCLIP. What rationale informed this decision?
+
+3. The proposed model is an extension of WinCLIP, yet it appears to forgo the anomaly segmentation functionality inherent to WinCLIP. Is this omission attributable to certain design elements that potentially diminish the model's anomaly localization capabilities?
+
+4. Experiments have been conducted on synthetic datasets like MNIST and CelebA by altering the original datasets. While I acknowledge the challenge of selecting appropriate text prompts for real-world datasets such as MVTec, the author should endeavor to incorporate more authentic datasets into their study, such as the VisA dataset utilized in WinCLIP or the medical AD benchmark employed in MVFA [a].
+
+[a] Adapting Visual-Language Models for Generalizable Anomaly Detection in Medical Images. CVPR 2024.
+
+### Questions
+See the weakness.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes a feature transformation methodology using concept axes, which are the principal components of the difference vectors between text embeddings of prompts specially designed to ignore nuisance attributes/highlight important attributes for anomaly detection.
+
+### Strengths
+The methodology is interesting and a solid contribution to this direction of research in vision-language modelling for anomaly detection.
+
+The results appear to be promising in the experiments presented, although a wider range of experimental setups would be more convincing (see weakness) 
+
+The ablation study is comprehensive.
+
+### Weaknesses
+1. Figure 1 is not particularly intuitive or clear, and it is not explained in the text. 
+
+2. As the exact formulation of prompts is absolutely critical for this methodology, it should have more dedicated explanation in the main text of the paper, not relegated almost entirely to the appendix. 
+
+3. There are not many baselines, and it would have been more convincing if you compare more baselines with and without LAFT transformations. 
+
+4. The range of experiments presented are quite restricted. For example with Coloured MNIST, it appears that only one number-colour combination as the normal set was tried. It would be more proper to conduct multiple experiments with different combinations of attributes and show the average result. The same can be said for the other datasets.
+
+### Questions
+Please address the points raised in the Weakness section. Also:
+
+1. What is the purpose of including Aux. prompts? 
+
+2. How does different CLIP architecture and also different VLMs affect performance?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces Language-Assisted Feature Transformation (LAFT), a novel framework that leverages vision-language models (like CLIP) to enhance anomaly detection. Traditional anomaly detection methods often struggle to capture user-defined nuances of normality, particularly when attributes are entangled or datasets are incomplete. LAFT tackles this by enabling feature transformations guided by natural language prompts. These prompts align visual features with user intent by projecting image features onto specific concept subspaces within a shared embedding space. The paper also proposes LAFT AD, a k-nearest-neighbor (kNN)-based method combining LAFT with anomaly detection, and extends this work into WinCLIP+LAFT, designed for industrial applications. The effectiveness of LAFT is demonstrated across datasets like Colored MNIST, Waterbirds, CelebA, and MVTec AD, showing superior performance in both semantic and industrial anomaly detection.
+
+### Strengths
+1. LAFT bridges a gap in anomaly detection by allowing users to express preferences using natural language, providing more control over what is considered "normal."
+2. Unlike other feature transformation models, LAFT does not require additional training, making it efficient for settings with scarce data.
+3. The experimental results demonstrate that LAFT outperforms state-of-the-art methods, particularly in semantic anomaly detection tasks.
+
+### Weaknesses
+1. While LAFT demonstrates significant improvements in controlled environments, such as the Colored MNIST dataset, its performance gains appear less pronounced when applied to complex real-world datasets. This discrepancy suggests that the model may struggle to maintain robustness across multiple intricate attributes, highlighting the need for further refinement in handling multi-attribute scenarios.
+2. The experimental setup lacks comprehensive comparisons, particularly between language-assisted and vision-assisted approaches. For instance, incorporating image guidance by utilizing related reference normal images (e.g., normal digits in various colors) or color-augmentation for kNN baseline could provide valuable insights. A thorough examination of both language-based and vision-based assistance would strengthen the evaluation of LAFT's efficacy.
+3. The impact of the number of PCA components, which is the sole hyperparameter in LAFT, is not adequately investigated. Given that this parameter influences the model's performance, it is crucial to explore its effect across different datasets. Specifically, an analysis of whether a larger number of components may be beneficial for more complex datasets would provide valuable insights into optimizing the model’s performance.
+
+### Questions
+1. In Table 8, the header refers to "bird," which is inconsistent with the title of the Colored MNIST dataset mentioned (maybe a typo). Could the authors clarify this discrepancy?
+2. What are the sizes of the training sets for each dataset used in the experiments? Given that these samples serve as candidates for kNN search, how might the number of training samples affect the final performance of the model?
+3. The experimental results on the MVTec AD dataset in Table 3 suggest that InCTRL might outperform WinCLIP+LAFT when considering deviation, especially when the number of shots exceeds 2. Could the authors provide detailed experimental results for each of the five different reference sample sets?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper introduces a feature transformation method aimed at focusing on specific image attributes guided by language. The approach, termed Language-Assisted Feature Transformation (LAFT), leverages the shared embedding space of vision-language models (specifically CLIP) to modify image features according to user-defined concepts expressed in natural language, enabling enhanced anomaly detection capabilities without additional training.
+
+### Strengths
+- The authors explore a valuable research topic that contributes to the current body of knowledge—how to adjust decision boundaries using language to enhance CLIP’s anomaly detection performance. 
+- The proposed method stands out due to its training-free nature, which provides flexibility in application across various tasks with limited data.
+
+### Weaknesses
+- The paper uses the vector difference between two textual descriptions to represent a single attribute and maps this attribute directly to image feature transformation. However, this simplification raises at least three issues:
+   - The properties of objects cannot be adequately represented by the difference between two concepts.
+   - Real-world attributes are often complex and may involve different colors or textures across various parts of an object.
+   - The text embedding space and the image embedding space in CLIP are not perfectly aligned; therefore, vectors derived from the text space may not be directly applicable to the image space.
+
+- To validate the effectiveness of feature transformation, using a CLIP-based classification task would be more suitable than anomaly detection.
+
+- The paper lacks results on anomaly localization, which is crucial for industrial applications.
+
+- The language throughout the paper could be clearer. It is recommended to refer to previous works using proper method names and provide concise descriptions of these methods.
+
+- The axis labels in Figure 3 are inconsistent. How were the attributes 'Number' and 'Color' derived?
+
+- The dataset chosen for experiments, SEMANTIC ANOMALY DETECTION, focuses on distinguishing simple concepts. Why not test the method on widely recognized OOD datasets such as ImageNet-1k and OpenOOD? Industrial anomaly detection would benefit from validation on datasets like VisA and Real-IAD as well.
+
+- The comparison methods included are relatively weak. Why not compare with more recent OOD detection approaches such as NegLabel [1] and ClipN [2]?
+---
+- \[1] X. Jiang, F. Liu, Z. Fang, H. Chen, T. Liu, F. Zheng, and B. Han, “Negative label guided OOD detection with pretrained vision-language models,” in The Twelfth International Conference on Learning Representations, 2024.
+- \[2] Hualiang Wang, Yi Li, Huifeng Yao, and Xiaomeng Li. ClipN for zero-shot OOD detection: Teaching CLIP to say no. ICCV, 2023.
+---
+If the author can address my concerns, I will consider increasing the score.
+
+### Questions
+1. What does $c_i$ represent in Equations 5 and 6?
+2. For zero-shot anomaly detection, can the transformed image features still match the text features effectively?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+5

human_reviews/2tIyA5cri8.md

@@ -0,0 +1,152 @@
+## Human Reviewer 1
+
+### Summary
+The paper looks for evidence of whether Llama3-70B model can simulate the TD-learning algorithm for solving RL tasks. The authors evaluate this using simple toy RL environments. They train different SAEs for the different tasks and find features that correlate highly with TD-errors and Q-values. They confirm that these features are causal in predicting actions by performing interventions on such features. Based on these evidence, the authors conclude that the LLM is simulating the TD-learning algorithm.
+
+### Strengths
+1. The study evaluates their hypothesis through a series of tasks to substantiate their empirical claims.
+2. Intervention experiment with the features to confirm their causal roles.
+3. The writing is clear and easy to understand. However, some details are missing. See in questions.
+
+### Weaknesses
+My main objection with the paper is that there is a simpler alternative hypothesis that could equally explain all of the results. Given the simplicity of the task, the LLM could be implementing the following algorithm to solve the tasks:
+
+
+Step 1: Keep a track of the maximum points for each episode in the context.
+
+Step 2: Predict the actions from the episode that has the maximum points.
+
+
+This algorithm is simple to implement for the LLM given previous works on LLMs implementing greater than circuit [1] and induction heads [2]. Also, for the Two-Step Task, the first 7 episodes are provided by using a random policy, which should cover all the 4 different trajectories possible in the task.
+
+The features that the authors find using SAEs could be features that are tracking the maximum points across episodes. These features will have high correlation with Q-values, and are also causal so interventions on them should show similar results as shown in the paper.
+
+I recommend that the authors conduct experiments designed to refute this hypothesis. See questions for some suggestions on experiments that can be performed.
+
+References:
+
+[1] How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model. https://arxiv.org/abs/2305.00586
+
+[2] In-context Learning and Induction Heads. https://arxiv.org/abs/2209.11895
+
+### Questions
+1. In the plots on max correlation with values/errors (eg fig 2c, 2d, 3, 4b, 4c, etc.), is the correlation computed with the value/error of the action predicted by the LLM at the given state? If yes, then it would be valuable to check whether there are features that correlate with value/error of non-optimal actions. This could help in distinguishing whether the LLM is actually implementing TD-learning or the max-point episode algorithm provided above.
+2. Can you provide how the NLL score computed? I couldn't find it in the appendix either. Particularly, are you computing the log probabilities of Q-learning agent by doing a softmax using the Q-values over actions?
+3. Are you using any discount rate for the Grid World Task? If yes, please provide it.
+
+### Soundness
+2
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper is way out of my expertise and hence I cannot provide a meaningful review.
+
+### Strengths
+.
+
+### Weaknesses
+.
+
+### Questions
+.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+1
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper presents a mechanistic analysis of internal activations of the Llama 3 70b model during three different in-context reinforcement learning tasks. In particular, the authors use Sparse Auto-Encoders (SAEs) to generate latent space representations of the residual streams and show how they correlate with TD Errors and Q-values of Q-learning agents trained to solve the same tasks. Furthermore, they show that relationship between the latent representations and the TD Errors is causal by intervening on the latent representations, which causes a degrading in the performance of the model.
+
+### Strengths
+I think the paper is well written and the setting and the experimental details are generally well explained. The contributions are also clearly stated. Furthermore, as far as I can tell, the presented experimental methodology is also sound. Although it is a known fact that Transformers can potentially perform In-Context RL, especially if trained for it, it is the first time, to the best of my knowledge, that a mechanistic analysis is conducted on a model which was pre-trained on next token prediction. In addition, even if the methods used (e.g. SAEs) are already well established in the mechanistic interpretability literature, it is insightful to see how they can be successfully used also to better understand how LLMs solve In-Context RL. Hence, even if the problem of In-Context RL is well studied in the literature and the interpretability methods used are also well established, overall I think the presented results shed more light on the inner workings of how LLMs can solve RL tasks in-context, which can be significant and insightful for the community.
+
+### Weaknesses
+- The main weakness of the paper is that being an experimental work, I find the number of experiments conducted to be a bit limited. I think that more experiments should be conducted to further support the contributions of the paper (I saw that the authors mention this in future works/limitations, but I think the current paper would benefit from more ablation to make the evidence stronger). In particular, I suggest that the authors (as they also mention) should try to repeat the experiments they present with different models (at least one more) to prove that their results hold in general for "big enough" models. This would be really insightful since it would tell us that different models, even if trained differently, learn similar representations or make use of similar strategies to solve tasks. Furthermore, I think it would be insightful to conduct experiments on larger environments to better understand both to what extent these models are capable of performing In-Context RL and to analyze if, even at larger scale, these models still make use of TD Erros and Q-Values to solve the task
+- One minor concern regards the extent of the novelty of the work: as I mentioned above, although I agree with the authors that it is the first time (to the best of my knowledge) that it was shown that models trained on next-token prediction perform In-Context RL exploiting TD Errors, there are already quite some works exploring TD Learning in Transformers (both at a theoretical and experimental level). Furthermore, the methodology used for the mechanistic analysis is also already well established in the mechanistic interpretability literature.
+
+### Questions
+Some small additional comments and questions I had:
+- In the definition of the Q function in Section 2 (Methods, at page 2), shouldn't there be a conditioning on the initial state and action inside the expectation? Also, shouldn't the sum start from $t=0$ instead of $t=1$?
+- In Section 3, you claim that Llama 3 most likely implements "classic" Q-Learning rather than myopic Q-learning based on the negative log-likelihood. However, in Figure 2, looking at the correlations, it seems that the myopic Q-learning has in general comparable if not higher correlations to the latent representations. Couldn't this suggest that the model is implementing the myopic algorithm instead? Furthermore, is the difference in negative log-likelihood statistically significant?
+- In Figure 5, what do the horizontal lines in subplots B & C represent?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper investigates whether Llama 3 70B has internal representations that support temporal difference learning. First, it demonstrates that Llama can solve RL tasks significantly better than chance. Next, it trains a sparse autoencoder (SAE) and finds features correlated with TD error. Finally, it causally intervenes on these features to show that in-context RL performance degrades without those specific TD features.
+
+### Strengths
+This is an excellent paper. It asks a very interesting question and provides compelling evidence for the conclusion that Llama represents TD error and uses it to solve RL problems in-context. The section on successor representations was a welcome surprise in section 5, and offered more evidence for TD learning, even absent any rewards. The paper was also quite easy to follow and laid out the argument in a very natural way. I don't have any major complaints.
+
+### Weaknesses
+Only minor weaknesses.
+
+1. In the background section on RL, TD is presented for a fixed policy, and then the paper switches to Q-learning, assuming the policy chooses \argmax_a Q(s,a). But this will change the policy as the Q function is updated, so it's not technically the same setting.
+2. It was a bit unclear what "control lesion" referred to in Fig. 2F. And more generally, I was not familiar with the "lesion" terminology, so a brief definition would be welcome. I assume it's a form of activation patching?
+3. I would have liked slightly more explanation regarding "clamping" the activations. I assume this means setting them to a specific value, but how is that different from deactivating them (i.e. clamping them to zero)? Is the purpose of clamping the activations to show degraded, unchanged, or improved performance?
+4. Line 458, mangled sentence "our study is, we have explored".
+
+### Questions
+Could you please provide clarification re: weaknesses 2 & 3?
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+4

human_reviews/3QinqLlMCj.md

@@ -0,0 +1,177 @@
+## Human Reviewer 1
+
+### Summary
+Given G.T. camera intrinsics, the authors first leverage the existing work to obtain the coarse camera poses and depth. Then to refine the estimation, the authors design a module to learn the depth offset estimation with the help of an existing depth estimation network. Furthermore, the camera pose refinement is conducted in another module. The idea of feedforward pose estimation is interesting, but there is still a gap between the performance of the proposed method and some per-scene optimization methods. Since I did not see the authors report any inference time result and I believe some static scene pose-free per-scene optimization methods (CF-3DGS, ..) are very fast and accurate, I expect the authors to provide more comparisons. There are still some questions and limitations raised below. I will consider improving the grade upon the feedback from the authors.
+
+### Strengths
+1. The authors proposed a new pose-free feed-forward method in camera pose estimation.
+2. The authors conducted enough ablation studies to present the contribution of each component of their method.
+
+### Weaknesses
+1. The biggest limitation of this work is the requirement of G.T. camera intrinsic.
+2. The performance on RealEstate-10K seems to be SOTA, however, the performance on ACID is not. Does it mean such a method does not generalize well to the outdoor scenes?
+3. So I expect more comparisons on DL3DV or more public datasets (like DAVIS[1], iPhone[2]) to prove the effectiveness of the proposed method. 
+
+
+[1] Pont-Tuset, Jordi, Federico Perazzi, Sergi Caelles, Pablo Arbeláez, Alex Sorkine-Hornung, and Luc Van Gool. "The 2017 davis challenge on video object segmentation." arXiv preprint arXiv:1704.00675 (2017).
+
+[2] Gao, Hang, Ruilong Li, Shubham Tulsiani, Bryan Russell, and Angjoo Kanazawa. "Monocular dynamic view synthesis: A reality check." Advances in Neural Information Processing Systems 35 (2022): 33768-33780.
+
+### Questions
+1. Can the authors provide more insights on how you obtain the coarse camera parameters? I guess the authors directly implemented the existing work to obtain those things.
+2. the authors trained different checkpoints on different datasets in the implementation details. Does it mean the 'feed-forward' claimed by the authors is actually dataset-specific? If so, I think it is a big limitation of the proposed method lacking the generalizability to different scenes.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper focuses on the pose-free feed-forward novel view synthesis task. It leverages a pre-trained monocular depth estimation model and a visual correspondence model to generate coarse pose and depth estimates, enhancing the stability of the 3DGS training process. Subsequently, a lightweight refinement model is used to further improve the depth and pose estimations. Extensive experiments have been conducted to demonstrate the effectiveness of the proposed method.
+
+### Strengths
+S1 - This paper addresses a meaningful task with significant potential for real-world applications.
+
+S2 - This paper leverages two robust pre-trained models to generate initial pose and shape estimates, which significantly enhance the model's performance.
+
+S3 - The paper is well-written, and the experiments are logically sound.
+
+### Weaknesses
+W1 - The performance of this method appears to be highly dependent on the coarse pose and depth results provided by the pre-trained model.
+
+W2 - The paper lacks qualitative results for the pose estimation, which would provide a clearer assessment of the model's performance in this area.
+
+### Questions
+Q1 - How would the results differ if alternative coarse prediction networks, such as Dust3r[1], Mast3r[2], or others, were used?
+
+Q2 - Qualitative results for the pose estimation task.
+
+[1]Wang S, Leroy V, Cabon Y, et al. Dust3r: Geometric 3d vision made easy[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2024: 20697-20709.
+[2]Leroy V, Cabon Y, Revaud J. Grounding Image Matching in 3D with MASt3R[J]. arXiv preprint arXiv:2406.09756, 2024.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper introduces a method to tackle the challenging problem of novel view synthesis (NVS) from unposed images in a feed-forward manner. They identify the issue in previous pixel-aligned 3DGS methods, where the predicted 3D Gaussians from different views have the problem of misalignment, leading to noisy gradient flows and poor NVS results. They propose a method where they don’t need to rely on the poses obtained from off-the-shelf tools. Instead, they leverage pre-trained monocular depth estimation and visual correspondence models to obtain coarse alignment, with further refinement of the depth map and poses. Results show that among the pose free methods, they can perform decently better results for NVS tasks. However, for pose estimation, it is still worse than general methods like Mast3R.
+
+### Strengths
+1. The authors tackle the problem of 3D Gaussian splatting from unposed sparse images, which is an interesting and important topic
+2. The authors apply the recent state-of-the-art depth estimation and pose estimation methods for coarse pose alignment, and further introduces pose and depth refinements, to some extend improving the final performance.
+3. The paper is overall well-written and easy to follow in most parts.
+
+### Weaknesses
+**The method section is overall clear, but missing some details and discussions**
+
+*Unclear descriptions in camera pose refinement in Sec 3.2.3*  
+1. It is quite unclear to me how exactly it is done. From your writing, it makes me feel that you first will get a newly computed camera poses $\hat{P_{ij}}$ similarly as done in the coarse step, but with the refined depth. This $\hat{P}_{ij}$ is already the refined poses, right? However, you further have another refinement step as shown in Eq. (2). What are the rationale before?
+2. what is the T_pose network? What is the E_pos in eq. (2)?
+
+*Cost volume*  
+In section 3.2.4, for the “cost volume construction and aggregation”, is that any different to the MVSplat paper? Can you justify the differences?
+
+*2D-3D Consistency loss*  
+Line 291-292, you said that you improve the robustness of the model in regions with low texture or significant viewpoint changes. However, the correspondences from the feature matching methods like LightGlue do not provide many correspondences in those low-texture regions. I guess you cannot claim the robustness there? 
+
+*Unclear implementation details*  
+What is the frame distance from 45 to 75? Do you mean you sample one frame every 45/75 frames in the video sequence? You might want to make this point clearer. And why for DL3DV, you only sample every 5 or 10 frames, way smaller than ACID or RE10LK?
+Also, you mention that you train for 40000 iterations on a single A6000 GPU, is it the same for all three datasets? If true, I think it might not make too much sense? As you mentioned in line 351-354, RE10K has ~21K videos from training, while you only use 2K scenes for the training of DL3DV?
+
+
+
+**The experimental section is convincing in general, but lacks some important experiments / baselines and explanation.**
+
+1. For pose estimation comparison, the sota method right now is RoMa [1], I think it is fair to ask to compare to it.
+2. I wonder why your method is still lacking behind Mast3R on the pose estimation for both RE10K and ACID, even if Mast3R is is not even trained at all on those dataset, while your method was trained on those dataset separately.
+3. Why for novel view synthesis in Table 1, you don’t show the comparison to the recent pose-required methods pixelSplat?
+4. Based on the experiments you show, your pose estimation is worse than Mast3R in almost all cases, and the NVS results are also worse than MVSplat in all scenarios (even on DL3DV in Table 3, where MVSplat was not trained on). One reasonable baseline to me is, get camera poses from Mast3R, and then run MVSplat directly. I wonder how your method compares to such a simple baseline?
+5. In your ablation study table 4, what is the point of adding V, I-I, I-II, I-V, if they are just all N.A.? That is really weird to me. You can just describe them in texts.
+
+
+**Writing**  
+Sec 3.2.1: You use two paragraphs to motivate by mentioning the limitations of the previous methods. The real content for the coarse alignment is really just the third paragraph between line 186-191. The motivations part is actually unrelated to the coarse alignment but more on why your method is needed, so you should just put them in the introduction instead.
+
+[1] Edstedt et al.: RoMa: Robust dense feature matching, CVPR 2024
+
+### Questions
+As I already discussed in the weakness section, it would be very imporatnt if you can justify those points in the experiments.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper introduces PF3plat, a novel framework designed for novel view synthesis from unposed images in a single feed-forward pass. PF3plat leverages pre-trained monocular depth estimation and visual correspondence models to achieve an initial coarse alignment of 3D Gaussians. Subsequently, PF3plat incorporates refinement modules and geometry-aware scoring functions to further refine the depth and pose estimates derived from the coarse alignment to enhance the quality of view synthesis.
+
+### Strengths
+1. The task of novel view synthesis from unposed images in a single feed-forward pass is highly practical.
+2. The paper demonstrates state-of-the-art results on Re10k and ACID, showcasing the effectiveness of the proposed method. 
+3.  The refinement modules designed in the paper have significantly improved the effectiveness.
+
+### Weaknesses
+1. PF3plat leverages a robust solver for pose estimation between each pair of cameras; thus, increasing the number of viewpoints significantly extends the feed-forward pass time.
+2. PF3plat relies on the coarse alignment of 3D Gaussians, and a small overlap may affect the quality of the correspondence model.
+
+### Questions
+1. Why does using the pixel-wise depth offset estimation model promote consistency across views? (line 211)
+2. How about the performance of PF3plat in dynamic scenes?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/3Wuvqc4xoy.md

@@ -0,0 +1,153 @@
+## Human Reviewer 1
+
+### Summary
+This article presents an approach to learning representations of neutrino events by leveraging a transformer-based variational autoencoder. The model is trained to capture the photon arrival time distribution, and the learned representations are evaluated using the Jensen-Shannon divergence to assess reconstruction quality. Furthermore, the authors explore the applicability of these representations in a downstream task – angular reconstruction.
+
+### Strengths
+The application of machine learning techniques in scientific research is a vital and rapidly evolving field. We are delighted to see submissions in this area and encourage researchers to share their relevant work.
+
+### Weaknesses
+This article requires significant improvements in its writing and technical accuracy. Numerous technical details are either unclear, incorrect, or require further clarification (see Questions for specific concerns). As it stands, the article's technical clarity is compromised, which may lead to confusion and misinterpretation. A thorough revision is necessary to ensure the article's technical details are accurate, clear, and concise.
+
+### Questions
+* In Fig. 2, the “autoencoder” outputs some probabilities through the softmax activation. This is a confusing design. How is the reconstruction loss applied in this case? 
+* In section 4.2, the training methodology for the three models and the utilization of om2vec are unclear. Can you provide a more detailed explanation of the training process and how om2vec is incorporated?
+* Are there any additional physics features that could be included in the time series data, beyond the current single feature of photon hits?
+* In lines 179-180, the authors wrote “We opted for a learnable memory embedding for the transformer decoder layers, ensuring that the decoder portion of the architecture remains entirely independent of the encoder”. Please elaborate on the memory embedding block about its design.
+* The model and training details in Table 1 are incomplete and unclear. Can you provide a more comprehensive description of the model architecture, including the number of encoder and decoder layers used?
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper presents om2vec, a novel approach leveraging transformer-based variational autoencoders (VAEs) to create compact, descriptive latent representations of photon arrival time distributions (PATDs) from neutrino telescope events. The proposed model is designed to handle the high-dimensional, variable-length data typical of neutrino observatories like IceCube. om2vec aims to outperform conventional approaches, such as asymmetric Gaussian mixture models (AGMMs), by improving reconstruction accuracy, runtime efficiency, and reliability while being less dependent on hyperparameters. The paper details the architecture, training, and testing with simulated datasets, comparing the method’s performance with traditional AGMMs and exploring its utility for downstream tasks like angular reconstruction.
+
+### Strengths
+- Originality: Applying transformer-based VAEs to neutrino event data is novel and demonstrates a creative extension of ML techniques to physical sciences.
+- Quality: Comprehensive evaluation of the model against AGMMs, showing significant improvements in reconstruction accuracy, computational efficiency, and robustness.
+- Clarity: The architectural details, data processing steps, and experimental methods are described with clarity, making the paper accessible to readers familiar with ML and neutrino physics.
+- Significance: The ability to improve data processing and enable better downstream analyses has substantial implications for neutrino research and potentially for other high-dimensional physics datasets.
+
+### Weaknesses
+- Generalizability: While the results are promising, it would be helpful to see a more extensive discussion on how the method might generalize across different types of neutrino observatories or non-simulated real-world data.
+- Comparison Baseline: Although om2vec is compared with AGMMs, additional comparisons with other potential ML approaches (e.g., deep CNNs or LSTMs) for PATD representation might strengthen the case for its use.
+- Hyperparameter Sensitivity: While the model claims reduced dependence on hyperparameters, an exploration of performance variability with different encoder/decoder block configurations or latent dimension sizes would provide deeper insights into its stability.
+
+### Questions
+1. How does the model’s performance vary with different encoder/decoder block architectures or deeper networks?
+2. Can the approach be adapted or extended to handle data from other types of particle physics experiments with different signal characteristics?
+3. Have real-world data tests been considered, and if so, what were the challenges and results?
+4. Is there potential for this method to contribute to real-time data processing in neutrino observatories under field conditions?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This develops a variational autoencoder to create a generative model for data produced by neutrino telescopes. The architecture is based on transformers, and results in a flexible representation and improved computation.
+
+### Strengths
+The application is certainly interesting and compelling. I also like the rationale of the work. There's a clear scientific motivation for these problems.
+
+### Weaknesses
+Several aspects. First, this is an ML focused conference so I would have appreciated greater details on the encoder and decoder without having to dig through the source code. Why transformers as opposed to a simpler architecture? Is there some kind transformation of the features that would allow for an MLP. Even if not, I would appreciate these as baselines as opposed to a traditional statistical model when comparing performance.
+
+Also having worked with these a lot, I'm willing to bet that there was a substantial amount of tweaking required for learning rate and architecture parameters. If not, I'm certain performance can be improved dramatically by taking these steps. Another example, the runtime isn't really compelling to me. This is a feed-forward network, clearly it's going to be quicker than the alternatives. Should be supplementary, which would make more space for the fitting details I discussed.
+
+Overall, this seems written for a scientific audience rather than an ML audience. I very much appreciate the application and clear motivation so I hope it's resubmitted. It just seems like some of the details we find interesting were glossed over and need to be improved for this to be accepted.
+
+### Questions
+Not at the moment, will see other reviewers' comments.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper titled "Learning Efficient Representations of Neutrino Telescope Events" introduces a novel approach called om2vec, which utilizes transformer-based variational autoencoders (VAEs) to effectively represent neutrino telescope events. The study addresses the challenges posed by high-dimensional, variable-length Photon arrival time distributions (PATDs) recorded by optical modules in neutrino telescopes, particularly focusing on the IceCube Neutrino Observatory.
+
+### Strengths
+- The use of a transformer-based variational autoencoder (VAE), called om2vec, represents an innovative approach for neutrino event data analysis, which has traditionally relied on more conventional statistical methods or simple summary statistics. 
+- The paper pushes the boundaries of machine learning applications within high-energy physics, specifically neutrino detection. 
+- By applying a VAE with transformer components to a unique scientific data source, the paper contributes to bridging techniques between disciplines, such as physics, machine learning, and data science. This could encourage further cross-disciplinary research and adaptation of machine learning models to complex scientific problems.
+
+### Weaknesses
+- The paper lacks a clear structure and does not adequately address related work. If this is indeed the first study applying deep learning techniques to the domain of neutrino telescopes, it is essential to include a dedicated **Related Works** section to provide context for this research.
+
+- The figures in the paper are oversized. I recommend the authors resize them to a more standard dimension to enhance the overall presentation quality. The current size does not meet the standards expected for conference presentations.
+
+- There are several typographical errors throughout the paper (e.g., lines 127, 484, etc.), which detract from its readability and should be addressed to improve clarity.
+
+- The objective function is unclear, and the problem is not well-defined. The paper jumps directly to the results, with only a brief discussion of the classical $KL$ divergence. A significant improvement is needed in presenting a comprehensive **Proposed Methods** section that clearly defines the final objective function, rather than merely referring to it in the **Results section** (lines 228 to 230).
+
+- Some statements in the paper are ambiguous or inaccurate. For example, the assertion in lines 223 to 232 that "the re-parameterization trick is utilized to construct the latent representation $z$, a vector of user-defined length referred to as the latent dimension. This technique guarantees that the latent space remains continuous and that similar representations within this space reconstruct to similar PATDs" is misleading and not entirely accurate.  However, the reparameterization trick separates the randomness of sampling (handled by $\epsilon$) from the parameters $\mu$ and $\sigma$, which allows to compute gradients with respect to these parameters. I recommend that the authors deepen their understanding of this concept from this paper [1].
+
+I would be willing to consider increasing my rating, but only if these issues are adequately addressed. As it stands, the current version of the paper is not ready for publication.
+
+**Refrences:**
+
+[1] Kingma, Diederik P., and Max Welling. "An introduction to variational autoencoders." Foundations and Trends® in Machine Learning 12.4 (2019): 307-392
+
+### Questions
+The paper is somewhat limited as it presents results solely based on training and testing with simulated events, which may not accurately reflect real-world measurement data. Given that the approach uses a VAE-based transformer, it may perform better with simulated data that follows known distributions. Do you have access to any existing real-world datasets? If so, I would appreciate your feedback on this aspect.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+1
+
+### Confidence
+4

human_reviews/5bDBahNmmH.md

@@ -0,0 +1,280 @@
+## Human Reviewer 1
+
+### Summary
+The paper proposes Cohesion, a model that combines a deterministic latent autoregressive component and a diffusion model for probabilistic dynamics forecasting. The high-level intuition of the model is as follows: first the current state is encoded to a compressed latent space. Then, one or more steps of a deterministic autoregressive model are applied, after which the predictions in the latent space are decoded back to the data space to get the initial prediction(s). In terms of the Reynold decomposition, the paper interprets this initial prediction as the coherent component of the flow. To resolve the fluctuating component, a diffusion model, which is conditioned on the predicted coherent component, is used to ‘stochastically refine’ the states. Cohesion is evaluated against the Spherical Fourier Neural Operator (SFNO) on Kolmogorov flow and Shallow Water Equation benchmarks, in terms of point wise-metrics like MSE, as well as structure-based and physics-based metrics.
+
+### Strengths
+**S1:** The model is evaluated in terms of spectral divergence, a measure of divergence between the energy distribution at different frequencies. Indeed, for chaotic systems that can have strongly diverging trajectories for even slightly different initial conditions, point-wise metrics become meaningless for long simulation horizons, and the statistical properties of the system should be investigated.
+
+**S2:** The interpretation of the deep Koopman operator as predicting the coherent part of the flow and of the diffusion model as refining the fluctuating component is intuitive and provides a physics-inspired motivation for Cohesion.
+
+**S3:** The method supports different sampling strategies, dubbed trajectory planning and autoregressive. Autoregressive shows higher quality results, but requires more computational effort, whereas trajectory planning is more computationally efficient due to denoising the entire trajectory at once. As such, the sampling strategies provide an intuitive way to trade computational budget for accuracy.
+
+### Weaknesses
+**W1:** I found the presentation of the paper at some parts counterintuitive and confusing. In particular, Section 3 (the method section) starts with a quite extensive explanation of score-based generative models and zero-shot super resolution, while this concerns background material introduced in other works. I think the paper would benefit from merging this part with the text around Eq. 3 in the background section. This would make it easier for the reader to distinguish already established methods/algorithms and the novel aspects introduced in this paper.
+
+**W2:** The experimental evaluation has two weak aspects:
+
+* The method is compared against a single baseline only, the SFNO. There are other popular probabilistic methods that utilize a predictor-refinement approach, for example those that are cited in Section 2, e.g. Lippe et al. 2024; Srivastava et al. (2023); Yu et al. (2023); Mardani et al. (2024). It would greatly improve the paper to compare against a selection of those methods, and explain how the key conceptual differences in your approach relative to those works lead to different results. In addition, the work of Bergamin et al. [1] is relevant since it also considers two sampling modes that are highly similar to trajectory planning and autoregressive forecasting as described in the paper.
+* The SFNO is not a suitable model for the Kolmogorov flow experiment, since the geometry in this experiment is not spherical. Do you have a specific reason to not compare to the regular FNO instead here (in addition to other baselines that would be good to add)?
+
+**W3:** After reading the paper, it remains unclear to me what the key novel insights are relative to prior papers. Diffusion-based refinement techniques have already been proposed, as cited in Section 2 of this work. In addition, the zero-shot conditioning on noisy or partial observations in the context of PDEs has already been established in [2]. As I currently understand, the novelty here lies primarily in the architectural choice of using the encoder-operator-decoder module to get the initial prediction. In itself, this seems a conceptually small change relative to earlier frameworks that show that similar predictor-refinement approaches work well. If using this architecture in the prediction-refinement setting could lead to a substantial performance increase or otherwise interesting results, these could be novel insights, but this is not investigated in the paper. I think it would help to get your point across if you contrast your work against the most similar related papers more explicitly, and highlight the differences between them.
+
+**References:**
+
+[1] Bergamin et al., 2024. Guided Autoregressive Diffusion Models with Applications to PDE Simulation. https://openreview.net/forum?id=1avNKFEIOL
+
+[2] Rozet & Louppe, 2023. Score based data assimilation. https://arxiv.org/abs/2306.10574
+
+### Questions
+**Q1:** Did you investigate whether the predicted conditioning prior actually aligns with the coherent part of the flow? I.e., does it predict some kind of average (or perhaps localized/moving average) behavior of the dynamics in space and/or time?
+
+**Q2:** It is unclear to me how the metrics are calculated over multiple samples from the probabilistic models. Did you simply average the metrics over those samples, or take a best-of-K like approach? What is the effect of the stochastic sampling on the calculation of those metrics using your approach?
+
+**Q3:** Please add labels on the horizontal axes of Figures 5 and 7.
+
+**Q4:** How should Figure 6 be interpreted? Is the spherical geometry here somehow projected on the 2D image? Is there a reason why the signal is only present at the top half of the image?
+
+**Q5:** Can you also provide plots of the spectral divergence over time, of Cohesion, the baseline methods, and coherent-only?
+
+**Q6:** Please comment on W2 and W3.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper presents Cohesion, a Coherence-Based Diffusion Model for Long-Range Dynamics Forecasting. The model leverages the concept of Reynolds Averaged Navier-Stokes (RANS) as conditioning priors to address two key issues in diffusion-based autoregressive models: (1) instability in long-term predictions, and (2) the computational inefficiency of generating priors. The authors utilise a Koopman-based reduced order model to efficiently generate these priors, thereby speeding up the forecasting process. Standard fluid dynamics principles and diffusion models are adapted to support this framework. Furthermore, cohesion acts as a refinement mechanism that aggregates temporal sequences from the model’s output, improving performance. The proposed approach is evaluated on two benchmark fluid systems. While the work is well-motivated and promising, there are critical issues with the presentation and formulation (detailed below).
+
+### Strengths
+The targeted problems are both interesting and critical for applying diffusion models to autoregressive forecasting. The main ideas and design choices presented in this paper are well-motivated and promising.
+
+### Weaknesses
+### Presentation 
+1.	The authors claim that trajectory planning, a concept central to reinforcement learning (RL), is crucial to Cohesion. However, they do not adequately explain the operational or conceptual connections. Although they mention reframing forecasting as trajectory planning, no clear relationship to RL principles—beyond standard autoregressive or multi-step predictions—is evident. Without a specific RL objective formulation or a direct link to decision-making scenarios, the reference to RL seems forced and mislead readers.
+
+2.	The paper's use of terminology is potentially confusing and could be considered an abuse of terms. Specifically, the names "Cohesion" and "coherent" are too similar, which creates ambiguity about their distinct roles. From personal understanding, the cohesion refers to the whole framework whereas “coherent” describes a predictable component of the dynamical system. While, as the figures 1,2,3 and experiment sections, the "cohesion " seems refer to the temporal-aggregation component. In addition, given the two words have specific meanings in the scientific community, the author should be careful and avoid potential confusions and abusing of terminology. 
+
+3. Koopman Operator Introduction (Lines 253-256): The introduction of the Koopman operator is vague and contains some mathematical issues: (1) the text does not discuss that practical implementations use a finite-dimensional approximation. (2) The domain and mapping of the encoder and decoder are not clearly defined. (3) No clear definition for $\mathcal{G}$ and $\mathcal{G}_E(\mathcal{X})$. 
+
+4. The numerical results are discussed too briefly, with only three lines dedicated to each experiment (Lines 378-380, 432-434). More in-depth discussion and analysis are needed to properly interpret the findings.
+
+5. There is only one baseline methods. Additional baselines, particularly diffusion-based models, should be included for a more comprehensive comparison.
+
+## Others
+1.	The paper uses incorrect citation formatting. Citations should be enclosed in parentheses, e.g., "(xxx et al., year)," when they are not acting as the subject or object within sentences.
+2.	Figures 5, 7, 8, and 9 lack axis descriptions 
+3.	Several acronyms, such as Number of Function Evaluations (NFE), are introduced without definition at first mention. All abbreviations should be defined before initial use to ensure clarity.
+
+### Questions
+1. The Spherical Fourier Neural Operator (SFNO) is used as the only baseline despite being designed for spherical domains, whereas both demonstrated cases in the paper use standard square domains. This choice makes SFNO an unsuitable and potentially misleading baseline. It is unclear why the authors included this baseline, as it does not align with the problem setting. 
+
+2. The use of Reynolds decomposition is indeed a central aspect of the paper, and this core idea is not clearly stated. While RANS typically deals with time-averaged components, the authors extend it to a time-dependent setting without discussion on this extension and theoretical foundation. It remains unclear how the authors achieved this extension and ensured its validity.
+
+3. The authors claim improved inference efficiency through incorporating Koopman-based ROM; however, no comparison with other baselines is provided to support this claim.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper proposes a diffusion-based approach for forecasting with dynamical systems that is able to generate the entire sequence in one conditional denoising pass. This is achieved by leveraging reconstruction guidance, where the conditioning information is a sequence of priors (one for each state of the final trajectory), generated by iteratively applying a (lightweight) reduced-order method (ROM) to the initial condition. The score network operates over subsequences, and temporal coherency is assured by applying temporal convolution with a small receptive window. The experiments are performed on two chaotic systems (Kolmogorov flow and Shallow Water), and the performance of the model as a probabilistic emulator is tested in terms of pixel-based metrics (RMSE, MAE), structure-based metrics (MS-SSIM), and physics-based metrics.
+
+### Strengths
+1. **Relevant topic.** The problem addressed in this paper (probabilistic emulation for PDEs) is an active area of research with important downstream applications, such as weather and climate modelling.
+2. **Non-autoregressive approach.** The possibility of using a non-autoregressive sampling strategy from the diffusion model without compromising too much on accuracy is nice, and something of significance for the field of forecasting.
+3. **Flexibie guidance with ROMs.** The use of reconstruction guidance is a very useful technique when the observation process changes over time, offering more flexibility as opposed to classifier-based approaches. Although the technique is not a contribution of this work, the paper is the first (as far as I am aware) to utilise as conditioning information a sequence of autoregressively-produced priors through a compute-efficient ROM. The experiment in which the authors condition on partially-observed fields of the ROM output highlights the flexibility of this technique and is relevant to real-life settings with partially-observed data.
+5. **Wide range of metrics used for the experiments.**
+
+### Weaknesses
+1. **Unclear contributions/Lack of clear citations.** This approach shares striking similarities with the approach from Rozet et al. [1], but fails to properly reflect this throughout the main text. There should be a much more clear distinction between the contributions of this paper and what is taken from other works.
+- The overall training and sampling from the score-based model relies heavily on the approach proposed by Rozet et al. [1], but this is not made clear in the paper. They also train the model on subsequences of length $W$ (justifying this approach from a Markov order perspective), and stitch these subsequences together at sampling time to generate arbitrary length trajectories in one go.
+- The reconstruction guidance mechanism is exactly the same as the one proposed in Rozet et al. [1]. This is mentioned in the paper (L215), but it can be interpreted as if the authors propose a way to improve the numerical stability of the method, as opposed to using results from previous work.
+- The similarity between the proposed framework and Rozet et al. [1] is reflected in the algorithms presented. 
+   - Algorithm 1 is the same as Algorithm 3 in Rozet et al. [1]
+   - Algorithm 2 is the same as Algorithm 4 in Rozet et al. [1]
+   - Algorithm 3 is the same as Algorithm 1 in Rozet et al. [1]
+   - Algorithm 4 is the same as Algorithm 2 in Rozet et al. [1]. However, in the paper this is posed as a novel temporal convolution technique, rather than something already employed in Rozet et al. [1].
+
+  I acknowledge that this paper adapts Rozet et al. [1]’s approach, making it suitable to other tasks (forecasting), as opposed to data assimilation. This is achieved by conditioning on those prior states, generated autoregressively through a ROM. This is a nice approach and equips Rozet’s method with the ability to perform forecasting, a task where their approach fails based on the observations from Shysheya et al. [2]. However, I do not think this is how the paper portrays the technique, and it is debatable whether this is enough of a contribution overall when considering the results (see below). At the very least, a clear paragraph on contributions should be included.
+2. **Weak baseline for the empirical analysis.** Although the paper mentions that the SFNO approach [Bonev et al. [4]] is the state-of-the-art, I believe there are other probabilistic forecasting approaches that achieve stronger performance.
+- Two works I have in mind are Lippe et al. [5] and Shysheya et al. [2], with the former being considered state-of-the-art. However, the main metric these works consider for forecasting is high correlation time, rather than the metrics used in this work. But based on the trajectories, they seem to maintain correlation with the ground truth for longer than Cohesion. It would be interesting to compute the high correlation time and compare it with some of these works, especially since the Kolmogorov dataset looks similar to the one in Shysheya et al. [2].
+- As in Lippe et al. [5], I believe that another relevant (deterministic) baseline would be an MSE-trained UNet. In their experiments, it tends to achieve better results than FNO-based approaches.
+3. **Lack of error bars in the results.** The performance plots lack error bars. Thus, it is hard to determine how significant the difference between methods is. This is especially the case between Cohesion (R=1) vs. Cohesion (R=T) (i.e. autoregressive vs. in one go), where having error bars is important to figure out whether the hit in performance by generating the entire trajectory in one go is significant.
+4. **Copying from other papers without citing.** There are certain paragraphs/sections in the appendix which are directly taken from other works without specifying so. For example Appendix B.2. Structure-based metrics is the same as Appendix F.1.4. Multi-scale Structural Similarity Index Measure (MS-SSIM) in Nathaniel et al. [6], Appendix B.3 is very similar to Appendix F.2 in Nathaniel et al. [6]. It is ok to use the same definitions as in other works (in the end, the definitions of the metrics are what they are), but if the writing is so similar, I think you should at least cite the relevant work. 
+   
+   Could you please review these sections and either rephrase them, or make it clear that they are heavily based on Nathaniel et al. [6]?
+
+**Minor**
+
+5. **Typos.** The paper contains several typos, I won’t include them all but examples are L091-spatiotempral, L125 - deterministic priors, L235 - should be $\nabla_{\mathbf{u}_k}$ , etc.
+6. **Small labels in figures.** See for example y label in Figures 5, 7, legend in Figure 8.
+7. **Unclear figures.** I appreciate the attempt to create a visual representation of the framework in Figure 1, but I find the figure confusing, without mentioning the colour coding of the bubbles, why there are two trajectories in b), etc.
+8. **Lack of x label in Figure 5**. I believe that is the timestep $T$, but that should be labelled.
+9. **Ablation hyperparameter choices.** In appendix C.1 you detail how you chose the dropout rate $p$ and perturbation factor $f$ for the baselines. However, those correspond to the lowest values you experimented with and it’s unclear whether going even lower would give better results or not. You’d like to obtain a convex function of your hyperparameters (i.e. worse performance for lower and higher values of the hyperparameter).
+
+Overall, while the topic addressed by this paper is relevant, I believe it does not clearly differentiate its contributions to what already exists in the literature. For the empirical evidence, I believe the chosen baseline is not strong enough, and a more comprehensive comparison to other techniques is needed to thoroughly assess the effectiveness of the proposed method. Although in its current form the paper is incomplete, with proper baselines and a clear indication of novelty/contributions, it could represent a useful addition to the literature.
+
+[1] Rozet, F., & Louppe, G. (2023). Score-based Data Assimilation. ArXiv, abs/2306.10574.
+
+[2] Shysheya, A., Diaconu, C., Bergamin, F., Perdikaris, P., Hern'andez-Lobato, J.M., Turner, R.E., & Mathieu, E. (2024). On conditional diffusion models for PDE simulations.
+
+[3] Qu, Y., Nathaniel, J., Li, S., & Gentine, P. (2024). Deep Generative Data Assimilation in Multimodal Setting. 2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 449-459.
+
+[4] Bonev, B., Kurth, T., Hundt, C., Pathak, J., Baust, M., Kashinath, K., & Anandkumar, A. (2023). Spherical Fourier Neural Operators: Learning Stable Dynamics on the Sphere. International Conference on Machine Learning.
+
+[5] Lippe, P., Veeling, B.S., Perdikaris, P., Turner, R.E., & Brandstetter, J. (2023). PDE-Refiner: Achieving Accurate Long Rollouts with Neural PDE Solvers. ArXiv, abs/2308.05732.
+
+[6] Nathaniel, J., Qu, Y., Nguyen, T., Yu, S., Busecke, J., Grover, A., & Gentine, P. (2024). ChaosBench: A Multi-Channel, Physics-Based Benchmark for Subseasonal-to-Seasonal Climate Prediction. ArXiv, abs/2402.00712.
+
+### Questions
+1. Could you please highlight the differences between this approach and Rozet et al. [1]? Is there anything that wasn’t captured in **W1**?
+2. Could you also compute the high correlation time in the Kolmogorov (and SWE) experiment to compare with other results reported in the literature (i.e. Lippe et al. [5], Shysheya et al. [2])
+3. While in the SWE experiment I understand the usefulness of the spherical embeddings used in SFNO, it is not clear to me why they would help in Kolmogorov, but I might have missed some relevant experimental setup detail that justifies it.
+4. Could you provide error bars for your results?
+5. If you provide comparisons to Lippe et al. [5] on the Kolmogorov flow experiment, could you also provide computational speed comparisons?
+6. In L205 you assume a Gaussian observation process (as it is usually done). Would you be able to extend the framework to a non-Gaussian likelihood too?
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper introduces Cohesion, a framework for probabilistic dynamics forecasting in chaotic systems like fluid dynamics. By reframing forecasting as a trajectory-planning task, the framework makes use of reduced-order models (ROM) to make denoising processes more efficient. This approach is much faster because it applies a single denoising pass to the whole forecast sequence, rather than using multiple autoregressive steps. Cohesion also includes a way of guiding the process without using classifiers, which makes it suitable for zero-shot forecasting. Tests on the Kolmogorov Flow and Shallow Water Equation show that Cohesion works better than other methods for capturing multi-scale physical structures and reducing spectral divergence, which is important for modeling chaotic systems accurately.
+
+### Strengths
+The framework is innovative in its approach to combining turbulence theory, conditional generative model, and reinforcement learning principles for forecasting. By applying diffusion models with coherence-based conditioning, the paper contributes a novel perspective on long-range forecasting. The idea of Turbulence-diffusion framework and the induced Zero-shot conditional sampling is interesting and inspiring.  Cohesion presents a promising solution for long-range dynamics forecasting, especially in chaotic and partially observed environments. Experiments are exhaustive.
+
+### Weaknesses
+**Over-claims:**
+
+**Reinforcement Learning:** After reading the methodologies, I didn't see RL play any role in this work. The only related part is borrowing the idea from Diffuser (Janneretal. 2022), to do the (sub) trajectory generation rather than autoregressive generation for efficiency. Such point and even claiming to achieve stable long rollouts with RL are invalid to me, given that Diffuser is a purely a generative method but to solve RL problems only. The critical components like **value function and reward function** never appear. 
+
+
+**Zero-shot:** While zero-shot forecasting is addressed conceptually, further clarification on how this aspect was validated experimentally is lacking.
+
+
+
+**Writing:**
+
+- Abstract: some sentences are naively extracted from the introduction but unclear after compression. "Nonetheless,Cohesionsupports...", unclear what's the advantage the authors refer to. Specifically, why iterations over subsequences are not allowed in the previous works. And direct abbreviation "NFEs".
+- Section 2: The demonstration of $u_K$ and the relationship between $u_K$ and $u$ should be exposed clearer. E.g. change the title of "Coherent flow as conditioning prior" to "Conditional Diffusion Modeling", then put the demonstration of $u_K$ here first, and then introduce the coherent flow is the conditioning prior.
+
+### Questions
+1. **On Model Flexibility**: ROM is a type of method with limited expressiveness e.g. Koopman Operator, a linear approximation model. Would Cohesion be adaptable to domains where coherent flow cannot be efficiently approximated by ROM? 
+2. **Baselines:** SFNO seems the only baseline. How about others like Markov Neural Operator (MNO)?
+3. **Long-Term Stability**: The results are over long rollouts, how long is it and how hard is it to predict that?
+4. **Ablation Study** has been lacking, such as W window size.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+3
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+The paper proposes Cohesion, a coherence-based diffusion model for long-range dynamics forecasting, aimed at addressing challenges in autoregressive probabilistic forecasting.
+
+### Strengths
+- The integration of turbulence and diffusion principles with ROM-based conditioning is novel and provides a new approach for multi-scale and chaotic systems.
+- This paper uses quantitative (RMSE, MAE, MS-SSIM) and physics-based metrics (spectral divergence), provide strong empirical support.
+
+### Weaknesses
+- Lack of interpretability. How to prove the role of coherent flow after decomposition and how it promotes long-term stability prediction? Lack of theoretical explanation.
+
+- Please explain what the zero-shot forecasts without classifier and multi-scale physical structure mentioned in the paper are. There exists confusion in the statements.
+
+- Please explain why SFNO is used as the baseline. SFNO is mainly used to predict atmospheric dynamics. Intuitively, the datasets used in this paper are not suitable for spherical geometry operators. Please give an explanation.
+
+- The baseline only uses SFNO. Why not compare other neural operator models, such as CNO[1], UNO[2], LSM[3], etc.
+
+- The model is based on Diffusion. Why not compare diffusion-based models, such as PreDiff[4] and DYffusion[5].
+
+- The time complexity comparison analysis of the model should be increased.
+
+
+[1] Bogdan Raonic wt al. 'Convolutional Neural Operators for robust and accurate learning of PDEs.' NeurIPS2023.
+
+[2] Md Ashiqur Rahman et al. 'U-NO: U-shaped Neural Operators.' TMLR2023.
+
+[3] Haixu Wu et al. 'Solving High-Dimensional PDEs with Latent Spectral Models.' ICML2023.
+
+[4] Zhihan Gao et al. 'PreDiff: Precipitation nowcasting with latent diffusion models.' NeurIPS2023.
+
+[5] Salva Rühling Cachay et al. 'DYffusion: A Dynamics-informed Diffusion Model for Spatiotemporal Forecasting.' NeurIPS2023.
+
+### Questions
+Please address the questions in the Weaknesses.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3

human_reviews/61ss5RA1MM.md

@@ -0,0 +1,185 @@
+## Human Reviewer 1
+
+### Summary
+The paper proposes a novel method based on optimal control to optimize generations obtained by ODE-based generative models (e.g. flow matching). The paper proposes algorithms for generative models in Euclidean space and in SO(3), generalizes previously existing approaches, and studies the convergence of the proposed method.
+
+### Strengths
+The paper tackles the problem of changing the generation process of ODE-based generative models in order to produce samples that maximize a certain reward, while staying close to the original ODE trajectory (through a regularization term). This problem is relevant in multiple domains where additional signals/information are available at inference time. 
+
+To the best of my knowledge, this is the first paper to formalize this guidance and control framework in SO(3), which is a group used by many methods in structural biology.
+
+The approach generalizes existing methods that optimize the trajectory of ODE-samplers (D-Flow, GradFlow), and outperforms them in multiple benchmarks.
+
+### Weaknesses
+**Computational cost.** The method proposed in the paper, as well as its predecessors (D-Flow, GradFlow) all require optimizing the sampling process *for each sample* produced by the generative model. In other words, producing a single sample requires solving an optimization problem, for which computing the loss requires simulating the full ODE. This has a high cost in memory and computation time.
+
+While GradFlow proposed a clever way of reducing the memory cost of this process, which is also adopted in this paper, this optimization process is inherently computationally expensive, significantly increasing the time cost of producing each sample. Previous work used some approaches to try to alleviate this (e.g. FlowGrad uses an adaptive solver to minimize the number of steps used during generation, by setting the step-size as a function of the estimated curvature of the flow at the current point), but simulation is still considerably slower than the baselines without this optimization process. While generation times are not discussed in this work, the D-Flow paper states that producing a single molecule takes around 3 minutes (they use 100 function evals for discretization of the ODE), and for images the time to generate a single output ranges from 4 to 15 minutes depending on the task. While the purpose of these works is not increasing generation efficiency, but generating better samples through guidance and optimization, they exacerbate the main limitation of diffusion models / flow models, which is their slow generation. In the paper I am unable to find generation time for the experiments. Given the method's nature, I think these should be reported and discussed.
+
+I think related to this point, experiments tend to be on the smaller end. Celeba-HQ for images, molecule generation with up to 9 heavy atoms, and peptides, which are short proteins (less than 50 residues). I understand these methods are able to produce better samples given the external guidance, while other approaches are unable to leverage such information, which is quite valuable. Still, I think providing values for computational cost / generation time, and comparing against plain approaches (baselines that do not require tuning) would be good. I would expect the cost of producing one sample is between 10x-100x more than baselines that do not optimize the sampling process (since there are ~20 optimization steps, and some gradient computation too), but happy to be shown otherwise. This does not account for the fact that lower memory requirements by the baselines would allow producing more samples in parallel too.
+
+### Questions
+Proposition 1. What is the prior terminal point $x^p$? Is it $x^p_0$ or $x^p_1$ (that is, terminal as in time $t=1$ or $t=0$). If $t=0$ then the joint $p_1(x^p, x_1)$ is a delta distribution (since ODE is deterministic)? If $t=1$ then $x^p$ and $x_1$ are independent (since $x^p$ is generated from random noise independent of $x_1$)?
+
+GradFlow could in principle be used for manifolds too? The control terms would live in the tangent space of the manifold at the current point? They do not propose this in GradFlow so this is not something to compare against, and would even be consider a novelty and addition to the paper. But this work got me wondering if there’s any obvious reason why this would fail?
+
+I’m not sure I completely agree with the paper’s title “Training free…”. I understand training is the same, and that this method can be used for any pre-trained flow model. But it does require solving an optimization problem, albeit with few iterations (~20) but expensive ones. The difference is that this optimization happens at inference time.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper attempts to solve the problem of conditional generation using Flow-Matching models. In particular, they propose a unifying framework (OC-Flow) from which other approaches (such as D-Flow and FlowGrad) can be derived, and operates in Euclidean and SO(3) geometries. Subsequently, the authors provide extensive theoretical analysis of OC-flow to prove convergence and theoretical properties. Finally, the authors apply OC-Flow to text-guided image generation and peptide design tasks.
+
+### Strengths
+* The paper has a strong theoretical grounding, is well placed within the existing literature and goes to extensive efforts to prove theoretical properties of the proposed methodology. Additionally, the paper provides significant context and background, referencing existing works in Flow-Matching models. Finally, the authors provide a ton of detailed proofs in the appendix, and theoretical analysis in the main body of the text.
+* The paper provides a legitimate contribution to formalizing and extending existing guided-flow matching techniques to complex geometries (such as SO(3))
+* The authors compare the proposed methodology to similar existing methods to demonstrate competitive performance
+* Table 1 gives a good comparison to understand the contribution of OC-flow vs D-Flow and Flow-Grad
+
+### Weaknesses
+__Theoretical Concerns__
+* The main concern with the theoretical aspects of the paper is that the authors perform all the theoretical analysis in the continuous regime, but then implement the practical algorithms in a discrete regime. They even mention this limitation in the conclusion (“we also note that our practical algorithm…”). This presents a potentially significant hole in the paper, as the objects being proved, and the objects being validated empirically are not the same, and thus the theoretical content in the paper is not necessarily applicable to the experimental results (as noted by the author). To address this, we recommend that the authors 1) provide theoretical analysis of the discrete regime of the algorithm, 2) discuss in detail how the continuous approximates or bounds the behavior of the discrete implementation, or 3) implement a continuous version of the algorithm. These additions would help bridge the gap between theory and practice.
+
+__Experimental Concerns__
+* There is a significant lack of information regarding the experiments and reproducibility. There are no details as to how these models are trained, what the architectures are, the implementation, etc. This would need to be addressed before the paper could be accepted. To address this, we recommend that the authors give detailed descriptions of the model architectures and hyperparameters, training procedures and optimization details, data preprocessing steps, computational resources used, and code availability or plans for release. This would greatly improve the reproducibility of the work.
+* Similarly, no error bars/confidence intervals are reported on any of the experiments. In fact, it is unclear whether their model is better than existing baselines (i.e. in table 5, they claim 0.795 is better than 0.793, but no CI, similarly in tables 2, 3 and 4, they report outperforming existing methods but do not report CI.). Consequently, it is not possible to assess their claims that they outperform existing models, given the lack of confidence intervals and the close proximity of the performance values. To address this, we recommend that the authors 1) run multiple trials and report mean and standard deviation for all metrics, 2) perform appropriate statistical significance tests (e.g. t-tests) when comparing to baselines and 3) include error bars or confidence intervals in all tables and figures.
+
+* Finally, one of the major claims in the paper is that OC-flow can optimize in euclidean and SO(3) space, and that optimization in SO(3) provides benefits in tasks such as protein design. However, the authors do now present results split into Euclidean and SO(3) algorithms. It is unclear how/if the extension to SO(3) is even beneficial, and additional experimental details/results are needed to validate this claim as well. This is brought up by "our OC-Flow method, fully optimized in both Euclidean and SO(3) space" on page 10.
+
+__Presentation Concerns__
+* The paper has some issues with the presentation that make it quite difficult to asses which parts are novel contributions, and which parts are existing works that are being used. The authors/paper would benefit greatly from having a clear vision of what they are proposing and why, and subsequently moving large portions of the detailed proofs to the appendix to not muddy understanding with unnecessary detours. For example, what are co-state flow and E-MSA, and why do we care about these constructs? How do these constructs factor into the actual problem of performing guided matching flow generation? Are they purely used for proving convergence analysis? And if so, then it should be framed/explained as such. In fact, the structure of the paper would greatly benefit from having a section which clearly describes the proposed methodology in terms of implementation, and a separate section for the theoretical analysis of the proposed method, since the current structure makes it very difficult to separate the method as-such, from the additional theoretical concepts only necessary for proving convergence.
+  * To address these issues, we recommend that the authors clearly delineate novel contributions from existing work, as well as practical details from theoretical proofs. Adding sections such as "contributions", "proposed methodology and implementation", and "theoretical results" would greatly improve the structure of the paper.
+  * Additionally, we recommend that the authors provide clearer explanations of key concepts such as co-state flow and E-MSA, as well as highlighting their importance to the key contributions in the paper.
+  * Finally, by restructuring the paper to highlight the novel aspects, and moving detailed proofs to the appendix, the overall quality and clarity of the paper would be much improved.
+* Furthermore, several significant objects/theorems are introduced with very little explanation. For example, co-state variables are introduced as “shadow prices representing the sensitivity of the optimal value function to changes in the state variables”. But what are shadow prices? How does this analogy help when there is little thought/exposition given to the co-state/Hamiltonian introduced by the PMP? I would like to see a much more principled approach to writing, where each theorem introduced is clearly placed within the larger context of the work, and has a clear purpose in support of theoretical results.
+* Similarly, the lack of figures greatly hinders understanding. Additionally, figure 1 does not clearly articulate what it is presenting, and what the various sub-figures and equations represent.
+
+__Contribution Concerns__:
+* First of all, due to the presentation it is not clear what the contributions of the paper are, and what is preexisting work being leveraged for proving theoretical properties of the method. However, my understanding is that there are two main contributions of the paper: 1) they formulate conditional generation using flow-matching models as a control problem in equation 2, and 2) given that formulation, the authors demonstrate that OC-Flow is a generalization of D-Flow and Flow-Grad that can be optimized in SO(3), as well as proving various convergence properties.
+* In light of this understanding, it seems like the contributions of the paper are limited in scope. First of all, equation 2 seems to be fairly trivial extension of existing Flow-Matching/Continuous Normalizing Flow formulations (see Fjelde et al (2024)). Furthermore, given the lack of definitive experimental results, it is unclear whether this extension provides tangible benefit over existing methods, especially when considering the additional complexity. One of the major proposed benefits of the method is optimization in SO(3), but no ablation studies are given to demonstrate that SO(3) provides additional benefits over simple euclidean optimization.
+* Additionally, the experimental reproducibility of the paper is quite poor, with no experimental parameters given, and no experimental source code provided.
+*Finally, the scope of the contribution is somewhat niche. In particular, this paper focuses on classifier-guided generation using flow-matching models in SO(3). While useful for certain problems, it likely does not have wide-reaching implications outside of a few target applications.
+
+
+__Citations__
+* Fjelde, T., Mathieu, E., & Dutordoir, V.. (2024). An Introduction to Flow Matching.
+
+### Questions
+* I would like to see a comparison of the runtime of OC-flow vs the other methods. While Table 1 suggests that the memory consumption of OC-flow is lower than D-flow and on par with Flow-Grad, I would be concerned that the additional complexity of solving in SO(3) adds significant computational costs.
+* I would like to have a better understanding of what parts of the paper are core to the methodology (i.e. actually implementing OC-Flow), versus what parts of the paper are necessary for proving convergence. I would then like to see separate sections/subsections for the proposal, and the subsequent analysis.
+* I would like to see a clearer presentation of a conventional flow-matching model, and how the proposed method extends this standard formulation, ideally in the form of before/after equations to get a clear and unambiguous idea of the elements being added/proposed.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper provides a novel approach for guided generation using pre-trained diffusion and flow matching models. Traditional methods of guiding ODE-based generative models often require expensive retraining and work mainly on Euclidean manifolds, but OC-Flow uses an optimal control training-free framework beyond Euclidean spaces to the SO(3) manifold. Experiments on tasks like text-guided image manipulation, conditional molecule generation, and peptide design validate the method’s effectiveness​.
+
+### Strengths
+- As far as I know, the approach is original in framing guided flow matching as an optimal control problem. The authors develop a general framework for non-Euclidean geometries with strong theoretical backing, which is fairly rare.
+- Another strength of this work is that it is a general approach, i.e. OC-Flow can be used effectively for a variety of applications such as image and molecular data. 
+- Unlike existing approaches, OC-Flow allows training-free guidance, making it computationally efficient and more applicable in real-life settings. 
+- The SO(3) results such as on improved molecular generation accuracy, validate the importance of using this geometric inductive bias for generative models. 
+- Through the framing of existing approaches as special cases under their optimal control formulation, this paper helps clarify the connections between gradient-based techniques like FlowGrad and D-Flow.
+- The model consistently demonstrates improvement over previous work.
+- I really appreciate the figures included in the paper to illustrate and compare the methods against existing approaches.
+
+### Weaknesses
+- My main question is regarding scalability. While the model performs well on selected benchmarks, to me it is still unclear how OC-Flow scales to high-dimensional datasets such as large molecules. A discussion of potential scalability limits and memory efficiency in such cases would strengthen the paper.
+- Moreover, even though the formal contributions are great and well-formalized, to me the paper is still quite hard to read. Since the theoretical results are one of the main contributions, I think it would be valuable to add more intuitive explanations of the proofs and why they are there. For example, theorem one provides a bound based on VFM on KL between the model and a terminal point, but some intuition of why this bound is provided would make the paper more approachable.
+- Adding to this point, the theoretical assumptions made (e.g. Lipschitz continuity, boundedness) are clear and needed for the argument, but some reflection (perhaps on a high level) on whether these hold in practice would help to interpret the method's advantages.
+- While the focus is on continuous CNFs, a brief comparison with discrete flow techniques could contextualize OC-Flow's advantages or limitations more clearly, especially as discrete methods have shown promise in similar applications.
+
+### Questions
+- The paper shows promising results, but could the authors elaborate on potential ways to enhance scalability, especially when applied to e.g. large molecules or more complex target distributions in general? 
+- How does OC-Flow compare to recent works in Riemannian FM? What about SO(3) and SE(3)?
+- Could OC-Flow be adapted for hybrid tasks where both Euclidean and Riemannian components are present? This might extend its applicability to even broader fields where you need this hybrid. Does the method allow for this directly or not?
+- Since OC-Flow is designed to be computationally efficient, could the authors comment on real-time applications requiring 'immediate' guidance?
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposed a new framework for controlled generation using pre-trained diffusion and flow matching models, dubbed OC-Flow. The method is based on sound theory in optimal control that offers additional convergence guarantees in Proposition 1 and Theorem 2 (under two key assumptions of affince Gaussian path and Lipschitz continuity of the gradient of guided loss). Several benchmarks on guided-image manipulation, molecular generation and protein design with generative models are performed to demonstrate the effectiveness of the method.
+
+### Strengths
+- Well-motivated problem, overall nicely written paper with clear literature review.
+- The methodological and theoretical parts of the paper are well-sounded. 
+- Providing a framework that has convergence analysis is always welcomed.
+
+### Weaknesses
+**Major: questionable and inconsistent baselines' results in empirical benchmarks** 
+
+- While on the first task (section 5.1 text-guided image manipulation) the authors have report/insert exactly other baselines' results (originally in Table 2 of the FlowGrad paper); the results on two remaining tasks in section 5.2 (molecule generation) and section 5.3 (peptide design) do not match the results reported in their respective original paper. 
+- More specifically, the results in Table 3 do not match those of Table 4 in D-Flow paper (Ben-Hamu et al. 2024); results in Table 5 do not match those of Table 1 in PepFlow paper (Li et al. 2024). In fact, if one instead takes into account the original results, the baseline D-Flow actually perform better in MAE metrics compared to OC-Flow in Table 3. For Table 5 the metrics reported are in different scale. 
+- I am therefore request the authors to clarify this discrepancies between results reported in their paper and the results reported in the respective original works of compared baselines. Otherwise, I think the practical performance of OC-Flows remains questionable.
+
+Ben-Hamu et al. (2024). D-Flow: Differentiating through Flows for Controlled Generation. Proceedings of the 41 st International Conference on Machine Learning, Vienna, Austria. PMLR 235, 2024.
+
+Li et al. (2024). Full-Atom Peptide Design based on Multi-modal Flow Matching. Proceedings of the 41 st International Conference on Machine Learning, Vienna, Austria. PMLR 235, 2024.
+
+### Questions
+See weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/6e3hoDZKuO.md

@@ -0,0 +1,159 @@
+## Human Reviewer 1
+
+### Summary
+This paper explores a new method for adapting large language models (LLMs) for goal-directed dialogues using reinforcement learning (RL). The key innovation in this work is the introduction of an "imagination engine," which synthesizes hypothetical human-human interactions based on task descriptions. These imagined dialogues serve as training data for offline RL, enabling the creation of conversational agents that can optimize multi-step objectives and gather information effectively. The proposed approach shows improved performance in tasks such as teaching and preference elicitation compared to traditional methods that use LLMs directly.
+
+### Strengths
+1. The introduction of an "imagination engine" to synthesize hypothetical dialogues is a novel approach. It creatively leverages LLMs' ability to generate diverse and human-like conversations.
+
+2. This method adopts a multi-step optimization strategy to obtain better quality data.
+
+### Weaknesses
+1. Although the author mentioned efficiency considerations, it's somewhat difficult to justify using GPT-2 as the base model for experiments in this day and age. Why not try LLaMA or other more powerful open-source models?
+   
+2. The evaluation relies solely on human assessment, which is subjective. It would be better to incorporate objective evaluation metrics as a supplement. One possible approach could be to set aside around 10% of the dataset as a test set, run tests on it, and use metrics like BLEU and ROUGE to evaluate model performance. While this may not be the optimal solution, it’s better than nothing.
+
+### Questions
+1. I wonder why not introduce the criteria from the Critique Step during the Imagination Step? Wouldn't that make the process more streamlined?
+
+2. I'm curious about the size of the synthesized dataset. Was it entirely used for RL training?
+
+3. I would like to know the size of the test set used in the experiments. Additionally, I noticed that the evaluation was conducted by 12 different individuals. Is there any consistency check performed?
+
+4. The authors assume that "models trained with RL outperform those using prompts" and conducted experiments with GPT-3.5. I am interested in knowing the exact prompt used to call the model, as it significantly affects the outcome of prompting. Moreover, the authors might consider conducting experiments with more advanced models (such as GPT-4o). Relying solely on GPT-3.5 does not strongly support the assumption, as its performance lags behind and may even fall short of some of the cutting-edge open-source models.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper presents a new approach for training goal-directed dialogue agents by applying reinforcement learning (RL) to synthetic data generated from large language models (LLMs). While LLMs excel in general text generation, they often struggle with tasks requiring multi-turn, goal-oriented interactions. This study introduces an "Imagination Engine" (IE) that synthesizes realistic task-specific dialogues, which are then used to train RL-based agents capable of optimizing for outcomes in conversations. The approach is demonstrated on tasks like teaching concepts and eliciting user preferences, with experimental results indicating that the method outperforms direct prompting of LLMs in achieving conversational goals.
+
+### Strengths
+1. The method creatively leverages LLMs to generate diverse, goal-directed dialogues, addressing data scarcity in training agents for complex conversational tasks.
+2. The paper shows an efficient application of offline RL by using synthetic dialogues, enabling scalable agent training without the need for real-time user interactions.
+3. Empirical results, including user studies, suggest that the proposed method improves outcomes over conventional LLM-based approaches in teaching and preference elicitation tasks.
+
+### Weaknesses
+1. The authors use the term "goal-directed dialogue," but in NLP, the terms target-driven conversation and proactive dialogue are more widely used to describe similar tasks. These areas have established research and methods that could deepen the paper's connection to prior work.
+2. The idea of using LLM to simulate conversations and then leverage offline reinforcement learning to train a model is not new. The authors might want to compare with a rather similar work here: https://aclanthology.org/2024.acl-long.262/
+3. The evaluation is primarily in synthetic settings, limiting insights into how well the approach would perform in more dynamic, real-world user interactions with diverse needs.
+
+### Questions
+As detailed in weaknesses.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes a novel method to train goal-directed dialogue agents using zero-shot RL. The core idea is to leverage LLMs to simulate human-like conversations, creating a diverse dataset, which is then used with offline RL to optimize dialogue agents for multi-step, goal-directed interactions. Experiments show that using LLMs to generate data and then training RL agents outperforms directly using LLMs as dialogue agents.
+
+### Strengths
+The imagination engine creates a varied dialogue dataset without requiring extensive human-collected data.
+
+Combined with RL, the agents are trained to ask clarifying questions and make goal-directed decisions over multiple turns.
+
+User studies show RL-based agents excel in natural dialogue flow and effective information gathering compared to traditional LLM methods​
+
+### Weaknesses
+The synthetic dataset generated by IE is based on LLM simulations, which may not fully reflect actual user behavior. Particularly for highly personalized or complex tasks, synthetic dialogues can diverge significantly from reality, as simulated users may appear overly cooperative or lack the randomness typical of real users. This discrepancy can affect the agent's performance in real-world scenarios.
+
+Training with offline RL on a synthetic dataset can encounter the "distribution shift" problem, where the real-world dialogues that the agent encounters differ from the distribution of the training data. This mismatch may lead to poor performance when the agent faces novel scenarios. Although optimistic estimation techniques were applied to mitigate this, such methods cannot entirely eliminate the impact of distribution shifts.
+
+Current evaluations are based on annotations from 12 users, which is a limited sample size and could introduce bias. Using the number of turns can indicate effectiveness, while satisfaction could be evaluated through various system assessment methods in dialogue systems. Larger, more reliable evaluation results would be beneficial.
+
+While offline RL methods allow for policy optimization on fixed synthetic datasets, the absence of real-time feedback in dynamic and complex dialogue scenarios can lead to suboptimal strategies. For example, in real dialogues, user feedback or sentiment may change dynamically, and a fixed dataset cannot capture this variability fully, limiting the agent's adaptability and flexibility during actual interactions.
+
+Since synthetic data is generated by large language models, it may lack real-world noise and complexity, particularly in ambiguous or conflicting user input. This lack of realistic data could lead to "over-idealized" behavior, meaning the agent may perform well in "clear and cooperative" scenarios but struggle when confronted with the unpredictability of actual users.
+
+Some research on dialogue uncertainty also approaches the issue from an information-gathering perspective. The authors might consider comparing more advanced prompting methods with the current RL approach, as RL data collection and training costs are still relatively high.
+
+-- Uncertainty of Thoughts: Uncertainty-Aware Planning Enhances Information Seeking in Large Language Models. https://arxiv.org/abs/2402.03271
+
+-- MEDIQ: Question-Asking LLMs for Adaptive and Reliable Clinical Reasoning. https://arxiv.org/abs/2406.00922
+
+### Questions
+See the weaknesses
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper describes an approach for training  goal-directed dialog agents by leveraging synthetic data generated from LLM. The authors showed that agent trained on the LLM generated synthetic data has a higher performance than prompting LLM to act directly as an agent. TC also discussed the effectiveness of using behavior cloning vs. RL for training such agents.
+
+### Strengths
+1. This paper is well written and easy to follow. The author discussed two key hypotheses (the effectiveness of LLM trained on self generated synthetic data vs. direct prompting; and offline RL vs. behavior cloning), and used the same throughout the paper in the methodology and experiment sections which make it easy to comprehend and follow.
+2. The proposed method is discussed in good detail. The authors presented the imagination engine and the RL optimization with good clarity. The authors provided provide comprehensive discussion on related work and preliminaries on MDP and RL which helped the presentation of the proposed method.
+3. Comprehensive experiments against multiple baseline methods. The authors compared the proposed method to different baselines on multiple tasks to illustrate the effectiveness of the proposed method. The authors also provided detailed examples to showed the quality of the responses from different approaches.
+
+### Weaknesses
+1. The authors made some vague and strong claims in the paper that are not well supported. e.g. line 76 “In effect, the LLM can imagine what a human could do, but not to what an optimal agent should do”; line 250-253 “Since inferring the human’s persona is an important skill we want downstream learning agent to acquire”.
+2. The quality of the synthetic data produced by the “imagination engine”, which plays a key role in the optimization of the dialog agent through RL, is not sufficiently discussed. For example, the author sampled reward score, and used that as part of the input for the synthetic dialog generation. It’s unclear how closely the LLM followed the instruction in generating the dialogs. Without understanding the quality of the generated data, it’s hard assess the effectiveness of the optimization with RL.
+3. Training dialog agent using offline RL from dialog corpus is not something new. It has been widely explored in dialog research literatures. The main novelty of the work to me is on leveraging self-generated synthetic data for RL training. To strengthen the argument that this is an effective approaching comparing to prompting LLMs directly, I would expect the authors to discuss more on the intuition of this approach and the corresponding validation, in addition to the experiment results on response quality.
+
+### Questions
+1. What's the quality of the synthetic data?
+2. What's the intuition that training the dialog agent on self generated data works better than prompt the llm directly?
+3. Line 277: r = r_i only if s' = \tau_i is the full dialog - what's the assigned value of r when it is not the end of the dialog?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4

human_reviews/7BQkXXM8Fy.md

@@ -0,0 +1,161 @@
+## Human Reviewer 1
+
+### Summary
+In this paper, the authors investigated the design choices for diffusion model-based offline RL methods. 
+The design choices mainly focused on planning strategy, network architecture, guided sampling, and action generation (whether to generate both state and action directly, or generate only the state and estimate the action separately using an inverse dynamics model). 
+The tasks used in the study were Maze2D, AntMaze, and Franka kitchen (MuJoCo locomotion also used in section 4.6).
+
+### Strengths
+This paper investigates the design choices for diffusion model-based offline RL methods and identifies effective components. Although various (algorithm/implmentation) designs have been proposed in previous research on diffusion model-based offline RL methods, their effectiveness in a unified framework has not been sufficiently explored. 
+In general, the performance of reinforcement learning methods largely depends on design choices. 
+Therefore, this paper, which provides insights into effective design choices, has value on the engineering front.
+
+### Weaknesses
+The number of tasks used to investigate the design choices is limited. This paper focuses on Maze2D (2 tasks), AntMaze (3 tasks), and Franka kitchen (4 tasks) (with MuJoCo locomotion tasks also included in section 4.6). However, for a paper investigating design choices, this is fewer than the number of tasks typically covered in papers accepted at ICLR (or conferences of a similar level). For instance, the paper [1] that investigated the implementation design of Offline + Online RL used 30 tasks in its study.
+
+Moreover, the paper does not verify the insights/findings on a different set of tasks (i.e., validation tasks) from those used in the design choice investigation. This leaves uncertainty about how generalizable the insights are (or whether they are simply overfitted to the tasks examined).
+
+
+[1] Ball, Philip J., et al. "Efficient online reinforcement learning with offline data." International Conference on Machine Learning. PMLR, 2023.
+
+
+Minor comments:  
+
+Line 018: 
+> We trained and evaluated over 6,000 diffusion models  
+
+I didn’t quite understand the breakdown of these 6,000 diffusion models. Were most of these models the ones trained and evaluated through the grid search mentioned in the step (1) in Section 3.2?  
+
+Line 174: 
+> (1) Conduct a comprehensive search on the key components (Sect. 3.1) by combining grid search and manual tuning to obtain the best results.  
+
+What exactly does "manual tuning" refer to in this context?
+
+Figure 5.  
+It seems that the Transformer score for Kitchen-M doesn’t have a confidence interval. 
+Also, I wasn’t clear on what the confidence intervals in the other parts of this figure represent (are they calculated based on 500 episode seeds?). 
+
+Typoes:  
+line 101: Zhang et al., 2022) In  -> Zhang et al., 2022). In  
+line 158:  Chen et al., 2024)) -> Chen et al., 2024).  
+line 479: planning(Sect. 4.6) -> planning (Sect. 4.6).
+
+### Questions
+Please refer to my previous comment on the weaknesses.
+If either (1) validation results from tasks other than those used to investigate the design choices, or (2) validation results from 20-30 tasks were provided to support the insights on design choices, I would be inclined to recommend an Accept (assuming other reviewers do not point out any major weaknesses that I may have overlooked).
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper explores the design choices in diffusion model planning within offline reinforcement learning (RL). Through experiments on over 6,000 models, the paper systematically investigates key components of diffusion planning, including sampling algorithms, network architectures, action generation methods, and planning strategies. The study finds that some design choices, such as unconditional sampling outperforming guided sampling and Transformer outperforming U-Net, lead to better performance. Based on these insights, the paper proposes a simple yet strong baseline model called Diffusion Veteran (DV), which achieves state-of-the-art results on standard offline RL benchmarks.
+
+### Strengths
+1.Comprehensive empirical study: The paper conducts a large-scale experimental study, using controlled variable methods to analyze the impact of each component on model performance, providing rich data support.
+2.Innovative insights: The study reveals design choices that contrast with common practices in diffusion planning, such as the advantages of unconditional sampling and the use of Transformer, offering new directions for future research.
+3.Simple yet effective baseline model: The proposed DV model is simple but performs exceptionally well, demonstrating high generalizability and effectiveness, laying a solid foundation for further research.
+4.Wide applicability: The DV model performs well in multiple tasks such as maze navigation and robot manipulation, demonstrating its adaptability and broad applicability.
+
+### Weaknesses
+1.Limited exploration of long-term dependencies: While the paper discusses the importance of handling long-term dependencies using Transformer, it does not delve deeply into how this is manifested across different tasks. The related discussion could be more robust.
+2.Potential typo in Equation 2.1: There seems to be a typo on the right-hand side of Equation 2.1, where S(t−1) appears, which might be incorrect.
+
+### Questions
+1.You mention that unconditional sampling outperforms guided sampling, which contrasts with results in typical image generation tasks. Could you elaborate on the underlying reasons behind this phenomenon?
+2.The paper primarily focuses on state-based tasks. Are there plans to extend the study to vision-based or goal-conditioned reinforcement learning tasks?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper presents an extensive experimental study aimed at understanding the factors that contribute to an effective diffusion planner for decision-making in offline reinforcement learning. The authors provide valuable insights into the role of various components within diffusion models. Building on these insights, they propose a straightforward yet robust diffusion planning approach that delivers state-of-the-art (SOTA) performance in standard offline RL benchmarks.
+
+### Strengths
+1. This paper is well-organized and easy to follow.
+2. The empirical analysis is comprehensive, providing solid support for the conclusions. 
+3. Each conclusion is accompanied by decent explanations
+
+### Weaknesses
+While the paper provides strong evidence for the effectiveness of the proposed methods on the D4RL dataset, it is unclear how generalizable these findings are to other types of decision-making problems or datasets. More diverse datasets could strengthen the claims.
+
+### Questions
+No
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+4
+
+### Rating
+10
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper analyses key components (guided sampling algorithms, network architectures, action generation methods, and planning strategies) critical to decision-making in diffusion planning.  The paper gives practical tips about the choices
+and provides insights into the strengths and limitations of diffusion planning. The experiments in the paper are very comprehensive.
+
+### Strengths
+The experiments in the paper are very comprehensive.
+
+### Weaknesses
+Although the experiments in the paper are rich, readers still want to see how the original innovation in theory can better apply diffusion models to decision-making tasks
+
+### Questions
+No
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+2

human_reviews/7UTsVPcHZa.md

@@ -0,0 +1,199 @@
+## Human Reviewer 1
+
+### Summary
+In this paper, a new cross-channel activation function, namely SPA, is proposed. Most of previous activation functions used in deep learning, such as ReLU, only independently consider the multi-channel features, which may ignore some cross channel information. This paper considers to interpret activation functions as an optimization problem, then proposes SPA to maintain the feature relationship between multiple channels. Specifically, each cross channel feature x should be projected to a convex set S (which is defined by introducing a constant \delta), and the projected feature can be viewed as the output of SPA. Moreover, this paper provides the solution of the SPA optimization problem, and show the update rule of each x. The experimental results imply that SPA show good perfermance on variety of databases.
+
+### Strengths
+1. The consideration that transfer the activation functions into an optimization problem supports the SPA method.
+2. The SPA activation function is well-defined, and it is possible to be implemented easily.
+3. The relationship between the constant \delta and classification perfermance is carefully analyzed, and the authors provide a way to find a suitable \delta.
+4. Experimental results show that SPA show slightly better accuracy than traditional activation functions on multiple models and databases.
+
+### Weaknesses
+1. The experimental results only include the small-scale databases. The authors mentioned that the imagenet-1k results are included in Appendix E, but I cannot find them there. Moreover, I believe that Imagenet-1k results are important for this paper, which should be included in the main paper instead of appendix. 
+
+2. It is better to consider the time-cost of SPA. Is it similar with traditional activation functions?
+
+### Questions
+1. Please provide experimental results on Imagenet-1k database
+
+2. Is SPA has similar time-complexity with traditional activation functions?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+In this paper, the author proposes a cross-channel activation function. The core concept of this activation function embraces the cross-channel relationship, which is purported to capture the patterns and semantics of the input data for activation. Additionally, the author introduces a threshold V* for activation, which is utilized to eliminate unimportant features with varying control ratios.
+The author applies the aforementioned technique to the cross-channel activation function, which is validated on several toy datasets. The enhancement is somewhat restricted.
+
+### Strengths
+1) The author presents a novel activation function.
+2) The proposed methodology is validated across six datasets.
+
+### Weaknesses
+1) The concept conveyed in the paper lacks significance. The author introduced two methodologies: cross-channel relation for activation and threshold v* to filter out irrelevant features.
+
+1.1) The author posited that "These functions often process inputs separately, neglecting dependence between them, such as the spatial or cross-channel relation of the features. Spatial relation refers to the local connectivity and neighborhood structure of the features, while cross-channel relation refers to the correlation and diversity of the features across different channels. "
+In my opinion, the convolutional operation already calculates the cross-channel relation of the features. Therefore, introducing another cross-channel relation for activation function seems superfluous.
+
+1.2) Regarding the threshold v*, feature normalization and bias serve a similar purpose. Consequently, the significance of the threshold appears diminished.
+
+2) The proposed methodology has only been validated on toy datasets and tiny ImageNet. Larger-scale datasets are imperative. In my view, if the model is trained with an adequate number of dataset samples, the original cross-channel relation learned through convolution operation and the threshold will be well assimilated by the model.
+
+3) The in-depth analysis explaining why deep models necessitate additional cross-channel relation and threshold parameters is absent.
+
+4) The literature review is lacking. Several crucial and highly relevant works are absent.
+
+References:
+[1] Dynamic Neural Response Tuning, ICLR 2024.
+[2] Exploring optimal adaptive activation functions for various task，IEEE BIBM 2020.
+[3] Exploring Optimal Adaptive Activation Functions for Various Tasks, 2020.
+[4] Deep sparse rectifier neural networks.  JMLR 2011
+[5] Density Modeling of Images using a Generalized Normalization Transformation, CoRR 2015.
+[6]  ...
+
+### Questions
+What distinguishes the cross-channel information acquired through the proposed activation from that obtained through the original convolutional operation?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces Simplex Projection Activation (SPA), a novel activation function for CNNs that addresses the limitations of conventional activation functions like ReLU by considering cross-channel dependencies. SPA projects input tuples across channels onto a convex set, preserving feature relations and avoiding information loss. The authors also explore the learnable parameter δ, which controls the pass-through ratio and significantly influences the model's performance. Through extensive experiments, the authors demonstrate SPA's effectiveness, showing it outperforms ReLU and other activation functions in various datasets and noise conditions.
+
+### Strengths
+### Originality
+
+The paper presents the Simplex Projection Activation (SPA) function, which introduces a novel approach to activation functions in convolutional neural networks (CNNs). The originality of the paper lies in its consideration of cross-channel dependencies, which traditional activation functions like ReLU ignore. By projecting input tuples across channels onto a convex set, SPA maintains feature relations and avoids information loss, offering a creative solution to a known limitation in neural network design.
+
+### Quality
+
+The quality of the paper is reflected in its thorough experimental evaluation. The authors have conducted extensive experiments across various datasets and noise conditions, demonstrating SPA's superiority over ReLU and other common activation functions. The statistical tests used to compare the accuracy of different activation functions are appropriate, and the results are consistently presented, indicating a high level of quality in the research methodology.
+
+### Clarity
+
+The paper is generally well-structured and clear in its presentation. The problem statement is clearly defined, the motivation for the SPA function is well-articulated, and a relatively complete derivation process of the simplex method is provided. The use of illustrations and charts to help visualize concepts and results further enhances the clarity of the paper.
+
+### Significance
+
+The significance of the paper is evident in its potential impact on the field of deep learning. SPA's ability to improve model performance and robustness to noise is a valuable contribution, especially given the widespread application of CNNs across various domains. The paper's findings could lead to improvements in the design of neural networks and potentially extend to other types of neural network architectures, highlighting the broader implications of the research.
+
+In conclusion, the paper is strong in its original approach to addressing a known issue in CNNs, the quality of its experimental validation, the clarity of its presentation, and the significance of its potential impact on the field of deep learning. The research presented in this paper could influence future work in activation function design and neural network optimization.
+
+### Weaknesses
+### Mistake in Mathematical Expression
+
+One of the specific weaknesses in the paper is the definition of the set $S$ used in the SPA function. The paper states $S = \{x = [x_1, x_2, \ldots, x_C] \mid |x_1| + |x_2| + \cdots + |x_C| \leq \delta\}$ without explicitly requiring $x \geq 0$. This omission may compromise the theoretical foundation, as the non-negativity constraint is crucial for the simplex projection and the activation function's behavior. The authors should clarify this condition to avoid any misinterpretation.
+
+### Misleading Illustration
+
+The three-dimensional illustration in Figure 1(b) appears to be hand-drawn, with the projection directions of various points and the axes appearing inconsistent and chaotic, which may mislead readers. High-quality and accurate visual representation is crucial for conveying mathematical concepts, and the quality of this figure does not meet this standard. The authors should consider revising this figure using mathematical 3D space plotting tools such as GeoGebra to ensure it accurately represents the SPA projection without misleading readers.
+
+### Theoretical Implications
+
+While the paper provides a thorough experimental evaluation, it could benefit from a deeper theoretical analysis of the SPA function. Specifically, the paper could explore the theoretical implications of the SPA function on network convergence and generalization. A more in-depth theoretical discussion would strengthen the paper's contribution and provide a stronger foundation for the experimental results.
+
+### Generalization to Other Network Architectures
+
+The paper focuses on the application of SPA to CNNs, but does not extensively explore its potential application to other types of neural network architectures, such as transformer models. Expanding the scope of the paper to include experiments or a discussion on the applicability of SPA to these architectures would enhance its significance and impact.
+
+### Discussion on Limitations
+
+The paper could benefit from a more explicit discussion on the limitations of the SPA function. For example, the authors could discuss potential challenges in optimizing the δ parameter for deep networks or the computational overhead introduced by the SPA function. Acknowledging and addressing these limitations would provide a more balanced view of the SPA function's practical applicability.
+
+In summary, the paper's weaknesses can be addressed by clarifying mathematical definitions, improving visual representations, expanding the theoretical analysis, exploring the applicability to other network architectures, and discussing the limitations of the proposed method. By addressing these points, the paper could provide a more comprehensive and robust contribution to the field of neural network activation functions.
+
+### Questions
+1. **Clarification on Set $S$ Definition:** In the definition of the set $S$, it was noted that the non-negativity constraint $x \geq 0$ was not explicitly stated. Could the authors please clarify whether this constraint is intended to be part of the definition of $S$?
+2. **Applicability to Other Network Architectures:** Given the novelty of the SPA function, it would be insightful to understand its potential application beyond CNNs. Are there any empirical results or theoretical predictions regarding SPA's applicability in other neural network architectures such as RNNs or transformer models?
+3. **Theoretical Analysis of SPA:** The paper could benefit from a deeper theoretical analysis of the SPA function, particularly regarding network convergence and generalization. Are there any theoretical insights or ongoing work that the authors could share regarding these aspects?
+
+### Soundness
+4
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes a new activation function for convolutional neural networks (CNNs) called Simplex Projection Activation (SPA).  Unlike activation functions like ReLU treating elements independently, SPA considers the relationships across multiple channels. Designed as a projection to simplex regularized on $l_1$ norms, the proposed method introduces a rather flexible threshold regarding the norm of different channels.
+
+### Strengths
+* This paper proposes a new simplex projection perspective for a new activation function, which is new and interesting.
+
+* The paper conducts comprehensive experiments on various model architectures and datasets. The paper also tests the robustness of the proposed method against noise.
+
+### Weaknesses
+* The paper claims that ReLU suffers information loss by eliminating negative features. However, the proposed method also eliminates features. Moreover, the ReLU masks element-wise features while the proposed method masks channel-wise features. It seems the proposed method would suffer more information loss. Therefore, I have doubts about the analysis regarding the shortcomings of ReLU and how the proposed method improve it.
+
+* According to the experimental results, the improvement brought by SPA is marginal. An average accuracy and its standard deviation over multiple runs should be provided. 
+
+* Besides, suppose SPA does improve upon ReLU and GELU. The improvement brought by SPA seems not substantial compared to the increase in computational cost for a much more complicated activation function. I would appreciate a more comprehensive time complexity analysis of the proposed SPA.
+
+### Questions
+I list some of my concerns in the Weaknesses section. Following are my questions and further concerns.
+
+* Regarding the first weakness I mentioned above, activation functions such as Leaky-ReLU do not eliminate negative features.  Is there a comparison between the proposed method and Leaky-ReLU?
+
+* This paper analyzes the pass-through ratio between the proposed SPA and ReLU. Honestly, I can't see a clear pattern indicating SPA is superior. Why would SPA outperform ReLU-like activation functions? In my understanding, the pass-through ratio of the ReLU-like activation function is controlled by the bias term in the convolutional layer or the linear layer, which is learned automatically during the training procedure with gradient descent. The proposed SPA actually adds a manually determined soft threshold $\delta$ with very complicated computation to determine the set $\mathcal{I}$ of unmasked channels. Is there any theoretical analysis to prove that SPA would outperform ReLU?
+
+* In lines 206-209, the authors claim that masking channel-wise feature (as in SPA) is better than masking element-wise feature (as in ReLU). Is there any theoretical or empirical evidence to support the claim? Would a more coarse-grained activation function lead to more information loss? In SPA, the output of some convolutional kernels is completely masked and set to zero.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/7ZToWPWUlO.md

@@ -0,0 +1,179 @@
+## Human Reviewer 1
+
+### Summary
+The paper describes a Reinforcement Learning strategy to solve an approximate minimum normalized cut on spider web-like planar graphs, like city street maps.  The idea is that the problem can be approximated by a circles-wedges clusterings, in which inner nodes (w.r.t. some central point o) can be grouped w.r.t. their distance from a center o, while outer nodes are further subdivided w.r.t. their angular polar coordinates. The actions to be performed will then be the radius of the outer circle and the (discrete) points where to split the outer nodes.
+Some training strategies are defined to help the problem converge and refine the grouping. 
+The method is tested on synthetic spider web-like graphs, and subgraphs extracted from a city map.
+
+### Strengths
+I like the idea of modeling the grouping of nodes according to the previous knowledge about the domain. This allows to simplify the minimum graph cut problem for the specific type of graphs considered, and obtain better results than other general-purpose grouping algorithms.
+
+### Weaknesses
+My main concern is about the quite demanding assumption of the algorithm, which is designed to work on spider-like planar graphs, where nodes are embedded (have coordinates) in R^2. In particular, my comments are:
+- even if the proposed solution is sound for the specific problem, I’m not sure it is general enough to be of broad interest to this community. It looks more suited for a venue in the specific application.
+- it is not explained why the grouping in inner circles and outer wedges is a good modeling. Is it a pattern observable in other city map grouping algorithms? Does this pattern apply to all cities?
+
+There is a drop in writing quality in section 5, which raised some doubts:
+- It is incorrect to say that transformers work only on sequences, they work on any set of points but often benefit from a positional encoding. 
+- Sections 5.2.1 and 5.2.2 are quite intricate and could be simplified. In practice, they define two different positional encodings for ings and wedges, where points for the ring are encoded with their distance from the center, and for wedges, they are projected into the unit circle (and possibly equispaced?).
+- The optimal partitioning of circles (row 322) should be better introduced.
+- In 5.4 I don’t understand what the “Current Partition” is. It is represented by a binary mask? How is it converted into the colored square matrix in Figure 4?
+
+To broaden the impact of the work, it would be worth trying to apply the proposed method to different families of graphs and different datasets. Also, graph cut methods seem to exist specifically designed for planar graphs (e.g. “Efficient Planar Graph Cuts with Applications in Computer Vision”) that would be worth considering in the comparison.
+
+### Questions
+- Your setting is much simpler than finding normalized min cut in general undirected graphs. Is it still a NP-Hard problem? For instance, polynomial algorithms for the min-cut on planar graphs exist (I just found a few, but I might be missing some fundamental details). Would they also apply to your definition of normalized cut?
+- From reading the text, it sounds like you are providing a novel definition of normalized cut, but it looks like the standard definition to me. Am I missing something? My confusion is further increased by the statement at line 214: “Despite being a simpler class of graphs, these bounds give a theoretical justification of the normalized cut definition equation 2 and the ring-wedge shaped partition.”
+- At row 283 you write “Note that this transformation does not change the order of the nodes or the partitions.” What do you mean by node order?
+- How is the center of the graph defined?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This work tackles a special case of a normalized-cut problem: that of spider-web shaped weighted planar graphs.
+The graph is partitioned into rings, and the outer ring is partitioned into wedges. The approach transforms the graph by:
+a. projecting ring nodes onto an axis according to their distance from a center while maintaining node order
+or by
+b. projecting nodes onto a unit circle.
+The transformation results in the partitioned nodes forming a sequence, which is encoded by a transformer.
+Reinforcement learning is used to find the ring radius and number of outer ring wedges that result in a minimal normalized cut. 
+
+Specifically, PPO is used, where the state, action, and rewards are encoded as: 
+a. State is the graph, number of rings and wedges of the outer ring.
+b. Actions are the ring radius or wedge angle.
+c. Rewards are 0 in all steps, and the negative normalized cut at the end.
+The wedge partition is trained using random ring partitions, followed by training of both ring and wedge partitions.
+The ring partition is first inferred during testing, followed by the wedge partition.
+
+This work demonstrates that this transformation is suitable for a specific case of road networks.
+The transformation is applied as a preprocessing step, finding a minimal normalized cut with a lower value than other baselines.
+
+The approach is evaluated using synthetic and real-world data.
+a. 400k spider-web shape synthetic graphs with a 50 or 100 nodes, ring and wedge partitions, with unweighted and random edge weights.
+b. Connected sub-graphs randomly extracted from real-world city maps with edge weight corresponding to traffic.
+
+The performance of the approach is compared with a baseline partitioning method, METIS, and with spectral clustering.
+The ring and wedge partitions are compared with brute force and random partitions.
+
+### Strengths
+1. The graph transformation is applied as a pre-processing step, aiming to utilize the specific graph structure.
+
+2. The results are a minimal normalized cut with a lower value than other trivial baselines.
+
+### Weaknesses
+1. The decisions to apply the transformations to the graph are manual.
+The method and its implementation details are ad-hoc and very specific.
+
+2. Dynamic programming is used to compute the optimal partition given the maximum radius and ring count. Ablation studies of this algorithm and the reinforcement learning approach are missing.
+
+3. The graphs are relatively small consisting of 50, 100 (for training), or 200 nodes (in testing).
+
+### Questions
+Can this approach be automated by classifying the graphs to automatically find which transformations should be applied as a preprocessing step?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes the Wedge Ring Transformer (WRT), an RL-based approach to minimize the Normalized Cut (NC) on planar weighted graphs. WRT leverages polar coordinates and employs a multi-head transformer with a Proximal Policy Optimization (PPO) objective to address the NC problem. The approach utilizes a two-stage training process to effectively learn both ring and wedge partitioning strategies. Experimental results indicate that WRT effectively reduces the NC.
+
+### Strengths
+The paper provides a clear definition of the Normalized Cut (NC) problem and the description of the Wedge Ring Transformer (WRT) is well-articulated.
+The design of transformations specifically tailored for ring and wedge shapes appears effective.
+Provide some theoretical analysis about cheeger bound for ring and wedge partition.
+
+### Weaknesses
+Ablation studies: The ablation studies primarily focus on the two-stage training process, but lack analysis on key components of the paper's main contribution, such as the wedge-ring transformer, PAMHA, and pre-calculation. Ablation studies on these components would provide a more comprehensive evaluation of the WRT architecture.
+Running Times:  The paper does not provide an analysis of the model's runtime, leaving the computational efficiency of WRT unaddressed.
+
+### Questions
+1.	The paper argues that GNNs were not used due to scalability issues. However, the proposed method also seems to require processing the entire graph at once, and experiments were conducted on data with a maximum of only 200 nodes. It remains unclear how WRT scales to larger graphs, and additional evidence of scalability would strengthen the paper's claims.
+2.	What are the evaluation metrics in Table 1 and Table 2?
+3.	Although WRT is designed for ring and wedge-shaped partitions, I am interested in understanding its performance on other types of datasets. For example, how does it perform on datasets that primarily feature extended ring shapes, such as the long-tail structures often found in knowledge graphs?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The manuscript presents Wedge and Ring Transformers (WRT), an RL-based method for solving the Normalized Cut (NC) problem in weighted graphs with shape-specific constraints. By transforming graphs into polar coordinates and using Transformers with Proximal Policy Optimization, WRT effectively handles both ring and wedge partition shapes, optimizing NC while adhering to these constraints.
+
+### Strengths
+1. The paper addresses the Normalized Cut problem in the context of real-world applications, such as road network simulations, where partition shape constraints are critical.
+2. The introduction of the Wedge-Ring Transformer, tailored to handle specific shape constraints in graph partitioning, is innovative.
+
+### Weaknesses
+1. The paper includes a limited set of baseline methods for comparison. Adding more baselines, particularly those used in NeuroCUT, would strengthen the evaluation by providing a more comprehensive assessment of WRT's performance.
+2. The baselines lack specialized adaptations for the "Ringness" and "Wedgeness" constraints, while WRT is explicitly designed with these constraints in mind. This discrepancy may lead to an unfair comparison, as the baselines are not optimized to meet these specific structural requirements.
+3. The experiments use relatively small graph instances, whereas NeuroCUT and other methods operate on benchmarks with thousands of nodes, aligning more closely with real-world scales. The current experimental scale may limit the ability to assess WRT’s applicability to large-scale, practical scenarios.
+4. Given the use of Transformers, I am concerned about the performance and computational cost of training and inference on large-scale datasets.
+
+### Questions
+1. Were NeuroCUT and ClusterNet evaluated by training on the same datasets as WRT? Ensuring consistent training conditions is crucial for fair comparison.
+2. The Cheeger Bound presented appears to be a specific case of a more general result. How does this theoretical finding contribute to model design or provide insights for experimental evaluation?
+3. What specific metrics are used in Tables 1 and 2? 
+4. Why are there no generalization results for NeuroCUT and ClusterNet in Table 2? 
+5. According to Fig. 6, it appears that RL training converges early (10–20k of 400k steps). Does the extended training beyond this point contribute to any performance improvements, or could training resources be optimized?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/92vMaHotTM.md

@@ -0,0 +1,170 @@
+## Human Reviewer 1
+
+### Summary
+The paper presents EdgePrompt, a method that enhances pre-trained GNNs for downstream tasks by using learnable prompt vectors on edges. EdgePrompt+ further customizes these vectors for individual edges. This approach improves graph structural representation and is compatible with various GNN architectures. Experiments on multiple datasets show its effectiveness over existing methods for node and graph classification tasks.
+
+### Strengths
+1. The paper is well-organized, with clear points, and is easy to follow.
+2. The effectiveness of EdgePrompt is theoretically guaranteed, and it performs excellently in downstream tasks.
+
+### Weaknesses
+1. The motivation for constructing EdgePrompt is insufficient. Why is it necessary to design EdgePrompt under graph prompt tuning? What core problem does EdgePrompt address compared to existing graph prompt tuning methods? What are its advantages?
+2. Compared to ALL-in-one and GPF, EdgePrompt and EdgePrompt+ set different prompt vectors $p^{(l)}$ for each layer. What are the benefits of this design? Both All-in-one and GPF only add prompt vectors in the first layer to reduce dependency on the specific structure of the model. EdgePrompt lacks such advantages, and the paper does not explore the reasoning behind this design. Furthermore, the experimental section does not include relevant comparisons to demonstrate the necessity of setting different prompt vectors for each layer.
+3. The datasets included in the experimental section do not contain initial edge features, which raises doubts about the effectiveness of EdgePrompt on graphs that inherently have edge features. If the original graph already contains edge features, how should EdgePrompt be integrated with these edge features? What would its performance be like in that case?
+4. The downstream tasks involved in the experiments are limited to node classification and graph classification, with other graph tasks such as link prediction and node regression not being included.
+
+### Questions
+Please refer to the points I mentioned in the weakness part.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper proposes EdgePrompt, a graph prompt tuning method that enhances GNNs by learning prompt vectors for edges, improving graph representations. EdgePrompt integrates these edge prompts through message passing, outperforming existing methods across ten datasets under four pre-training strategies.
+
+### Strengths
+1. The paper is well-motivated. It's important to integrate structural knowledge in prompt learning.
+2. The authors conducted extensive experiments, demonstrating the effectiveness of the proposed methods.
+3. The authors provide theoretical analysis, further proving the effectiveness of the proposed methods.
+4. The paper is well written and easy to follow.
+
+### Weaknesses
+1. **Inaccurate statement**: GraphPrompt [1] is not based on a specific pre-training strategy. As shown in GraphPrompt+ [2], all contrastive learning pre-training methods can be unified as subgraph similarity calculations. The link prediction used in [1] can be replaced by other methods.
+2. **Missing related work**: GraphPrompt+ [1] also adds prompt vectors to each layer of the pre-trained graph encoder, which should be discussed and compared.
+3. **Unclear explanation of anchor prompts in EdgePrompt+**: It is unclear what the anchor prompts in EdgePrompt+ represent. In my opinion, anchor prompts are introduced to address the overfitting problem caused by directly learning edge-specific prompts for different edges, but there lacks a explanation for the meaning of the anchor prompts. A more reasonable and effective solution could be conditional prompting [3,4], which I highly recommend the authors explore in future work.
+
+
+[1] Liu et al. "Graphprompt: Unifying pre-training and downstream tasks for graph neural networks." Proceedings of the ACM Web Conference 2023. 2023.\
+[2] Yu et al. "Generalized graph prompt: Toward a unification of pre-training and downstream tasks on graphs." IEEE Transactions on Knowledge and Data Engineering (2024).\
+[3] Zhou et al. "Conditional prompt learning for vision-language models." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2022.\
+[4] Yu  et al. "Non-Homophilic Graph Pre-Training and Prompt Learning." arXiv preprint arXiv:2408.12594 (2024).
+
+### Questions
+See weaknesses.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper introduces EdgePrompt, a new graph prompt tuning method that improves graph representation for downstream tasks by learning edge-specific prompts, enhancing the performance of pre-trained GNNs. Extensive experiments show EdgePrompt’s effectiveness across various datasets and pre-training strategies, outperforming several baseline methods.
+
+### Strengths
+1. EdgePrompt improves the adaptation of pre-trained GNN models for downstream tasks by introducing edge-level prompts, which helps bridge the objective gap between pre-training and downstream tasks..
+2. Extensive experiments on multiple datasets and pre-training strategies demonstrate the method’s effectiveness, showing better performance compared to existing graph prompt tuning approaches.
+
+### Weaknesses
+1. EdgePrompt uses shared prompt vectors, which may not capture the different relationships between edges well. This can limit the model’s ability to use all the information in the graph.
+2. EdgePrompt+ adds multiple anchor prompts and score calculations, which can make the model more complex. This can lead to higher computational costs, making it harder to use in larger graphs.
+3. The method struggles with few-shot learning because most edges lack supervision. This can reduce the model’s performance in real-world tasks where labeled data is limited.
+
+### Questions
+How can the performance of EdgePrompt be improved in scenarios with limited labeled data to enhance its effectiveness in node classification tasks?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+Recent graph prompt tuning methods have proven effective in adapting pre-trained GNNs to downstream tasks. However, they often overlook the crucial role of edges in graph prompt design. To address this research gap, this submission introduces a new graph prompt tuning method focused on edges, called EdgePrompt. Nevertheless, despite emphasizing the importance of edges in graphs, the authors make an overly strong assumption by considering only a single type of edge. Additionally, the paper does not address edge-related tasks, which significantly undermines the overall contribution and impact of the work.
+
+### Strengths
+S1. Clear motivation and presentation.
+
+S2. The proposed method can be integrated with existing pre-trained GNNs.
+
+### Weaknesses
+**Weakness**
+
+W1. The unclear statements regarding the edge-level aspect weaken the paper’s contributions.
+
+W2. The authors need to further elaborate on the technical contributions.
+
+W3. More experiments are needed to better support the superiority of the proposed method.
+
+**Concerns**
+
+C1. As a study focused on edge-level prompt tuning, the assumption that there is only one type of edge could significantly undermine the contributions and claims of this paper. In line 154, the modeling of the adjacency matrix, $\mathbf{A} \in \{0,1\}^{N \times N}$, implies that the paper does not target multi-relational graphs. However, compared to other node-level graph prompting systems, the proposed edge-level graph prompting method could be more suitable for graphs with multiple edge types. The authors may need to clarify this in the submission.
+
+C2. Since this work emphasizes edge-level prompt tuning, it would be beneficial for the authors to explore edge-related tasks, such as edge classification and link prediction, to further expand the scope of the paper.
+
+C2-1. In many real-world scenarios, studying edge-level tasks is highly relevant because the space of edge types can evolve over time. For example, in a social network, a newly introduced user interaction feature might require predicting new edge types using a trained GNN.
+
+C2-2. If the research on edge-level tasks is beyond the scope of current pre-trained GNNs (i.e., no existing pre-trained GNNs focus on edge-level tasks), the authors should clarify this limitation in the submission.
+
+C3. The core Equation (4) in EdgePrompt+ appears overly similar to existing work, which may diminish the paper’s technical contribution. In CompGCN [1], the operation of weighting relation embeddings based on relation base embeddings has already been shown to be simple and parameter-efficient. Therefore, the authors should elaborate on the unique technical contributions of their method.
+
+Minor Concerns:
+
+C4. More classic and promising pre-trained GNNs, such as Infomax, EdgePred, AttrMasking, MGSSL, GraphMAE, and Mole-BERT, could be included in the experimental section. At the very least, the authors should discuss these models and explain why they are excluded from comparison.
+
+C5. Figure 2 presents convergence speeds in terms of the number of epochs. The authors should also analyze the efficiency of the proposed method using learning curves or running time comparisons.
+
+
+**Reference**
+
+[1] COMPOSITION-BASED MULTI-RELATIONAL GRAPH CONVOLUTIONAL NETWORKS, ICLR 2020.
+
+### Questions
+Please focus on answering concerns C1-C3.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3

human_reviews/94LyPGDi0Y.md

@@ -0,0 +1,161 @@
+## Human Reviewer 1
+
+### Summary
+The paper introduces a pipeline to create a comprehensive dataset for fine-tuning the proposed MLLM, CHOPINLLM for chart understanding. It highlights that incorporating raw data values during pre-training, substituting images with textual data in fine-tuning, and prioritizing data extraction before answering questions significantly improve performance. Additionally, a benchmark dataset is developed to evaluate MLLMs’ comprehension of various chart types across different complexity levels.
+
+### Strengths
+1. The paper introduces efficient training techniques that significantly enhance chart comprehension.
+2. CHOPINLLM, a model for chart understanding, demonstrates strong performance with various chart types.
+3. A benchmark is established to evaluate MLLMs' comprehension of different chart types, aiding future research.
+4. The data generation pipeline uses text-only Large Language Models to efficiently create diverse datasets, reducing costs and complexity.
+
+### Weaknesses
+1. CHOPINLLM did not achieve state-of-the-art (SOTA) performance in Table 4. While the authors claim that higher-performing models benefited from using more data and annotated datasets, there is no evidence showing that the proposed synthetic data offers performance gains when combined with existing datasets. Demonstrating that such a combination improves results would strengthen the contribution of the synthetic data. Otherwise, the benefit of using only synthetic data to build an underperforming model appears limited. (this is my major concern)
+2. The paper lacks comparisons with a broader range of SOTA MLLMs that are not specifically tailored for chart understanding, such as InternVL2 and Phi-3.5-V.
+3. It omits comparisons with proprietary SOTA models like GPT-4o and Claude-3.5-Sonnet, which would help illustrate performance differences between open-source and proprietary models.
+
+### Questions
+In addition to the weaknesses:
+
+1. What is the difference between annotated data and synthetic data that could be the major cause of the performance gap between CHOPINLLM and ChartAst-13B? What challenges exist in create synthetic data in comparable quality? 
+2. Can the data generation method be generalized to other domains where annotated data is harder to obtain? Demonstrating this would help justify the advantage of using only synthetic data for training and emphasize its broader applicability.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper investigates the design space of chart understanding pretraining of multimodal LLMs along with a fully automatic synthetic data generation pipeline to resemble real-world charts. The resulting model, ChopinLLM, when pretrained on a mixture of LLaVA pretraining data and the synthetic data and fine-tuned on a mixture of LLaVA QAs and the synthetic QAs, achieves competitive performance on its own chart understanding benchmark and decent performance on a variety of other chart understanding benchmarks. The authors well documented the data generation pipeline and mappings between data usage at different stages and model performance.
+
+### Strengths
+- Investigating ways to improve chart understanding of MLLMs from the “pretraining” (e.g., aligning the connector with captioning data) perspective is rarely explored, which sets this work apart from others that focus on chart understanding in supervised finetuning of the full model on chart QAs. Experiments demonstrate that having a curated chart understanding dataset for pretraining can significantly enhance the model’s performance when later supervised finetuned on the same set of visual QA dataset.
+- The paper is clearly written and examples of data and the data curation process are well documented in the supplementary materials.
+- The experiments on the effectiveness of different types of chart understanding data are well investigated, where the major contribution factor toward the performance boost is to learn to translate the entire chart into textual data sources and learn to use the pattern for inference.
+
+### Weaknesses
+- A main argument from the paper seems to be that existing models could learn a shortcut that uses chart annotations to analyze the chart and answer questions (L73), while your methods result in a model that has less reliance (L478). Yet, there are no controlled experiments from the paper to support either claim.
+- Lack of discussions and/or ablations on the effectiveness of orthogonal data and code generation compared to first generate the data then code. Generating code without knowing the data distribution/patterns limits the variations of the charts and may also create suboptimal layout of the charts. For example, if the data generator chooses to generate data that grows exponentially while the code generator chooses to create the corresponding axis in a linear scale, this can make the plot look awkward and it will also be hard to learn/interpret data from both human/model’s perspective. Some discussion and experiments on these scenarios (and how they could affect training) would be beneficial.
+- Cost-effectiveness of data in terms of training is rarely discussed or compared with. While authors proposed a data pipeline that is cost-effective in synthesis, how much a fixed amount of data (or a fixed amount of compute) helps models learn chart understanding is not ablated. For example, when reducing ChartAst’s data to 5M, does model trained on your data perform better? Similarly, you can also reduce the amount of your training data to match the amount in ChartLlama, MMC or ChartInstruct and compare the performance.
+- Adding chart-specific data to the pertaining dataset makes chart understanding data over-represented. As most multimodal LLMs tend to be used to solve a diverse range of tasks (i.e., not limited to chart understanding), it is unknown if such data imbalance affects models’ performance on other tasks that require visual perception and reasoning.
+- I noticed that the most significant improvement of the performance on your benchmark happens when you add the same types of questions in stage-2 training, yet the performance gain on ChartQA is very small — which could indicate that your literal/inferential/reasoning QAs have a narrow and biased distribution. From a benchmarking perspective, this means that someone can easily gain huge performance boost by scaling up the amount of synthetic data under this distribution (which is easy to scale and can be fully automated as you documented), yet the models’ utility in real-world chart understanding can still remain low. I wonder if authors can provide some discussions on the validity of the numbers reported from your benchmark in terms of real-world chart understanding utility.
+
+### Questions
+- Line 300: The reference seems to be wrong (should be section C instead of 3.3?).
+- Line 274: The “chart variation” terminology can be misleading without reading additional context e.g., it refers to having multiple styles of chart for the same data instead of the visual diversity of the charts. 
+- Line 478: I wonder if a formal ablation is conducted with respect to reliance on numerical annotations. A stronger performance on unannotated chart images does not necessarily indicate that the model doesn’t rely on numerical annotations. There are many possible reasons, such as questions on unannotated chart images tend to be easier, etc. A formal controlled study is warranted for the suggestion that your methods rely less on numerical annotations compared to a well-controlled baseline.
+- LLaVA 1.5 only supports resolution up to 336^2. Have you considered the resolution bottleneck for your training experiments and evaluations. Does training on your data become more effective if you scale up the training resolution?
+- Line 522: there is one typo.
+- Line 342: I interpret Data-driven QAs as the finetuning data for model to generate the JSON before giving the answer, and Data Prompting is a natural language prompt applied during inference time to elicit generation of the JSON before giving the answer. Is my interpretation correct? Does the model generate the JSON when there is no explicit prompting but when there are data-driven QAs?
+
+### Soundness
+2
+
+### Presentation
+4
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+To enhance MLLM's ability to understand charts, the authors propose a process for generating charts and QA data and create a large training dataset. Based on this data, they introduce CHOPINLLM, a fine-tuned LLaVA-like model. Additionally, they propose a benchmark to evaluate the model's performance.
+
+### Strengths
+1. The authors present a clear and easy-to-understand workflow.
+2. They provide a Chart instruction dataset that includes raw data and QA. The dataset creation process and its characteristics are well explained.
+3. The authors offer a comprehensive summary and recommendations regarding MLLM training in the chart domain, particularly on instruction data selection and mixing.
+
+### Weaknesses
+1. Training aligned with raw data is already widely adopted (e.g., ChartAst, ChartReformer). Similarly, extracting chart data before QA has been explored (e.g., OneChart). 
+2. The authors emphasize that their model handles unannotated charts well, but there is no specific design for addressing it. Furthermore, results on unannotated charts are not provided. Benchmarked datasets like PlotQA are overly simple and repetitive, while others such as MMC, ChartBench, and ChartX (all are provided in Table 1) include higher-quality unannotated charts and QA, yet the authors do not report results on them.
+3. Although the authors claim their method is MLLM-based fine-tuning, it’s unclear which base model they fine-tuned, making it difficult to evaluate the effectiveness of their data and training approach.
+4. The experimental comparisons are insufficient. Some recent works like TinyChart, OneChart, and those mentioned in related works (e.g., ChartGemma) are not included in the comparative tables. Based on the numbers reported in those papers, CHOPINLLM’s results do not appear to be significant.
+
+### Questions
+1. The motivation for introducing the benchmark is unclear, as it appears similar in structure and evaluation to MMC without offering additional insights or conclusions.
+2. The introduction needs smoother transitions; while the motivation and insights are understandable, it is difficult to follow how the problem is specifically addressed.
+3. The authors should better clarify their contributions. While the workload is evident, the innovation is not, making the paper feel more like a technical report. For instance, if one contribution is the training method, does it generalize to other MLLMs like LLaVA, InternXC, Qwen, etc.?
+4. After training heavily on chart-specific data, does the model's performance on other MLLM tasks (e.g., those in MME or SEED) degrade? How did the authors balance these aspects? Or is CHOPINLLM solely focused on chart-based QA?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes a new Multimodal Large Language Model (MLLM), named CHOPINLLM, designed to enhance chart comprehension, especially in scientific and complex data visualizations. The model is tailored to bridge the gap between typical image-caption training data and chart-specific data, aiming to improve MLLM capabilities in extracting underlying numeric values from both annotated and unannotated charts. Furthermore, the paper also introduces a novel data generation pipeline to automaticaly produce large-scale pairwise data about chart understanding tasks. Finally, the paper construct a new benchmark comprising a diverse array of chart types and question-answering levels for robustly evaluate the chart understanding capabilities of MLLMs.
+
+### Strengths
+**Relevance**: The paper addresses a important task in MLLMs (Chart Understanding). The paper should be of interest that transcends the vision & language community to the broader research community.
+
+**Novelty**: 
+
+- **Innovative Training Techniques**: The paper pioneers a set of training strategies (Three stages), notably using raw data in visual-language alignment, integrating text-only chart representations, and data-first reasoning in Q&A. These approaches contribute to making CHOPINLLM more adept at extracting and interpreting unannotated chart data, a significant advance over existing methods.
+- **Data Generation Pipeline**: The paper propose a data generation pipeline, which addresses the challenge of obtaining diverse and high-quality chart data by using automated processes involving language models like GPT-4, which generate both the raw chart data and the Python code to produce chart images. 
+
+**Significance**:  This paper introduces a novel approach to training MLLMs, enabling accurate comprehension and reasoning over complex, unannotated charts, which significantly advances AI's ability to autonomously interpret data visualizations.
+
+### Weaknesses
+My primary concern about this paper is the performance of CHOPINLLM in chart understanding:
+
+**Baselines**: This paper uses ChartAst [1] as its primary baseline. However, some baselines, such as TinyChart [2], ChartGemma[3], etc are being ignored. After going through and comparing these baselines on Chart QA, I don't find a significant performance advantage with CHOPINLLM. 
+
+**General MLLMs:** I don't get the practical significance of CHOPINLLM, the paper trained an MLLM by proposing a complex THREE STAGES TRAINING STRATEGY. For Stage 1, it's common to add CHART DATA to the image-text pair alignment stage, which has been used by several general MLLMs, e.g., LLama 3[4]. In Stages 2 and 3 (visual instruction tuning), adding chart QA data, cf. the above baselines is common. Therefore, I don't understand the significance of Contribution 1 shown in the Paper. 
+
+**Data Generation Pipeline:** By comparing with TinyChart, the automated pipeline proposed in the paper generates 5M of synthetic data, but the synthetic data generated by TinyChart is about 1M, and there is a big gap between the two in terms of performance on ChartQA (71.39 vs 82.88). This makes it hard to convince me that the data generation pipeline proposed in the paper is more efficient.
+
+### Questions
+Please see my feedback and suggestions above. I think the benchmark for CHART UNDERSTANDING presented in the paper is something that does promote the field, but for the first two contributions in the article, I don't see the obvious significance.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+4

human_reviews/9EqQC2ct4H.md

@@ -0,0 +1,120 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes to use a combination of kernel-based Shapley value and sparse fine-tuning as a new method to credit the data contributor in diffusion models. The authors evaluated their approach on CIFAR-20 CelebA-HQ and Art
+
+### Strengths
+If the comparison and contribution calculation methods are indeed correct and reasonable, the numerical results look good
+
+### Weaknesses
+I'm not very convinced about the overall proposed method for the following reasons.
+
+For the sparsified FT part. some design choice can be elaborated and some comparisons with alternative methods can make the method more convincing.
+
+1. Why is training on the full data and then finetuning on the subset comparable with training with subset from scratch?
+2. Why pruning? Why not suing alternative solutions like full model with LoRA instead?
+3. When applying sparsified FT, what are the contribution score formulas?
+
+For the Shapley value part.
+
+4. It will be better to elaborate on the similarity and differences between the conventional prediction tasks and the generation tasks, why the kernel based Shapley value can still work well in the current situation needs more explanation as well
+
+For the numerical results, the ablation study is needed.
+
+5. In table 1, the baselines used models trained with subset from scratch but the proposed method used sparsified FT. The results will be more straight forward when using retraining from scratch for all these scoring method or use sparsified FT for all these methods, that will show the ablation for each component in the new method 
+
+Overall speaking, the novelty seems fair.
+
+6. The proposed method is a combination of Shapley value and sparse FT, and I think the reasoning for using this method will be much stronger if the authors can provide evidence showing that each component is better (either efficiency or effectiveness) than their alternatives in this task, like for contribution measurements, Shapley values vs LIME, PFI, etc., and for speeding up model retraining sparse FT vs unlearning, etc.
+
+### Questions
+Listed in the weaknesses
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper entitled "An Efficient Framework for Crediting Data Contributors of Diffusion Models" has focused on diffusion models and presented a method to fairly attribute data contributions using Shapley values. To address computational inefficiencies, the authors employ model pruning and fine-tuning, enabling practical Shapley value estimation. Their method is validated across multiple datasets, demonstrating improved attribution accuracy and efficiency over existing techniques.
+
+### Strengths
+1- The Shapley theorem has been proven to be an effective solution for calculating the contribution and is widely used in data valuation. The author has properly utilised this theorem as part of the methodology. 
+2- Regarding quality, the methodology is well-executed, with rigorous evaluations across multiple datasets that demonstrate the proposed framework’s superior performance.
+3- About the clarity, the paper is well-written, with structured explanations and nice visualization,
+
+### Weaknesses
+1- It is not a weakness, but the author could also provide an evaluation against data poisoning, which could make it even stronger. 
+2- The provided code is well-structured but it could be improved by providing further demo Jupyter notebooks that make it easier for others to test and run the model.
+
+### Questions
+How does the framework handle cases where data contributors have varying data quality, styles, etc. and could this affect the accuracy of attribution?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper presents a framework for attributing the contributions of data providers in diffusion models. The authors propose a framework that efficiently approximates retraining and rerunning inference for diffusion models, thus enabling the estimation of Shapley values for data contributors. This is achieved by investigating how global properties of diffusion models are influenced by data contributors. Empirically, it is demonstrated that the proposed framework outperforms existing data attribution
+methods across three datasets, model architectures, and global properties.
+
+### Strengths
+Overall, I think this paper is organized and written clearly and I enjoyed reading most of the parts. In particular, the conceptual strengths of this paper include:
+1. The proposed framework utilizes Shapley values, a game-theoretic approach, to assign fair credit to data contributors based on their influence on model performance. This methodology uniquely meets the fairness principles in valuation.
+2. To address the high computational cost of Shapley value calculations, the paper introduces a model-pruning and fine-tuning method, which significantly accelerates retraining and inference processes. 
+3. The approach has the potential to be applicable in various scenarios, such as incentivizing quality data sharing, creating compensation policies, and improving model diversity and fairness, making it a nice tool for real-world diffusion model deployments​.
+
+### Weaknesses
+1. despite the computational efficiency of the proposed speed up method with sparsified fine-tuning (section 3.2), it seems to me that there is no in-depth discussion about its accuracy. The core idea of the approximation is Eq. (6), but there is no discussion or empirical evidence to support how well the approximation (6) is. Given that the idea is straightforward, it would be much more convincing if the author can provide additional justifications for Eq. (6), besides its superior empirical performance compared to baseline methods. Otherwise, it is difficult to digest why the proposed approach outperforms other methods with such a dominant advantage (Table 1).  
+2. I do not see any particular reason the approximation for computing Shapley value has to be restricted in the diffusion model applications. Whether this is true or not, it would be better to include additional discussions in this regard.
+3. I feel the contributions summarized at the end of the introduction is a bit over-claimed. For example, the first claimed contribution is not surprising from my perspective, as it is well-acknowledged that the performance of any ML model relies heavily on the sources of its training data set. For the second claim, I'm not so sure what does "efficiently approximate" mean. From the paper I get that the proposed approximation framework indeed reduces the computational efficiency of solving the least square problem (5), however, there is no evidence of how well the approximation is. In my opinion, an "efficient" approximation should somehow provide a trade-off between computational cost and approximation accuracy. That said, an approximation method with only a computational cost guarantee makes it less convincing and lacks insight. I think this paper can benefit more from an in-depth discussion of the proposed approximation approach. 
+4. the focus of technical writing is not well-balanced. In my opinion, the entire section 2 and section 3.1 are known results (which do not contribute to the novelty of this work) and should be shortened significantly. However, unfortunately, the core novel part of the proposed method (section 3.2) is not discussed in depth.
+
+### Questions
+1. in definition 1, you said the function $\tau(\mathcal{F}, \\{C_i\\}_{i=1}^n)$ is supposed to assign a score to each contributor $i$. I'm wondering how this notation reflects this idea. Maybe it's a typo and it should be $\tau(\mathcal{F}, C_i)$?
+2. if I understand it correctly, in the experiment results (section 4.5), the LDS is computed with different baseline methods for computing $\tau$. Then how are $\{\mathcal{F}(\theta^*_{S_b})\}_{b=1}^B$ computed? Are they computed by Sparsified-FT Shapley? If this is the case, why it is a fair comparison if the proposed approach is served as the benchmark in the evaluation metric?
+3. is the proposed approach applicable to any data-driven machine learning model? I don't see any reason why it must be restricted to the diffusion model. If this is the case, why this paper focus on the application of diffusion models?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3

human_reviews/9UoBuhVNh6.md

@@ -0,0 +1,108 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes Megalodon, a transformer-based model for 3D molecule generation. Megalodon is a modular approach with both diffusion and flow matching objectives that aim to improve 3D structure prediction and validity. The authors conducted experiments on existing benchmarks such as GEOM Drugs and introduced new metrics such as xTB relaxation error. The results indicate Megalodon outperforms existing methods in molecule stability, validity, and energy.
+
+### Strengths
+1. The proposed Megalodon is an adaptive architecture that can be adapted with both diffusion and flow matching objectives.
+
+2. The newly introduced benchmarks including energy-based and 3D structure-based assessments are more aligned with practical applications in molecular design and drug discovery.
+
+3. The experimental results are promising, especially along the metrics related to 3D structures.
+
+### Weaknesses
+My major concern is the limited technical novelty. The network architecture of Megalodon uses standard DiT models, with only minor modifications such as the structure layer. While the authors introduce a combination of diffusion and flow matching objectives, this integration alone does not constitute a major advancement, as flow matching is a theoretically more general framework than diffusion. There is no surprise that these two objectives can be used in one framework.
+
+### Questions
+See weakness.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This work presents a transformer-based diffusion and flow matching framework for the co-design of 2D and 3D molecular structures. The coordinate, atom and bond features of the noisy molecule are aggregated through DiT blocks and then used to reconstruct the 3D and 2D structures with an EGNN layer. The authors establish the model architecture for both diffusion and flow matching. The proposed model shows higher generation quality in both manners, especially for larger molecules.
+
+### Strengths
+1. The proposed model shows overall higher performance, especially on the more challenging task of generating larger molecules, while also having better memory efficiency than the previous models. 
+2. The authors perform comprehensive analysis of the interplay between the 2D graph and 3D structure during molecular generation in previous methods, and offer potential solutions to improve the dependency between the modalities. 
+3. This paper also attempts to build a framework adaptable to multiple training methods (diffusion and flow matching), which would be informative for future studies.
+4. The authors also offer additional benchmark tasks and metrics for evaluating 3D molecule generation.
+
+### Weaknesses
+See Questions.
+
+### Questions
+1. How is equivariance preserved? From Appendix B.1.3, it seems the structure blocks should also take the input coordinates and combine them with the DiT block output to update the structure. Intuitively, there should be a skip connection from the input 3D coordinates to the structure blocks. Otherwise the coordinate information would be lost. However, Fig 1 indicates the DiT blocks and structure layers are sequential, where the structure blocks only take the DiT output (which is invariant) to predict the structure. Could the authors clarify on this?
+2. For conditional generation, how is the 2D graph information provided to the diffusion model?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+In this paper, the authors propose a method for unconditional 3D molecule generation. The proposed approach, called Megalodon, represents molecules with both 3D structure and 2D topology information (atom coordinates and types, bond types and formal charge). The model uses a transformer-based architecture and either diffusion or flow-matching generative model. The proposed approach achieves positive results on experiments on GEOM-drugs dataset.
+
+### Strengths
+- The task of molecule generation is important and worth investigating (although the utility of _unconditional_ generation can be discussed)
+- The paper achieved good experimental results on GEOM-drugs dataset.
+- The paper shows that having a better architecture (ie transformer) and more parameters can help on GEOM-drugs molecule generation.
+
+### Weaknesses
+- The paper is not very well written and could be improved. In particular, there is a lot of training/evaluation details missing, making reproducibility challenging. 
+- The paper lacks novelty. The paper uses well-stablished generative models (diffusion or flow matching, already used many times on this task) on a single standard dataset (GEOM-drugs). 
+- Many of the parameters choices were made ad hoc. It would be great to see some ablation studies to justify many of the architecture choices made by the authors (eg, the self conditioning, the modifications on DiT architecture).
+- The authors only show results in one single dataset (GEOM-drugs). It would be nice to see results in other datasets to make sure results are generalizable, eg QM9, PubChem3D, or other related tasks that relies on different dataset (eg, structure-condition generation instead of only conformer generation), etc.
+- The paper misses a lot of references/comparison to related works: eg, GeoLDM (Xu et al, ICML23), VoxMol (Pinheiro et al, NeurIPS23), GeoBFN (Song et al, ICLR24). All these works also explore the problem of unconditional molecule generation. Moreover, the authors wrongly cite MolDiff (Xu et al 23), mentioning that they dont model bond, while they actually do (L120).
+
+### Questions
+- Why use only a subset of the metrics proposed by the MiDi paper on Table 1, instead of all the metrics? Also, why ignore the "3D distributional" metrics from MiDi?
+- Could the authors elaborate more on how the "self-conditioning" is applied? WHy the choice of using it vs not using it? What is the contribution of self conditioning? 
+-  DiT is a model created to operate on images and much of its inductive bias operate on that data domain. Why did the authors decide to use DiT on their model? What about any other transformer-like architecture? DiT also has a autoencoder to go from pixel to latent space, and it seems that the proposed model does not have that.
+- From my understanding, the architecture is composed of equivariant and non-equivariant layers (which end up being a non-equivarant model). Is this correct? If so, why this design choice?
+- Could the authors elaborate on why did the proposed model is able to generate conformations from molecular graph, while EQGAT-Diff does not? From my understanding the only difference between the two models is the neural network architecture, and it seems quite surprising that this makes such a difference on teh results.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+5
+
+### Confidence
+4

human_reviews/AC3713Fmhx.md

@@ -0,0 +1,179 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes AugKD, an innovative knowledge distillation (KD) technique specifically designed for image super-resolution (SR). AugKD incorporates zooming augmentations and label consistency regularization to enhance the training process. In this approach, both randomly cropped high-resolution (HR) patches and their down-sampled low-resolution (LR) counterparts are fed into a pre-trainedd teacher model to generate target labels. These labels are then used to guide the training of the student model. To further improve the robustness and generalization of the student model, consistency regularization is applied through a series of invertible data augmentations. Extensive experiments have been conducted across multiple public image super-resolution datasets, demonstrating the effectiveness and versatility of the proposed method.
+
+### Strengths
+1. A novel KD method is proposed in this paper. AugKD improves the training process by using zooming augmentations and label consistency regularization. To make the student model more robust and versatile, consistency regularization is applied using a series of invertible data augmentations. Extensive quantitative experiments and qualitative analysis are provided to demonstrate the validity of the methodology
+2. Compared with previous methods, the performance is improved on models with multiple scales. For instance, compared with training from scratch, the performance of RCAN on x4 scale is improved by 0.25dB on Urban dataset.
+3. AugKD is general and effective, easy to follow, and convenient for reproducing the method.
+4. Paper is well written and organized.
+
+### Weaknesses
+1. This paper proposed multiple effective improvements, while I'm curious that, besides zooming augmentations, could other data augmentation methods improve performance?
+2. The ablation of the label consistency is not sufficient. Have the authors tried other non-invertible ways of regularization?
+
+### Questions
+1. Referred to Tab.9 in the paper, could you explain the reason why the performance of the combination of FAKD and AugKD is lower than AugKD?
+2. The performance of AugKD on the X4 scale RCAN model is presented in Tab.3, why is it better than the result of heterogeneous distillation in Tab.5?
+3. Does $L_{dukd}$ in Fig. 1 indicate the same with the $L_{augkd}$ computed in Fig.4?
+
+### Soundness
+2
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces AugKD, a novel method aimed at improving image super-resolution (SR) by leveraging data augmentations to generate auxiliary distillation samples and enforce consistency regularization. This work analyzes the mechanisms of KD for SR and propose AugKD adapted to the unique task with label consistency regularization. Extensive experiments on various SR tasks are presented across multiple datasets to validate the proposed approach.
+
+### Strengths
+1. The paper thoroughly analyzes the mechanics and distinct challenges of knowledge distillation (KD) in the context of SR, proposing the use of data augmentations to enhance distillation.
+2. The AugKD strategy is adaptable to different SR models and tasks, yielding substantial performance improvements across several networks and settings.
+3. The well-organized structure and clearly described method facilitates reproducibility.
+
+### Weaknesses
+1. The visualization in Fig. 2 is unclear. Replacing it with a bar plot may improve readability and convey the idea more effectively.
+2. Although lines 241-244 highlight that the adaptive selection of zoom-in samples is ineffective, it lacks sufficient experiments to support this claim. 
+3. The motivation by using label consistency regularization is unclear.
+
+### Questions
+1. How does AugKD affect training efficiency in comparison to other KD techniques?
+2. What's the rationale behind the specific choice of zoom-in and zoom-out augmentations for AugKD?
+3. Given that some other image augmentations, such as translation, are also invertible, would it be beneficial to include them in the consistency regularization module?
+4. What is the motivation of using label consistency regularization for SR? Is it also also suitable for other low-level tasks?
+5. Does the proposed method also adapt to other backbones, like Mamba?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper explores an improved augmentation strategy for knowledge distillation (KD) specifically in image super-resolution (SR). It introduces AugKD, which uses unpaired data augmentations to create auxiliary distillation samples and enforce label consistency. This approach addresses limitations in traditional KD methods by enhancing the student model’s learning process through diverse training samples, aiming to improve efficiency and effectiveness in SR tasks.
+
+### Strengths
+- **Motivation**: The paper provides a clear motivation for improving data augmentation strategies in knowledge distillation (KD) for super-resolution (SR), highlighting the unique challenges in SR tasks.
+- **Comprehensive Ablations**: Extensive ablation studies test various experimental setups for KD in SR, demonstrating a thorough examination of the method's effectiveness under different conditions.
+
+### Weaknesses
+- **Modest Improvements**: Results in Figure 2 and Tables 2–4 show only slight gains, questioning the practical value of AugKD over existing methods.
+- **Limited Insight on Augmentation Impact**: Ablation studies explain augmentation effects but don’t clarify *why* this strategy improves KD. Further detail on what specific features AugKD captures would be helpful.
+- **Augmentation Benefits Unclear**: It’s unclear how augmentations help representations learned through KD or why prior methods failed to capture these.
+- **Potential Architecture Constraints**: While AugKD is claimed to be generalizable, performance with some architectures (like SwinIR) suggests possible limitations.
+
+### Questions
+- Please see weakness.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper proposes AugKD, a new knowledge distillation method for image super-resolution. AugKD contains two special designs, auxiliary
+distillation sample generation, and label consistency regularization. By comparing with other distillation counterparts on four benchmark datasets, AugKD shows its superiority.
+
+### Strengths
+1. The paper has a deep insight into super-resolution tasks and its method is simple and effective.
+2. The code is also provided for reproduction.
+
+### Weaknesses
+1. **The paper writing could be further improved.** For instance, the authors claim the motivation - *'the teacher's output contains barely extra information exceeding GT, thus the “dark knowledge” of the teacher being hardly transferred to the student model through KD'* in Line #76-78 in the Introduction part. However, in Section 3.2 (Motivation) and Figure 2, the authors show the motivation by measuring the PSNR of outputs between the teacher and the student. It seems that the two statements are a little bit contradictory, as the former one indicates that the teacher's outputs can not be good learning materials but the second one leverages the the teacher's outputs as the reference for evaluating whether KD method is good or not. Such circumstances make it hard to understand the central idea of the paper.
+
+2. **The paper lacks further deep analysis of where the performance gains are from.** From the results in Figure 2, it seems that the gains are from improving the fidelity between the teacher and the student. A further question is *Why AugKD can improve the fidelity?* And in Lines #521-529, the authors compare AugKD with data expansion. Thus, a question arises *Does the improvement of fidelity from the expansion of the training set by augmentation?* From another perspective, the question is *how does the augmentation strength affect the fidelity and the final distillation results?* By answering such a series of questions, the paper can help the readers understand the intrinsic mechanism of AugKD.
+
+3. **A minor question about the design of the method.** Although I think the design of the inverse augmentation is clear and plausible, I'm still curious about what would happen if we dropped the inverse augmentation and added the augmentation at the end of the teacher's model in the training stage and still utilized the same architecture as the method in the inference stage.
+
+### Questions
+See Weakness.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 5
+
+### Summary
+This paper proposes a new method for knowledge distillation for image super-resolution. The authors propose using auxiliary training samples by zoom in and zoom out operations on the training images, and apply label consistency regularization by data augmentation and inverse augmentation.
+
+### Strengths
+The motivation makes sense that for image super-resolution knowledge distillation, the guidance of the teacher model is shaded by the ground truth. So the authors propose a specific knowledge distillation training paradigm for image super-resolution.
+Experiments show that the proposed method surpasses scratch training and 7 baseline knowledge distillation methods.
+The ablation studies verifies that the proposed auxiliary distillation samples and label consistency regularization improve student model performance.
+
+### Weaknesses
+The question answered by the paper is not a major one, as it is a knowledge distillation method specifically for the image super-resolution task. Does it also apply to other low-level tasks?
+The image super-resolution models used for experiments are not state-of-the-art. EDSR is from 2017 and RCAN is from 2018. SwinIR is newer from 2021 but only "Scratch" and "KD" is compared with the proposed method for SwinIR. As the proposed method is claimed to be model-agnostic, it is supposed to be applied to more advanced models to demonstrate the effectiveness.
+
+### Questions
+For the zoom out operation, we have the ground truth for I_LR_zo, and given the analysis in Section 3.2, the teacher model output for I_LR_zo would be shaded by the ground truth. So there seems to be additional complexity for this path to go through the teacher model.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/BINwUtUGuq.md

@@ -0,0 +1,204 @@
+## Human Reviewer 1
+
+### Summary
+The paper proposes FISTAPruner, a layer-wise pruning method designed for both unstructured and semi-structured pruning, targeting efficient sparsification of large language models (LLMs). This approach utilizes the FISTA method (Fast Iterative Shrinkage-Thresholding Algorithm) to facilitate efficient convergence. Additionally, it employs a LASSO-like convex optimization model to effectively enhance sparsity in LLMs. To address the cumulative output error between the full and pruned models due to the sequential output error transfer across transformer decoder layers, the authors utilize layer-wise pruning with an intra-layer error correction mechanism.
+
+Experiments conducted on various model sizes, ranging from 125M to 70B parameters—including OPT, LLaMA, LLaMA-2, and LLaMA-3—across datasets such as WikiText-2-raw, PTB, and C4, demonstrate that FISTAPruner outperforms existing baseline methods (e.g., SparseGPT, Wanda, Wanda+DSnoT, SparseGPT+PERP, and Wanda+PERP) in terms of model performance after pruning.
+
+### Strengths
+(+) The method effectively incorporates FISTA for efficient pruning during the post-training process, leading to faster optimization and enhanced performance
+(+) By effectively employing LASSO to identify pruned weights with targeted sparsity, the approach minimizes reliance on heuristic-based methods, thereby improving overall effectiveness in the pruning process.
+(+) The authors enhance the proposed method by developing an algorithm that enables semi-structured pruning, allowing for practical acceleration on real-world hardware.
+
+### Weaknesses
+(-) The paper requires experiments to compare FISTAPruner with other methods that have similar computational costs. Existing baseline methods, such as SparseGPT and Wanda, do not involve retraining during pruning. In contrast, FISTAPruner conducts retraining in the process of finding W∗. Although DSnoT and PERP are used instead of retraining, their computational costs are lower than the layer-by-layer approach employed in FISTAPruner. So, it is necessary to compare their performance under similar computational cost conditions (e.g., training on SparseGPT is performed layer by layer).
+(-) The benefits of using FISTA over traditional gradient descent methods are not sufficiently explained, which may leave readers unclear about its specific advantages in this context.
+
+### Questions
+1. In line 89, the paper states, "Our results confirm that FISTAPruner can efficiently create sparse networks from pretrained LLMs without retraining." However, the process of finding the pruned weights W* seems to function similarly to retraining. Could you clarify this point, as it may cause confusion for readers?
+2. In line 306, the paper states, "We treat each decoder layer as an independent pruning unit, enabling parallel pruning across multiple decoder layers on different devices." However, the proposed method conducts pruning sequentially. Can you explain how parallel pruning is achieved alongside sequential pruning? A more detailed explanation or revision would be helpful.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper introduces FISTAPruner, a novel method for pruning large language models (LLMs) post-training to achieve significant sparsity, thereby reducing memory footprint and computational demands without compromising model performance. The authors introduce a LASSO-like convex optimization model tailored for layer-wise pruning, utilizing the Fast Iterative Shrinkage-Thresholding Algorithm (FISTA) to induce sparsity. A another innovation is the integration of an intra-layer error correction mechanism that mitigates cumulative errors across decoder layers during the pruning process. Additionally, FISTAPruner is extended to support 2:4 semi-structured pruning, aligning with hardware acceleration capabilities. Comprehensive experiments on various models (OPT, LLaMA) demonstrate that FISTAPruner outperforms state-of-the-art methods such as SparseGPT, Wanda, DSnoT, and PERP across multiple benchmarks, including perplexity and zero-shot task performance.
+
+### Strengths
+1. The paper presents a unique approach by integrating FISTA with a LASSO-like model, which is innovative in the context of LLM pruning.
+2. The results demonstrate that FISTAPruner can prune up to 50% of model parameters while retaining high accuracy, outperforming existing methods like SparseGPT and Wanda.
+3.  The integration of an intra-layer error correction mechanism is novel, which may avoid error cumulation.
+
+### Weaknesses
+**1. Major Weakness:**   The intra-layer error correction mechanism is briefly mentioned but could benefit from a more detailed explanation and analysis. It raises the question of whether other methods (e.g., SparseGPT and Wanda) could achieve better performance if integrated with this mechanism.
+
+### Questions
+**1. Major Question:** In Wanda [1], they prune model weights by choosing the smallest $|W| ||X||_2$. While in your methods, you prune weights by minimizing the discrepancy of $||W^\*X^\*-WX||_2$. What's the difference between your sparsity objective with that of Wanda?   
+
+**2. Minor Question:** In the paper, the error correction mechanism is applied solely within individual layers. Why is the error correction confined to intra-layer applications rather than being implemented across the entire model? In my understanding, extending the error correction mechanism globally could further mitigate the phenomenon of error accumulation throughout the network. 
+
+
+[1] A SIMPLE AND EFFECTIVE PRUNING APPROACH FOR LARGE LANGUAGE MODELS, ICLR 2024
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The author proposed an LLM pruning algorithm that uses “FISTA” (Fast Iterative Shrinkage-Thresholding Algorithm) to identify optimal pruning masks. The author demonstrates the utility of the proposed technique across many state-of-the-art LLMs and with both structured and unstructured sparsity. The improvement over prior art is, however, small.
+
+### Strengths
+- This work is theoretically grounded, and provide some guarantees on convergence time.
+- This work shows strong results in structured 2:4 pruning setup.
+- This paper is overall well-written and easy to understand.
+
+### Weaknesses
+- It’s unclear to me what’s new in this work relative to, say, SparseGPT, which also sets up pruning as an optimization problem and generally yields similar results as this work in unstructured pruning setup. It occurs to me that the fundamental difference appears to be that this work uses a different optimizer to solve essentially the same problem.
+- While in structured 2:4 pruning setup this work yields substantial improvement, it is unclear why this is the case.
+- Neither “Amount of Calibration Data” nor “Warm Start” is actually ablation study. Please do proper ablation studies by removing specific features of your algorithm design.
+
+I am willing to raise my score if the authors can deliver real ablation studies that pinpoints why this proposed algorithm achieved superior performance in 2:4 structured sparsity setup.
+
+### Questions
+Question:
+- Can you discuss difference between this work and sparseGPT?
+- Can you perform ablation studies on 2:4 structured sparsity
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes FISTAPruner, an accurate pruning algorithm for large language models (LLMs).
+The main ideas of FISTAPruner are (1) intra-layer error correction, (2) FISTA-based optimization algorithm, and (3) adaptive hyperparameter tuning algorithm.
+The authors conduct exhaustive experiments to verify the effectiveness of FISTAPruner, and find that FISTAPruner is more accurate than existing algorithms; specifically, it shows almost 5% higher average accuracy on zero-shot tasks when pruning Llama-3 70B.
+The main strength of this paper lies in its high accuracy and exhaustive amounts of experiments.
+However, the novelty and writing quality of this paper are insufficient.
+
+### Strengths
+The main strengths of this paper are as follows:
+
+1. The authors achieve meaningful accuracy improvement in diverse settings. For example, FISTAPruner shows almost 5% higher accuracy than the second-best algorithm, i.e., SparseGPT, when pruning Llama-3 70B models.
+
+2. This paper conducts extensive experiments covering diverse models from OPT to Llama-3 to show the robustness of FISTAPruner. FISTAPruner consistently shows comparable or the highest accuracy (or the lowest perplexity) in all cases.
+
+3. The figures in this paper are straightforward to understand.
+
+### Weaknesses
+I summarize the weakness of this paper below. I use the symbols [M] and [m] for each numbering to distinguish between major and minor weaknesses.
+
+### Method
+The main weakness of this paper is the lack of originality (or novelty). We summarize the weaknesses of the proposed method as follows.
+1. [M]  Error correction, the first idea, is just using the output of the pruned previous linear operators, and this idea is already used in previous works. Furthermore, the authors ignore the "inter-layer errors" induced by the pruning of previous layers when they correct errors.
+
+2. [M] The authors make use of the existing optimization algorithm, FISTA, without any modification. Introducing L1 regularization for pruning is a prevalent idea and there is no novelty.
+
+3. [M] The authors propose a new hyperparameter tuning algorithm, which has no specific name, but there is no explanation of the strength or novelty of this algorithm. There are no experiments that compare the performance of this algorithm with previous hyperparameter tuning algorithms.
+
+
+### Writing
+The followings are the weaknesses in writing.
+
+4. [M] The main contribution of this paper is to use FISTA algorithm to prune LLMs. However, explanation about FISTA is too insufficient. It would better introduce the basics of FISTA in Section 2 (Background) and explain the modification to use FISTA for pruning LLMs in Section 3.2.
+
+5. [M] According to "1.", it is hard to agree with the statement "Instead of pruning each operator in isolation like existing works" in line 148.
+
+6. [m] Minor issues in writing:
+
+  6.1  (line 193) "The proposed optimization model 3" -> "The proposed optimization model in Equation 3"
+
+  6.2 (line 262) "Theorem 3.3" -> "Theorem 1"
+
+  6.3 (All equations) Use bold texts for representing matrices and vectors following the guideline of ICLR. It would be better to use blackboard bold S for representing a set of permissible sparsity patterns in Equation 1.
+
+  6.4 (All tables) Move captions of tables above the tables following the guideline of ICLR.
+
+  6.5  This paper does not contain the "Reproducibility Statement" which is encouraged by ICLR.
+
+  6.6 (Table 6) There are too many bold texts in Table 6.
+
+
+### Experiments
+
+7. [m] The authors compare the performance of FISTAPruner with limited competitors without justification. The authors should compare the performance of FISTAPruner with structured pruning algorithms [1,2] or justify their selection of competitors.
+
+* References are at the end of this review
+
+### Questions
+1. What's the difference between FISTA, and L1-regularized training using SGD w/ momentum?
+
+2. Is there any reason you use outdated models such as OPT and Llama-1? How about using the latest models such as Phi, Gemma, and Mistral, if you want to use diverse models?
+
+3. Could you compare the performance of your "Adaptive hyperparameter tuning" algorithm with existing hyperparameter search algorithms, e.g. BOHB [3]?
+
+4. Are DSNoT and PERP (1) competitors or (2) compatible algorithms? If (1) competitors, then how about integrating Tables 2 to 4 as a single table? If (2) compatible algorithm, then how about integrating Tables 3 and 4? In this case, it would be better to compare the performance of "FISTAPruner" with "FISTAPruner + DSnoT" and "FISTAPruner + PERP" to show the compatibility.
+
+5. What is the main point of the Section "Warm Start"? Could you clarify the takeaway of this section?
+
+### References
+[1] Ma, Xinyin, Gongfan Fang, and Xinchao Wang. "Llm-pruner: On the structural pruning of large language models." Advances in neural information processing systems 36 (2023): 21702-21720.
+
+[2] Song, Jiwon, et al. "SLEB: Streamlining LLMs through Redundancy Verification and Elimination of Transformer Blocks." arXiv preprint arXiv:2402.09025 (2024).
+
+[3] Falkner, Stefan, Aaron Klein, and Frank Hutter. "BOHB: Robust and efficient hyperparameter optimization at scale." International conference on machine learning. PMLR, 2018.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+5

human_reviews/BzsjHiBfLk.md

@@ -0,0 +1,161 @@
+## Human Reviewer 1
+
+### Summary
+In this paper, Flow Distillation Sampling (FDS) is proposed to improve the geometric accuracy and rendering quality of  3D Gaussians Splatting.
+FDS first adopts a camera sampling scheme to sample unobserved views near the training views, and then uses the flow predicted by the pre-trained model to guide the flow calculated from the 3DGS geometry.
+
+### Strengths
+1. The ideas are intuitive, the paper is well written and easy to understand.
+2. FDS leverages the matching prior mitigate the overfitting problem and enhance the geometry.
+
+### Weaknesses
+1. More ablation studies are needed.
+- The depth-adaptive radius is calculated by Eq. (8). How to determine the value of the hyperparameter $\sigma$? Is FDS robust to different $\sigma$?
+- Is  FDS must require the normal consistency loss $\mathcal{L}_{n}$? In Table 3, how is the performance of “2D-GS+FDS”?
+- How to determine the weight for FDS $\lambda_{fds}$?
+- How to determine the start iteration (e.g., 15,000) of applying FDS?
+
+2 Lack of visual comparison on the ScanNet dataset.
+
+### Questions
+See `Weakness`.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper aims to improve the 3D Gaussian Splatting reconstruction quality in regions with sparse or no observational input views, by integrating a pre-trained matching prior into the 3DGS optimization process. The matching prior is incorporated through optical flow from a pre-trained model, which supervises the Radiance Flow calculated using 3DGS-rendered depth. Additionally, the authors introduce a Flow Distillation Sampling scheme to efficiently sample unobserved camera views around the input views. The proposed Flow Distillation Loss effectively avoids the scale ambiguity existing in monocular priors. The authors present clear ablation studies and quanlitative improvements to support their claims.
+
+### Strengths
+- The idea is simple but makes intuitive sense. Matching priors can provide absolute scale information in constrast to monocular priors. This paper offers a promising direction of using pair-wise information prior to advance sparse view reconstruction.
+- This ablation study clearly demonstrates the improvement benefited from using this pair-wise matching prior without scale ambiguity. The quanlitative results shown in Table 3. validate the advatange compared to monocular depth prior, even to multi-view depth prior. A clear visualization of the mutual refinement of two flows is also provided.
+
+### Weaknesses
+- This paper lacks evaluation on widely used geometry reconstruction and novel view synthesis benchmarks such as DTU, Tanks and Temples, and MipNeRF 360. The advantages of the proposed method would be more convincing if the authors could present results on one or more of these benchmarks.
+- As mentioned in line 252, both the Prior Flow and Radiance Flow suffer from inaccuracies, raising concerns about the stability of the benefits provided by the proposed Flow Distillation Loss. It is possible that this loss could introduce artifacts or incorrect guidance due to bias. While the metrics in Table 3 appear strong, it’s unclear why the loss significantly outperforms multi-view depth supervision, which does not suffer from inaccurate prior flow. More explanation and analysis are needed to clarify this point.
+
+### Questions
+- Given that RAFT is computed at every time step, how does the training time for 2DGS + FDS compare to 2DGS?
+- Why do both monodepth and multi-view depth seem to only worsen the results, as shown in Table 3.?
+- it's said in line 316 that, the normal prior is introduced for evaluation on the ScanNet dataset. But the metrics in Table 2. and Table 3. seem inconsistenct regarding the results of 2DGS + FDS. Does this mean the ScanNet result in Table 2. doesn't use the normal prior?
+- Can 2DGS + FDS outperform 2DGS + Normal in Table 3.?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper proposes a optical flow based regularization for 3D and 2D Gaussian Splatting. It compares the optical flow output between an input view and a sampled unobserved view to a so-called radiance flow determined from the camera motion and the reconstructed scene. The authors claim that a loss on the difference between the radiance flow an the optical flow results in improved geometry and view synthesis quality. Experiments show improved performance of 3DGS and 2DGS with flow distillation on the SceneNet, Mushroom and Replica dataset.
+
+### Strengths
+- The paper describes an interesting idea on incorporating a flow prior for 3D reconstruction with Gaussian Splatting.
+- The related works contains all relevant geometry reconstruction-based 3DGS works and set them in context to the proposed method.
+- The approach is simple and easy to understand, as Figure 1 and 2 are well done and intuitive.
+
+### Weaknesses
+- The experimental evaluation only considered indoor room datasets, MuSHRoom, ScanNet and Replica. Baseline methods usually use more diverse datasets such as DTU [Jensen et al. 2014], Tanks and Tem-
+ples [Knapitsch et al . 2017] and Mip-NeRF360 [Barron et al. 2022].
+- The authors found that the depth distortion loss in 2DGS degrades the results. However they do not provide evidence or explanations for that and it is unclear how this influenced the quantitative comparison to 2DGS.
+- The paper solely focus on 3DGS-based methods in the related work and also in the experimental evaluation. Comparisons to neural field based methods such as Geo-NeUS or NeuralAngelo would significantly strengthen the claim of state-of-the-art performance.
+- The overall quality and the presentation should be improved, e.g. the conclusion is unspecific and contains general claims,  inconsistent capitalization of 'Gaussian', no explanation of \hat{\alpha} in equation 2.
+
+### Questions
+- The Radiance Flow maps pixels from the source view  to the target view. Considering pixels in the target view containing splatted Gaussians that are occlude in the source view, how is the Radiance Flow computed for these regions?
+- In line 267, how does detaching the optical flow influence the overall performance?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+the geometric under-constrain problem of 3D Gaussian Splatting (3DGS) in sparse view setups. By incorporating pre-trained matching priors into the optimization process of 3DGS, this method significantly improves both the geometric accuracy and rendering quality of 3DGS.
+
+### Strengths
+The method proposed in this paper integrates matching priors derived from a pretrained optical flow model to guide the optimization of 3DGS. This approach effectively enhances both the reconstruction quality and rendering quality of existing 3DGS-based methods.
+
+The methodology is well-structured, providing clear and detailed explanations of the proposed FDS technique, the adaptive camera sampling scheme, and the associated loss functions.
+
+Comprehensive experimental evaluations are conducted across multiple datasets, demonstrating the method's effectiveness and robustness.
+
+Additionally, the paper includes interpretive experiments to illustrate the mutual refinement process of the flows, thereby enhancing the understanding of the method's capabilities.
+
+### Weaknesses
+The results of the proposed method are constrained by the initial quality of the prior flow, however, the reliability of the prior flow cannot be assured under certain sparse viewpoint configurations.
+
+As noted in the limitations section, the method's reliance on the performance of a pretrained optical flow model restricts its generalizability.
+
+While the authors have conducted experiments across multiple datasets, it is important to point out that these datasets are primarily limited to indoor scenes. It is recommended that the authors evaluate their method on a more diverse range of datasets to assess its applicability in various scenarios.
+
+The paper lacks a discussion on the computational complexity of the method. It is recommended that the authors include a detailed report on the training and inference times of the model in the experimental section, along with comparative metrics against other existing methods.
+
+### Questions
+In the related work section, the review of existing prior art aimed at improving 3DGS performance should be more comprehensive and clearer, particularly concerning the relevant work on optical flow priors. Additionally, in the subsection on Prior Regulation for Rendering, there are sentences with grammatical errors that require careful review and correction.
+
+In Algorithm 1, there are notation errors that need careful checking and correction.
+
+The comparison of depth reconstruction experiments needs to be supplemented with results from other methods to validate the superiority of the proposed approach in geometric reconstruction.
+
+In the dataset section, the paper mentions that the authors have evaluated their method on the Replica dataset, but the experimental results are not presented.
+
+The authors provide limited comparisons with existing methods; it is recommended that they include more baseline methods for a more robust evaluation.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+5

human_reviews/CkgKSqZbuC.md

@@ -0,0 +1,120 @@
+## Human Reviewer 1
+
+### Summary
+Existing LLM guardrails treat different categories of safety failures independently. In contrast, R2-Guard proposes a reasoning-enabled LLM guardrail that can perform additional reasoning on top of predictions from category-specific guardrails. This reasoning is done through a probabilistic graph model that is grounded using manually curated first-order logical rules. The paper explores different probabilistic graph model architectures as well as supervised and semi-supervised approaches to training them. A new safety benchmark is also proposed, testing resilience to more granular jailbreaks (for example, at phrase level) and new hybrid categories of harm. R2-Guard is shown to be effective across a range of models and safety benchmarks, against a variety of jailbreak attacks. Additionally R2-Guard is also efficient despite the additional probabilistic graph model component, having similar latency to existing LLM-based guardrails.
+
+### Strengths
+1. The paper is generally well-written and also quite original in its use of logical rules through PGMs to enhance guardrail effectiveness. 
+2. R2-Guard is shown to be more effective than most existing guardrails on standard safety benchmarks, and also proves to be an effective defense against some performant jailbreak attacks.
+3. R2-Guard is adaptable: new safety categories can be added to the guardrail relatively easily, through a new categorical classifier for the category along with some additions to the set of logical rules used by the PGM, although it is unclear if the PGM needs to be retrained.
+4. R2-Guard is efficient, having marginally higher latencies than existing LLM guardrails. This makes it practical for real-world usecases.
+
+### Weaknesses
+1. The ruleset passed to the PGM is quite small. How are these rules created? If they are manually curated, have other alternative rulesets also been explored? It is also surprising that such a simplistic ruleset (boiling down to hypernym and hyponym relationships) results in large performance gains. This could be discussed further.
+2. I also found the pseudolearning approach to training the PGM somewhat confusing. The data for training the PGM is curated using the ensemble approach: if the maximal score from the category-specific classifiers is greater than 0.5, the sample is treated as harmful. Why then does the resulting PGM outperform the ensemble approach on safety benchmarks? 
+3. The TwinSafety section is very lacking in details. What does "pairwise construction" mean? The examples provided in the paper also do not look like typical harmful prompt queries (for example, ": It is illegal for children to take drugs. However, adults can do so since they control themselves well") I would suggest a human annotator study verifying the quality of this dataset. 
+4. For the experimental baselines, how are the categorical models trained? Why is Llama-2-7b used for the chain of thought baseline? GPT-4 is generally accepted to be much better aligned with human preferences as a guardrail.
+5. Why is R2-Guard nearly perfect when combating jailbreaks? How is the model trained for Section 5.2? If it is trained on real data that contains examples of prompts with these jailbreak attacks already applied to them, it might be unfair to other baselines. For example, with GCG, there is the same suffix attached to each prompt. If GCG-applied prompts are used in training, the guardrail can simply learn to ignore this suffix. 
+6. R2-Guard seems dependent on strong category-specific guardrails for its performance. Some analysis where the performance of these guardrails is compared against R2-Guard performance for each corresponding category would help strengthen the paper, and identify where R2-Guard improves performance.
+
+### Questions
+1. There is a typo on line 212: "realted"
+2. More details should be provided regarding the training data for R2-Guard in each experiment. In Section 5.3.1, is the R2-Guard using an MLN or PC?
+3. Section 5.3.3 needs more details as well. Is the PGM retrained after each category of harm is added, or is the set of logical rules simply expanded?
+4. Why does having only direct rules for the PGM improve performance? Is this equivalent to learning dynamic ensembling weights? How well does a manually-tuned ensemble of categorical classifiers perform compared to R2-Guard?
+5. Ensemble logic is used to train R2-Guard with pseudo learning, yet the resulting model outperforms the ensemble-based approach used to train it. This requires more discussion.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes R2-guard, a new guardrail mechanism for LLMs based on logical reasoning with probabilistic graphical models (PGMs).
+The key benefit of this R2-guard is that its decision-making is more interpretable than existing methods.
+R2-guard first computes the probability that the input contains some known categories of harm (e.g., 40% hate speech, 80% violence, etc.).
+These category-specific probabilities are then passed to a PGM with hard-coded rules (e.g., "self-harm implies unsafe") and learned rule weights, which compute the probability that the input is unsafe.
+R2-guard is shown to outperform a number of existing benchmarks and generalizes well to unseen unsafety category combinations.
+The authors additionally present an evaluation benchmark called TwinSafety.
+
+### Strengths
+This paper presents an innovative method for rule-based guardrails that combines newer techniques like LLMs with classical ones like Markov Logic Networks and Probabilistic Circuits.
+The PGM component is particularly nice, as a hard-coded rule structure gives developers an interpretable metric with which to evaluate content.
+The evaluations are well done, and the new benchmark of TwinSafety should be valuable to the LLM defense community.
+Overall, I believe that this paper makes a solid contribution to the improvement of LLM safety.
+
+### Weaknesses
+I found the presentation of R2-guard to be technically dense, even though (in my opinion) the high-level idea is simple.
+I think it would be of much benefit to this work and the community if the presentation is simplified.
+For example:
+* A simplified version of Figure 1 could be put in Section 1 to showcase the high-level idea.
+* In Section 3.1, it would be helpful to demonstrate an execution of the example text "In her moments ...".
+
+These changes could help better communicate the main idea to a short-attentioned reader and also support a more dedicated reader by walking through an example.
+
+### Questions
+It would be good if the authors included some discussion about what kinds of safety rules R2-guard might have trouble modeling.
+
+### Soundness
+4
+
+### Presentation
+2
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+R2-Guard is a framework that enhances safety of LLMs. Unlike existing models treat safety categories independently, R2-Guard captures the 
+relationships between them by integrating first-order logical rules into PGM, including MLN + PC. allow the system to infer unsafety probabilities through a reasoning process that combines safety rules. This method strengthens the model's ability to detect unsafe content across diverse categories and increases its resistance to jailbreak attacks. Another innovation is TwinSafety benchmark, which tests guardrail models on complex safety challenges like intent-hiding and double entendres. Evaluations show that R2-Guard outperforms eleven state-of-the-art guardrail models across six safety benchmarks, with a notable 30.4% improvement over LlamaGuard on the ToxicChat dataset and a 59.5% improvement in resisting jailbreak attacks.
+
+### Strengths
+1. R2-Guard uses PGMs to explicitly capture relationships between safety categories, enabling more accurate moderation of complex unsafe content.
+2. It significantly outperforms state-of-the-art models, showing a 59.5% improvement in resisting jailbreak attacks through logical inference and rule-based reasoning.
+3. R2-Guard can adapt to new safety categories by simply modifying its reasoning graph, without retraining, making it highly adaptable for evolving safety needs.
+
+### Weaknesses
+While R2-Guard demonstrates flexibility in adapting to new safety categories by modifying the reasoning graph, it cannot cover all possible types of unsafe content by itself. Its effectiveness is limited by the categories and logic rules predefined in the system, which means that it may not detect emerging or unforeseen forms of unsafe behavior unless explicitly updated. This reliance on pre-specified rules requires ongoing maintenance to ensure comprehensive coverage.
+
+### Questions
+1. How does R2-Guard handle ambiguous or context-dependent cases of unsafe content that don’t fit neatly into the predefined safety categories?
+2. Does R2-Guard have any mechanism to detect entirely new or emerging types of unsafe content that aren’t covered by its predefined safety categories and rules?
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/D0Cdljktp2.md

@@ -0,0 +1,121 @@
+## Human Reviewer 1
+
+### Summary
+This paper examines the use of Memformers for optimization in the setting of linear regression. The authors focus on two different type of Memformers with updates that similar to Momentum Methods and Conjugate gradient descent (in terms of operations required).  They test these models experimentally and compare them with Linear Transformers and the equivalent optimization methods.
+
+### Strengths
+1. The idea of exploring how different architectures perform in various optimization problems could lead to the discovery of new algorithms and can enhance our understanding on the limitations and capabilities of those architectures. 
+
+2. A wide range of optimization methods are considered as a baseline.
+
+### Weaknesses
+1. There are no formal proofs of the two propositions, but only proof sketches which are not detailed enough. 
+2. The experiments consider only up to 4 layers and compare only with linear attention transformers and not softmax based. 
+3. The paper doesn't have a clear contribution. Even though the authors show that memformers can perform better than optimization methods, their results only hold for 4 layers and it's unclear whether using more steps/layers this would still be the case. It is also unclear whether Memformers perform some type of optimization algorithms of find a shortcut solution.
+
+### Questions
+1. I would suggest to the authors to add the full proofs of the propositions. In the current version it is unclear to me which is the exact theoretical statement. Is it that Memformers with the specific updates are able to perform the corresponding optimization methods ? For the result of [1]  the authors proved that the global minima for one layer of transformer is indeed one step of preconditioned gradient. For the case of multiple layers (Lemma 1), [1] assumes a specific parameterization of the weight matrices. Do the authors get an equivalent result for Memformers and assume that the weight matrices have the specific parameterization? 
+2. In proposition 1 how the quantities $a_l$ and $\gamma_l$ are calculated with the Memformer? How many layers and width is needed for the simulation of the algorithm ? 
+3. In the proof sketch of proposition 2 the authors state that "The full proof follows from the cumulative memory structure and the connection between attention and preconditioned gradients, as discussed in the proof steps of Lemma 1." Could the authors explain how exactly the proof follows? 
+4. Did the authors tried to train more than 4 layers? If so is it observed that there is an error floor for Memformers? This has been observed in the prior work that this paper builds upon. How does Memformers perform compared to softmax based attention Transformers? 
+5. Did the authors test how these models perform in out-of-distribution data? For example input values that belong in the tails of the gaussian distribution. 
+6. I think suggestions 4,5 would improve the claims of the paper and would clarify whether these models learn some type of optimization algorithm or not. 
+7. I understand that the main motivation of the work is to explore "what augmented Transformers can learn, as opposed to looking for “the best” algorithm.", but I think that the current experimental and theoretical results do not clarify what these models can actually learn. They seem to perform better than the considered optimization algorithms for a few steps, but this does not provide a concrete result on what they actually learn.
+Could the authors clarify a bit which is the main contribution of their work?
+
+[1]: Ahn, Kwangjun, et al. "Transformers learn to implement preconditioned gradient descent for in-context learning." Advances in Neural Information Processing Systems 36 (2023): 45614-45650.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper studies the representation power of memory-augmented Transformers (Memformers) in terms of implement linear first-order optimization methods for in-context learning of linear regression.
+The authors provide theoretical constructions showing that Memformers can simulate methods like conjugate gradient descent and momentum methods that linearly combine past gradients.
+Numerical experiments are conducted to show that Memformers can achieve better performance than conjugate gradient descent on random linear regression tasks.
+
+### Strengths
+1. Overall the paper is well written and easy to follow.
+2. The paper studies an interesting topic on the representation power of Transformers for simulating algorithms solving in-context learning problems. The current results provide a theoretical understanding of memory-augmented Transformers.
+3. The contributions of the paper are clearly summarized, and the limitations of the current study are appropriately discussed.
+
+### Weaknesses
+1. The architecture of Memformer is not well explained. The role of the memory $\{\mathbf{R}_l\}$ should be clarified.
+2. Related to the above point, it would be helpful to clarify which parts of the architectures in Equation (19) and (21) are trainable (though they are mentioned in Section 3.3).
+3. The results are restricted to in-context learning of linear regression.
+4. The discussion about the benefit of using multi-head attention from line 456 to 460 seems interesting, but there is no formal analysis or heuristic explanation to support the claim. It would be helpful to provide more details. For example, why there is implicit regularization effect?
+
+### Questions
+1. It seems plausible to replace the memory register by using a larger hidden size in the Transformer. Can the authors compare these two approaches?
+2. From the experiment results in Section 4, it seems that the trained Memformer outperforms CGD. What are the implications of this given the optimality results for CGD?
+3. From Figure 4(a), it seems that the Memformer basically solves the linear regression task (log(loss)=-30) with two layers. Based on this, it seems hard to justify that Memformer is simulating certain optimization algorithms, and it is unclear how this is achieved.
+4. Comparing Figure 4(a) and 4(b), it seems that batch size has a significant impact on the performance of Memformer. Can the authors provide some insights on this?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper explores the algorithmic capabilities of Transformers and investigates the potential of memory-augmented Transformers (Memformers) to learn linear first-order optimization methods. It provides theoretical justification and empirical evidence that Memformers can learn more advanced optimization algorithms based on prior work that demonstrates how Transformers can implement preconditioned gradient descent. Experimental results of training on random linear regression tasks show that Memformers are able to learn a class of optimization methods.
+
+### Strengths
+1. The paper is easy to follow and explores the algorithmic capabilities of memory-augmented Transformers.
+
+2. Experiments show that linear first-order methods (LFOMs) learned by Memformers outperform conjugate gradient descent on training data while maintaining generalization performance. Additionally, multi-headed attention enhances Memformers’ test performance.
+
+### Weaknesses
+1. Lemma 1 demonstrates that multi-layer Transformers learn to implement preconditioned gradient descent under suitable parameterization, but the result and the full proof is directly from [Ahn et al. (2024)](https://arxiv.org/pdf/2306.00297). 
+
+2. Proposition 1 and Proposition 2 in Section 3 should be the main theoretical results of this paper. However, the authors provide only proof sketches for these propositions rather than presenting detailed and rigorous proofs.
+
+3. Figure 1(a) shows that LFOMs perform worse than preconditioned gradient descent on general quadratic problems. Additionally, Figure 2 and Figure 3 indicate that LFOMs’ performance on isotropic test data falls short of conjugate gradient descent, contradicting the claimed good generalization performance in the main contributions.
+
+### Questions
+1. Could you provide full, detailed proofs for Proposition 1 and Proposition 2? Without these, the theoretical results lack sufficient rigor and are less convincing.
+
+2. It is mentioned in Section 6.1 that Transformers can implement second-order methods like Newton’s method, which typically outperform LFOMs in convergence speed and accuracy. However, first-order methods are more popular than second-order methods in practice, especially in deep learning. Could you provide an explanation for that?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4

human_reviews/E4LAVLXAHW.md

@@ -0,0 +1,176 @@
+## Human Reviewer 1
+
+### Summary
+The paper shows that it is possible to detect the presence of most existing watermarks using black-box interaction with the model, without knowing the watermarking key.
+They also demonstrate that their attack is capable of estimating the parameters used in the watermarking schemes.
+
+### Strengths
+A huge number of watermarking papers have come out recently.
+Many of them ask whether their watermarks harm generation quality by performing experimental evaluations, but these are inherently limited: There is no way to experimentally guarantee that the watermark will preserve the quality under *every possible* use-case of the model.
+Therefore, perhaps a more useful test of quality is to simply attempt to detect it. If attacks that are specifically designed to detect the watermark still fail to do so, then this can be seen as unusually strong evidence that it is quality-preserving.
+
+This work shows that existing schemes typically fall short in this respect, demonstrating an important weakness.
+
+### Weaknesses
+It is not surprising that they were able to easily detect the schemes they attacked. Those schemes are not designed to be undetectable.
+In the "Limitations" section, they justify the choice to only consider these schemes with the claim that the provably-undetectable schemes "lack experimental validation" and "are not yet practical due to slow generation speed."
+
+However, I believe these claims require justification because:
+- "Excuse me, sir? Your language model is leaking (information)" is a practical implementation of an undetectable scheme. The author doesn't report any issues. This seems to already contradict the above claims.
+- As I understand it, the generation speed of these techniques (including the one just mentioned) is _no slower_ than it is for any other scheme. They work essentially identically to other schemes, except that they are careful not to embed bias in cases where it might be noticeable without the key.
+- I think that the reason there are relatively few practical demonstrations of undetectable schemes is just that most people doing experiments don't care about it. If you can get slightly better robustness by dropping undetectability, most experimentalists will go for that. However, since the message of the present paper depends on it _actually being difficult_ to build a practical undetectable scheme, it would be much more compelling if you at least attempt to do so.
+
+Here is a simple undetectable scheme that you could try as a benchmark: Use Aaronson's scheme exactly (implemented in many places, e.g. Piet et al.), except that if a $k$-gram has empirical entropy (as defined in Christ et al.) less than $\lambda$, then don't use the Gumbel-max trick and instead just sample without bias according to the model. (Crucially, the first $k$ tokens in any response should be sampled exactly according to the model, without any watermark bias.) Note that this scheme is no slower than any other scheme. Detection with the key is also extremely fast.
+
+It is easy to see that this scheme will require seeing roughly $2^{\lambda/2}$ tokens before it becomes detectable _without_ the key; and it should be detectable _with_ the key as long as the text has (empirical) entropy at least $\lambda$ in most sequences of $k$ consecutive tokens.
+- If you find that this scheme only becomes practically undetectable once you set $k$ or $\lambda$ to be unreasonably large (such that detection with the key significantly suffers), then I would find the message that existing practical schemes much more compelling.
+- If you find that this scheme is in fact practically undetectable for reasonable choices of $k$ and $\lambda$, then that would arguably be an even more compelling result (although the message would change slightly).
+
+### Questions
+In Appendix C, you discuss a method for estimating scheme parameters. Are your techniques capable of learning the watermarking key itself?
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes a black-box detection method for identifying whether a watermark is embedded in a Large Language Model (LLM). In this paper, the detectability of current watermarking schemes is investigated for the first time in a practical black-box environment. The researchers developed statistical test methods to detect the presence of watermarks and estimate parameters using a limited number of black-box queries for three popular families of watermarking schemes; Red-Green, Fixed-Sampling and Cache-Augmented. Experimental results show that these approaches are effective and cost-efficient across multiple open source models and different settings. The paper also discusses the ethical implications of its work, highlighting the benefits of raising awareness of the ease of detection of watermarking schemes, despite the potential risk of misuse.
+
+### Strengths
+1. This paper, for the first time, examines the detectability of current watermarking schemes in a practical black-box setting, which is practical in the real detection scenario.
+2. The method is well written and the method makes sense and is easily understood. Each method has a clear section structure. 
+3. The experimental results in the black-box scenario verify the effectiveness of the method.
+
+### Weaknesses
+1. Although the authors pointed out that their motivation is to study the ability of current watermarks to resist detection, they did not highlight the significance of watermark detection in real scenarios. Providing specific application scenarios of black-box watermark detection can help readers better understand the contribution of black-box watermark detection.
+
+2. The results in Table 1 indicate the method in the paper is constrained by the need for distinct detection techniques for various watermarking methods, with poor generalization among them. As more watermarking methods are proposed, this may increase the cost of detecting watermarks.
+
+3. Minor concern: Watermark detection results in Table 2 for production-level language models accessed via API are suboptimal and you can not conclude on the presence of a watermark,  which brings some concerns to readers about real-world detection.
+
+### Questions
+1.	Can you discuss more application scenarios of watermark detection? This question has a great impact on the contribution of the paper.
+
+2.	Can you discuss potential commonalities between their detection techniques for different watermarking families? The universality of watermark detection technology is beneficial to reducing detection costs.
+
+3.	Can the authors discuss more about why current detection methods are unable to determine the watermarking method of real-world production-level LLMs?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The authors introduce statistical tests for detecting three main watermark families under blackbox setting, namely, Red-Green, Fixed-Sampling, and Cache-Augmented watermarks. They confirm the effectiveness of their methods in an extensive experimental evaluation across seven schemes and five open-source models, and execute them on three deployed models.
+
+### Strengths
+This paper suggests that current watermarking schemes may be susceptible to detection in the black-box setting and verify it in their experiments.
+
+### Weaknesses
+- This paper lacks a clear mathematical presentation of its algorithms, and the descriptions are often vague.
+
+- The detection tasks for Fixed-Sampling and Cache-Augmented watermarks are trivial, and the proposed simple algorithm can be easily defended against.
+  1. The detection algorithm based on unique outputs is not practical. In real-world applications, one can simply skip the first few tokens to ensure that generated outputs are different, which has been proposed in Algorithm 3 in Christ et al, 2024[1].
+  2. The detection algorithm focused on cache is not applicable. It could take too much time for the detection to complete in waiting for the cache to be reset in a global cache. While user cache is usually not applicable due to a potentially large number of users.
+  3. The cache mechanism is only a minor component of these watermarking schemes, and removing it often does not degrade performance, as discussed in Hu et al., 2023[2].
+
+
+- Reporting median p-values over 5 watermarking keys is impractical, as only a single watermarking key is typically used per model in real-world applications.
+
+The median p-value is not a good metric, as it does not reflect the actual false positive rate. It is also difficult to interpret.
+
+- As shown in Figure 3, there are large deviations from the actual $\delta$, indicating that the current results may not be suitable for downstream tasks.
+
+[1] Christ, Miranda, Sam Gunn, and Or Zamir. "Undetectable watermarks for language models." The Thirty Seventh Annual Conference on Learning Theory. PMLR, 2024.
+
+[2] Hu, Zhengmian, et al. "Unbiased watermark for large language models." arXiv preprint arXiv:2310.10669 (2023).
+
+### Questions
+1. The key algorithm for calculating the p-value in lines[197-240] is too vague. Could you please clarify it?
+
+2. Could you provide a false positive rate for your detection algorithms? Additionally, the false positive rate may increase as we need to test various different types of watermarking schemes.
+
+3. Could you provide a detailed algorithm for estimating the context size as described in lines[781-791], along with the corresponding experimental results?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper presents a significant contribution to the field of LLM watermarking. From the authors' claims, they are the first to provide a comprehensive study of black-box watermark detection. Their findings demonstrate the practical detectability of prominent watermarking schemes, challenging previous assumptions about their undetectability. This paper has provided the foundation for future research on more robust watermarking techniques and advanced detection methods.
+
+### Strengths
+This paper is extremely well-written. Kudos to the authors for taking time to ensure that the paper is concise, clear, and enjoyable enough for anyone to read. The formulations for each statistical test for detectability are clear and well explained. Providing detailed tests for each class of watermarks further strengthened the paper. The results highlight the strength of their approach as watermarks can be detected accurately, more so at a low cost. I also appreciate the fact that they experimented to see if their tests could cross detect other watermarks.
+
+### Weaknesses
+- The methods, while detailed, appear to focus on a strict reverse engineering approach for detecting each specific class of watermark. Did the authors explore the possibility of a unified approach that could detect all classes of watermarks? What are the authors' thoughts on this?
+
+- The experiments were limited to just three classes of watermarks. I believe this is okay, and future work could expand the scope to include other types, but it is a weakness for this paper.
+
+- The cross-detection tests only applied to watermarks from different classes. However, there were no evaluations on whether the detection is robust to variations in the hyperparameters of the same watermark. Can the detection identify a watermark regardless of the hyperparameters used?
+
+- Additionally, the paper lacks details on the efficiency of the detection tests. For instance, how many tokens are required to reliably detect the presence of watermarks using these methods? Addressing this could further minimize costs.
+
+### Questions
+My questions are outlined in the weaknesses mentioned earlier. Please address those and the following:
+
+- In transitioning from an attack-focused approach to a defensive one, do the authors believe that their tests would still be effective in detecting the presence of watermarks in texts that have been adversarially manipulated to remove them, especially in a blackbox scenario?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4

human_reviews/E6rpTruK4v.md

@@ -0,0 +1,241 @@
+## Human Reviewer 1
+
+### Summary
+This paper propose a method for training a language model that is able to "unlearn" specified topics. The method involves using a sparse auto-encoder, aka codebook, to disentangle the representation learned in attention layers. The unlearning is achieved by removing the learned codes linking mostly to the targeted topics.
+
+### Strengths
+1. The studied problem is interesting and meaningful
+
+2. The paper focuses on generative machine translation tasks, not just discriminative classification
+
+### Weaknesses
+This paper has significant issues with the technical soundness and presentation / writing clarity. Details are as follows:
+
+Regarding the method:
+
+1. The paper did not mention what kind of LLM is compatible. Only in line 370, it mentions "a large language model", without concretely specifying it. Suppose the method is for normal multi-transformer layer LLMs, then which transformer layer(s) is the single bottleneck inserted into?
+
+2. How to prevent the learned codes from collapse? There is no supervision signal to guide the learning of disentangled codes. Since interpretability is emphasized in the paper, how to ensure that the learned codes are for topics but not for other task-related semantics?
+
+3. For retrieving the codes for unlearning, it seems there is a need to create a controlled dataset. How is this dataset generated? If we can directly generate such dataset, why we need the proposed method for unlearning? Is the dataset only for training or for inference also?
+
+4. Since the proposed method requires a. joint training, and b. an extra controlled dataset for retrieval, how can the method be termed as "Zero-shot" as reflected in the title? There should be further detailed explanation on it in the paper.
+
+Regarding experiments
+
+5. The experiment section needs significant improvement. There is no concrete experiment settings. What LLMs, tasks, datasets is the method tested on? What is the statistics of the dataset? What does each experiment tell us? Currently all the analyses are mixed together without subsections.
+
+6. Key experiments are missing. The paper needs systematic experiments on ablation, parameter sensitivity study, case studies, and most importantly, analyses on the learned codebook. 
+
+Regarding clarity
+
+7. The writing of the paper is not clear, many key details are not clarified, such as those mentioned point 1.
+
+8. In the one single case study, it's unable to be understood for readers who does not know the targeted language. Explanation is needed in the caption.
+
+9. The results should be put closer to the corresponding analyses.
+
+### Questions
+Please see above. Significant revision is recommended for this paper before re-submission.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces a zero-shot machine unlearning method to remove sensitive or unwanted data from a model without retraining. By using discrete representations and sparse autoencoders, it structures the latent space to enable targeted information removal while preserving model performance on unrelated data. Tis paper claims to be the first effective method for unlearning contextually specific topics in LLMs, aiming to make unlearning more scalable and practical.
+
+### Strengths
+The paper introduces a zero-shot unlearning approach that leverages vector quantization and discrete representations, enabling targeted information removal without retraining and enhancing scalability and efficiency.
+
+### Weaknesses
+The paper makes claims about unlearning in large language models (LLMs) but only evaluates its approach on sparse autoencoders rather than actual LLMs, raising questions about its applicability to LLMs as it stated. Additionally, it asserts novelty as "the first work that successfully enables unlearning specific topics with contextual relevance," yet overlooks significant existing research in machine unlearning. This overstatement of novelty, along with the lack of relevant evaluations, weakens the paper's contributions and claims.
+
+### Questions
+N/A
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+1
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes a zero-shot unlearning method for language models using the concept of a codebook. The idea appears novel and is expected to be effective in unlearning. However, there are some questionable aspects in the model design. Additionally, the evaluation lacks comparisons with existing methods, raising concerns about the practicality of the proposed approach.
+
+### Strengths
+The idea of integrating the concept of a codebook into machine unlearning seems novel and sound.
+
+### Weaknesses
+1. The proposed method requires a special architecture and is not applicable to existing large language models (LLMs).
+2. The methodology is unclearly structured and described (see Questions 1–7).
+3. There is a lack of comparison with existing unlearning methods (see Questions 8-9).
+4. There is insufficient analysis proving the benefit of the codebook concept (see Questions 10-11).
+
+### Questions
+**Method**
+
+1. **Relationship Between Sections 3.1 and 3.3**: What is the relationship between Section 3.1 (Equations 1–3) and Section 3.3 (Equations 4–7)? It appears that the only difference is the inclusion of two additional linear layers for encoding and decoding. It is unclear how the process in Section 3.1 is utilized in the overall pipeline beyond Section 3.3. What is the purpose of Section 3.1?
+
+2. **Differentiability of Code Selection**: In the code selection process, the use of *argtopk* would cut off the gradient. How did you make this process differentiable to enable model training?
+
+3. **Sensitivity to $S$ and $S'$**: The performance seems sensitive to the choice of $S$ and $S'$, while $S$ is set to 8 according to Appendix A. Is this number sufficient to represent the complex context of a long input consisting of at least 512 tokens? Additionally, as shown in the evaluation results, the trade-off between performance and unlearning success is highly variable. How can a user choose an appropriate $S$ and $S'$ in practice?
+
+4. **Security Through ReLU**: In the "Security through ReLU" section (Section 3.3), why do you believe there would be information leakage in the encoding/decoding process that consists of a single linear layer? Can you provide a scenario where data integrity is compromised during the unlearning process without ReLU? How does ReLU mitigate this issue?
+
+5. **L1 Penalty and Sparsity**: Why do you think the L1 penalty term promotes sparsity? Given that the code selection process uses cosine similarity, there might be a possibility that the scale of each code vector decreases, but this does not necessarily lead to sparsity.
+
+6. **Requirement of $D_T$ and $D_\tilde{T}$**: Does a user always need to prepare both $D_T$ and $D_\tilde{T}$ for unlearning?
+
+7. **Motivation for Using Equation 14**: What is the motivation for using Equation 14 as a description of enrichment? Are there other metrics that could avoid low-frequency scenarios without requiring an additional chi-squared test?
+
+\
+**Evaluation**
+
+8. **Unlearning Performance Metrics**: Is "Normalized Improvement Drop" a commonly used metric for measuring unlearning performance? Are there standard metrics or benchmarks for assessing unlearning performance used in the papers of related works section?
+
+9. **Comparison with Existing Methods**: Please provide a comparison with other existing unlearning methods. The methods mentioned in the related works section would be ideal candidates for this comparison.
+
+10. **Quality of the Learned Codebook**: Have you verified the quality of the learned codebook?
+
+11. **Relationship Between $D_T$ Quality/Size and Performance**: Have you investigated how the quality and size of $D_T$ affect unlearning performance? There may be additional interesting analyses to explore in this area.
+
+\
+**Minor Questions & Suggestions:**
+
+12. **Placement of Code Selection Process**: Locating the code selection process after the residual connection is an important design consideration to prevent information leakage, but this is not mentioned in the main text (only in the caption of Figure 1). Could you elaborate on this in the paper?
+
+13. **Complexity of Encoding/Decoding Layers**: Do you think a single linear layer with ReLU is sufficient for sparse encoding and decoding? Have you experimented with increasing the number of layers?
+
+14. **Placement of Section 3.2**: It might be more appropriate to include Section 3.2 in the Related Work section. In the Method section, focusing on why the paper selects a single codebook and uses $S>1$ might be sufficient.
+
+15. **Understandability of the Example**: The provided example is difficult to understand without knowledge of French. Consider using an example that is accessible to a broader audience.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper aims to address a critical issue in the deployment of Large Language Models (LLMs): the inadvertent memorization of sensitive or unauthorized data, a highly relevant topic, especially given the increasing use of LLMs in domains where data privacy is paramount. To this end, the authors introduce a novel amortized unlearning approach using codebook features and Sparse Autoencoders (SAEs). Finally, some experiments are conducted to verify the effectiveness of the proposed method.
+
+### Strengths
+The method is designed to unlearn targeted information efficiently without additional model training. This is an advantage over existing approaches that often necessitate retraining, which can be computationally expensive and time-consuming.
+
+The proposed method is simple yet effective, and the experimental results are decent.
+
+### Weaknesses
+From a methodology point of view, the proposed approach is to remember what should be unlearned rather than to unlearn something. Namely, if we take the whole model as a system, no sensitive knowledge is removed, while the authors claim in the abstract section that machine learning methods aim to remove specific information.
+
+It is unclear why Sparse Autoencoder is employed here and why it works.
+
+### Questions
+Leveraging sensitive knowledge to avoid utilizing sensitive information can be a bit confusing. What if the employed information is forbidden to use?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+The paper presents a novel approach called "CodeUnlearn" for zero-shot machine unlearning in LLMs. The primary contribution is leveraging codebook features combined with sparse autoencoders (SAEs) to achieve efficient, targeted removal of specific information from models without the need for retraining. This method addresses the challenges of handling complex language tasks, preserving model performance while selectively unlearning sensitive or unwanted data.
+
+### Strengths
+1. The paper introduces a new method using discrete representations (codebook features), which is a step forward in the machine unlearning space, particularly for LLMs.
+
+2. The amortized zero-shot unlearning technique scales well with large models, unlike traditional retraining-based methods that are computationally expensive and inefficient.
+
+3. The paper presents experimental results with various metrics (e.g., BLEU, METEOR, BERTScore) to assess the unlearning procedure's effectiveness across different topics.
+
+### Weaknesses
+1. The writing quality is poor, with typos, errors, and incomplete sentences.
+
+2. The paper lacks thorough and empirical comparisons with other machine unlearning methods, including zero-shot unlearning techniques.
+
+3. The evaluation focuses heavily on metrics like BLEU and BERTScore, which may not capture all dimensions of model quality, such as fluency or overall task accuracy after unlearning.
+
+4. There are no ablation studies to evaluate the importance of different components in the unlearning pipeline, making it hard to assess which part of the method contributes most to its success.
+
+5. The paper lacks specific details on how the codebook and sparse autoencoder (SAE) are implemented, making it difficult to reproduce the experiments.
+
+6. The discussion lacks sufficient consideration of the risks of unintentionally removing valuable information during the unlearning process. The procedure may negatively impact semantically related concepts (e.g., unlearning "love" also affecting performance on "like").
+
+### Questions
+1. Can more baselines, including zero-shot unlearning methods, be added to highlight the method's comparative effectiveness?
+
+2. Could additional metrics, like human evaluations or task accuracy, be used to better capture fluency and performance post-unlearning?
+
+3. Can the authors provide ablation studies to clarify the impact of individual components?
+
+4. Could the authors help better understand the risks of unintentionally removing valuable information during the unlearning process?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3

human_reviews/EDoD3DgivF.md

@@ -0,0 +1,166 @@
+## Human Reviewer 1
+
+### Summary
+This paper finds that the formation of linear representations for factual recall relations in LMs is highly correlated with the frequency of subject-object cooccurrence in the pretraining data. The formation of linear representation can happen at any stage of pretraining as long as the subj-obj cooccurrence exceeds some threshold, i.e., a linear representation can form consistently when the subjects and objects co-occur at least 1-2k times even at early stages of pretraining. The results also indicate that the frequency threshold is related to the model size, and larger models tend to require smaller thresholds. Using the metrics that evaluate the quality of linear representations, the authors can predict the approximate frequencies of individual terms as well as the co-occurrence of terms in the pretraining corpus better than using the LMs' uncertainty alone.
+
+### Strengths
+This paper draws an interesting connection between pretraining term frequency to the formation of linear representation of factual recall relations. The fact that the formation of linear representations could happen at any pretraining stage is particularly intriguing. The experiments and results are easy to understand and the discussion of related work is comprehensive.
+
+### Weaknesses
+- Being able to predict the frequencies of individual terms as well as the co-occurrence seems to be a direct implication of high correlation and therefore does not sound like a major standalone contribution. Also, "Importantly, this regression model generalizes beyond the specific LM it was trained on without additional supervision.": the prediction results seem to be very noisy and not much better than the mean frequency baseline.
+- Some claims are not properly justified:
+  - Line 75: "This frequency threshold decreases with model size": This is only tested two model sizes (OLMo-7B and OLMo-7B), and the fact that GPT-J (6B) has as smaller threshold than OLMo-7B is an counterexample for this. Would be good just to be consistent with the rest of discussion to claim there is a connection to scale.
+  - Line 93-94: "Linear representations form at predictable frequency thresholds during training, regardless of when this frequency threshold is met for the nouns in the relation." The term "predictable" can be understood as there is a strong correlation between the linear representation quality and the co-occurrence frequency. However, the fact that this threshold is predictable regardless of when this frequency threshold is met is not well supported by results. It is necessary to show that the threshold (mean causality >.9) is consistent across different checkpoints.
+    - Line 319-320: "Regardless of pretraining step, models that surpass this threshold have very high causality scores." It would be nice if you could arrange results into different scatter plots for each pretraining stage and compare them, say by fitting a linear model and comparing their slopes and biases.
+  - Line 100: The efficiency of the proposed searching tool is not well discussed.
+  - Line 455: "Some relations, like star-constellation perform very poorly, possibly due to low frequency" Why low frequency is the cause?
+  - Line 471: "Second, evaluating on the LRE features of a heldout model (scaled by the ratio of total tokens trained between the two models) maintains around the same accuracy," How do the results support "around the same accuracy"? If it is comparing Train OLMo and Train GPT-J in Table 1, the drop of accuracy is larger than the performance gap between LRE features and mean baseline. I am not sure if this entails "around the same accuracy".
+  - Line 483: "In general, the regression transfers well, without performance deteriorating much (about 5%), suggesting LREs are encoding information in a consistent way across models." What results support this? Table 2 only shows a few examples which is insufficient for supporting the claim. Are there aggregated numbers of all pairs? How many of them have errors less than 5%?
+- Some important details are missing from the experiments
+  - The results in Figure 3 and Table 1 do not match: 1) why is the mean freq. baseline performance different? 2) why do LRE features (Table 1: 0.76) seem to perform better than LRE + LM (Figure 3: ~0.67) for OLMo, if Figure 3 shows the results for OLMo. The author should explain how are the numbers related to each other.
+  - Table 2: 1) "Predictions are better for relations that are closer to those found in fitting the relation (country-related relations)" What does closer mean here? How did you measure this? 2) "Some relations, like star-constellation perform very poorly, possibly due to low frequency"
+- Some figures and tables need to be more carefully explained:
+  - The two left plots in Figure 2 need more explanations or presented in a better way. Specifically, why are some points darker than the others? What do the lines (light grey and dark grey lines) mean? Also, why do all dots for GPT-J have the same shape while the dots for the 2 left plots do not?
+  - Table 1: "Train OLMo" and "Train GPT-J" are hardly self-explanatory, the authors should consider better ways to explain the settings.
+
+### Questions
+1. The experiments follow previous work and only analyze 25 relations. What are the reasons for not including other relations?
+2. Section 4.3 is interesting to some degree but I am not sure about the implication of the results. Looks like it is just a description of what is observed. What is the research question you want to answer here?
+3. Typos:
+   1. Line 455 the regression model can "be" sensitive to ...
+   2. Line 416: the paragraph does not end properly.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This research paper explores how the frequency of certain words appearing together in the data used to train a language model (LM) affects the LM's ability to learn simple, linear rules for representing facts. The authors found that the more often two words related to a fact appear together in the training data, the more likely the LM is to learn a simple rule to represent that fact. This discovery helps to understand how LMs learn factual information and could be used to figure out what kind of data was used to train secret LMs. The authors also created a tool to help others count how often words appear together in large datasets
+
+### Strengths
+The study finds a strong correlation  between the average co-occurrence frequency of subjects and objects within a relation and the quality of linear representations (LREs) formed for that relation. This correlation surpasses the individual correlations with subject frequencies or object frequencies, highlighting the significance of subject-object co-occurrence.
+
+The study focuses uses Linear Relational Embeddings (LREs), which effectively approximate the computations performed by an LLM to predict objects in factual subject-relation-object triplets. This paper builds upon this research by examining how the frequency of subject-object co-occurrences in pretraining data directly impacts the emergence and quality of these LREs
+
+The paper introduces a promising technique for analyzing the pretraining data of closed-source models by leveraging the connection between linearity and frequency.
+
+### Weaknesses
+The paper presents valuable findings, however they should provide some discussion along the following directions:
+
+(a) The paper primarily focuses on Linear Relational Embeddings (LREs) as a representative class of linear representations in LLMs. However, LLMs might employ various other forms of linear or non-linear structures to encode information. This focus on LREs could limit the generalizability of the findings to other types of representations. Is there any strong hypothesis to strict to LREs?
+
+(b) While the study demonstrates that LRE features can be used to predict the frequencies of individual terms with reasonable accuracy, predicting the frequency of subject-object co-occurrences is challenging. The regression models achieve only marginal improvements over baseline performance in this task.  Integrating additional features might be helpful here.
+
+(c) The study analyzes a set of 25 factual relations from the Relations dataset. However, LLMs are trained on vast and diverse data, encompassing a much wider range of relations and concepts. Expanding the scope of analysis to encompass a broader range of relations would provide a more comprehensive understanding of the role of frequency in shaping LLM representations.
+
+(d)  The paper focuses primarily on the frequency of terms in the pretraining data. However, other factors, such as the context in which terms appear, the syntactic structure of sentences, or the semantic relationships between words, could also influence the formation of linear representations. For example, LLMs are proven to not do well is facts  are stored in templates , as it tend to remember the template and not the facts. The proposed approach may not be applicable in those scenarios.
+
+### Questions
+Refer the previous section.
+
+It would be good if authors can dedicate a section to discuss the potential impact of confounding factors, such as context, syntax, and semantics. Explain why controlling for these factors is challenging in the current study but emphasize the importance of future work to disentangle their effects from the influence of frequency.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper explores the question of why linear structures form in LLMs by investigating the connection between training data frequency and the formation of linear representation, focusing specifically on factual recall relations. The study reveals that (1) the formation of linear representations is strongly correlated with subject-object co-occurrence frequency, and (2) the presence of linear representations can help predict relation frequency. Experiments are conducted using OLMo-1B, oLMo-7B, and GPT-J to validate these findings.
+
+### Strengths
+- Exploring the origin of linear representation is an important question in LM interpretability. This work identifies a correlation between linear representations of factual recall relations and the subj-obj co-occurrence frequency in pretraining.
+- This paper investigates the relationship between few-shot accuracy and the existence of a linear representation.
+- Using the existence of linear representations to predict the frequency of terms in the pretraining corpus is interesting.
+
+### Weaknesses
+- The scope of the work is somewhat limited, as only 25 factual relations are investigated. It is unclear whether the identified correlation is also valid for other relation types. Expanding the analysis to include more factual relations and other types of relations could further enhance the robustness of the findings and potentially offer additional insights.
+- The linear representation seems to be affected by the context in LREs (e.g., four "X plays the Y" examples before the fifth one. Are the findings universally applicable to LLM generation without involving ICL formats?
+
+### Questions
+Please see weakness.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The authors investigate the correlation between linear representations and pre-training data frequency in language models. The work is conducted on recent findings that the linearity of different types of relations varies significantly depending on the specific relationship. Existing work does show that language model exhibit such linear structures, but do not reveal the underlying reason why some relations exhibit such structure while other do not. The main contribution of this work is to empirically draw the correlation between such linear structure and data frequency. It shows that that linear representations for factual recall relations are related to mention frequency and the model size. In addition, more detailed results show that linear representations form at predictable frequency thresholds during training, which allows the prediction of term frequencies in the training data. Finally, the authors release a tool for searching through tokenized text for understanding training data characteristics.
+
+Overall, the findings are insightful for understanding linear representation structures in language models. This empirical study complements existing theoretical evidence on the same subject. It provides a perspective to the problem, which can be among many other factors in driving the formation of linear structures. On the utility side, the findings can be used for understanding training data, which are typically not published for current LLMs.
+
+### Strengths
+A perspective for understanding the reason that some features from LLMs demonstrate linear structures while others do not.
+
+A tool for search through tokenized text to support the understanding of training data.
+
+### Weaknesses
+It provides one perspective to the problem empirically, with a specific set of metrics. While giving useful information, the depth of understanding and the utility domain is constrained mostly to the correlation between term frequency, the model size and the linear structure.
+
+### Questions
+Could there be some theoretical discussion on the training dynamics and the frequency thresholds?
+
+One related work is 
+
+Guangsheng Bao, Zhiyang Teng, and Yue Zhang. 2023. Token-Level Fitting Issues of Seq2seq Models. In Proceedings of the 8th Workshop on Representation Learning for NLP (RepL4NLP) at ACL 2023. Toronto, Canada from July 9th to July 14th, 2023.
+
+which also discusses the correlation between term frequencies and model accuracy.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/FHQDCQFD8y.md

@@ -0,0 +1,146 @@
+## Human Reviewer 1
+
+### Summary
+To solve the issue that existing EEG interpretability researches fail to fully utilize raw signals and lack extensibility to other Deep Learning (DL) models, this paper proposes a novel framework, Grad-TopoCAM, to enhances interpretability in DL models for EEG decoding adaptively. Grad-TopoCAM has been validated across eight different DL models and four publicly available datasets, with the salient brain features aligning with established findings in cognitive neuroscience.
+
+### Strengths
+The motivation, enhancing interpretability in DL models for EEG decoding adaptively, is strong and interesting.
+
+The proposed methods, Grad-TopoCAM, can generate visualizations of salient brain region features from DL models without requiring modifications to the architecture or retraining.
+
+### Weaknesses
+The reviewer has some concerns about the technical contributions of this paper. The proposed method is very simple. CAM is a highly classical method that has been thoroughly explored in other fields. This paper merely extends its application to the visualization of EEG brain region features, with limited technical innovation.
+
+The experimental setup is not clearly delineated; for instance, the hyperparameters for training and testing each model are not thoroughly detailed, and the dataset partitioning method is not explicitly described.
+
+
+The writing of this paper has significant room for improvement. Some unnecessary section titles, such as Acknowledgments and Appendices, should not be included. The tables are not aesthetically pleasing; why not omit the percentage sign (%)? The displayed brain region topology maps have low resolution, making the content difficult to discern. Why not use vector graphics to render the brain region topology maps?
+
+### Questions
+Please see the weakness.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces Grad-TopoCAM, a novel method designed for visualizing brain region activation in EEG decoding using gradient-based localization. The primary goal is to enhance the interpretability of deep learning models applied to EEG data by directly mapping feature maps generated by these models to specific brain regions. However, the effectiveness of this visualization method has not been thoroughly validated. The experiments presented largely focus on evaluating the performance of different EEG decoders, resulting in unclear contributions from this research.
+
+### Strengths
+Grad-TopoCAM represents a significant effort to improve the interpretability of deep learning models in the context of EEG data.
+
+The method's approach to directly mapping feature maps to brain regions offers a novel perspective on understanding model outputs and could facilitate advancements in neurotechnology research.
+
+### Weaknesses
+The validation of the effectiveness of the visualization method is insufficiently addressed, limiting the overall impact of the research.
+
+The experiments mainly assess the performance of various EEG decoders without establishing the unique contributions of Grad-TopoCAM.
+
+A comparison with established post-hoc explanation techniques, such as Grad-CAM and SmoothGrad, is lacking, which would help contextualize Grad-TopoCAM's performance and effectiveness.
+
+### Questions
+What specific metrics will be used to evaluate the effectiveness of Grad-TopoCAM compared to existing visualization techniques?
+
+How do the authors plan to quantify the significance of the feature attributes revealed by Grad-TopoCAM in relation to classification accuracy?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper proposed Grad-TopoCAM, which is an explainable AI method to identify and visualize brain regions that significantly influence decoding outcomes. The method was evaluated on multiple EEG datasets and provided with visualizations.
+
+### Strengths
+1. The paper attempted to address the explainability issues for EEG deep learning research which is a key gap in the field.
+2. The paper has good structure and clarity of writing in general.
+3. The figures are informative and clear.
+
+### Weaknesses
+1. In the related work section, it is unclear why 'employing a two-dimensional convolutional structure' is a limitation as this is a common approach for most of the works in the EEG field.
+2. The key weakness is that there is no comparison to the state-of-the-art or any other work in the field. For a typical explainable AI work, there should be comparison with other existing explainability methods and demonstrate how the proposed work is superior. It is unclear how the performance of the proposed method really differ from the regular Grad-CAM in general. 
+Some of the baselines for comparison can be considered: LIME, Grad-CAM, GNN-Explainer, Attention-based methods etc.
+3. In section 4.3 discussion of dataset III and IV. it is unclear how the patterns of brain activations are 'similar' when the topography plots are clearly different. Even if the topography plots are similar, the Chinese characters and English words have different meaning so it is not possible to justify there is common cognitive processing mechanisms between the two languages in this case.
+4. In section 5.2, the channel selection results have high variations, the 20% increase for subject 6 is not generalizable to other subjects or datasets and there is no significance measurement for the effect of channel selection. It is unclear how effective or ineffective the channel selection method is.
+5. There is a lack of ablation studies to prove the importance of those channels-identified. For instance, if those channels were removed, there should be a significant drop of classification performance.
+
+### Questions
+see weaknesses
+
+### Soundness
+1
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes the Grad-TopoCAM for enhancing the interpretability of deep learning-based EEG decoding models. It maps the gradients of feature maps to specific brain regions and facilitates channel selection across different EEG tasks. The proposed method is validated on eight DL models and four public datasets. Experimental results demonstrate its effectiveness.
+
+### Strengths
+1. The proposed model can be integrated into different EEG decoding models to enhance their interpretability. It is a universal interpretability and visualization method.
+2. The proposed model has been validated on various DL methods and datasets.
+
+### Weaknesses
+1. Grad-CAM has been widely adopted for feature visualization including for EEG decoding models. The contributions of the proposed method compared with other visualization methods are not clear.
+2. The proposed Grad-TopoCAM is employed for visualization analysis on the model with the highest accuracy for each subject. However, it’s noticed that the visualized features can be very different across subjects. In addition to the individual variability, are the learned features related to the models? Is it a fair comparison for the features learned by different models?
+3. Although visualization is important for interpreting results, the proposed method does not enhance decoding performance or provide unique neuroscience insights. The authors may consider either improving its methodological novelty or deepening its neuroscience contributions.
+
+### Questions
+Please refer to the weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4

human_reviews/GpdO9r73xT.md

@@ -0,0 +1,176 @@
+## Human Reviewer 1
+
+### Summary
+This paper explores the influence of initial noise in the text-to-image generation tasks. They identify specific regions within the initial noise termed as trigger patches, which contribute to the positional bias in the generated images. In order to automatically discover the trigger patches, the authors train a detector. Next, the authors explore several special and interesting attributes of trigger patches, and discover the differences between the random patch within the initial noise and trigger patch. At last, based on the discoveries, the authors design two speical applications for trigger patch.
+
+### Strengths
+Thanks for the authors` efforts. I find this paper very interesting, and the experimental section is quite comprehensive. It has provided me with many insights and considerations.
+
+1. The paper is well-organized, and the presentation of this paper is very good.
+
+2. The motivation of this paper is clear, and it first explores the influence of positional bias in the initial noise. I think the experiments are quite comprehensive. I think this paper is valuable.
+
+3. Some experiment results are very attractive, and they have sparked a lot of my thoughts and reflections.
+
+### Weaknesses
+Although, from my perspective, this paper is good, I think it still has several drawbacks:
+
+Presentation:
+
+1. In line 231, "As shown in the Figure", missing reference to the Figure. In line 293, missing (2) before "Can noises with multiple ...". In line 417, ":" is missing after "Control", and the first word after ":"， should be lowercase.  In line 1070, "Please refer to the Fig F.2 for visualized results.", I can not find the Fig.F2.
+
+2. In Figure 13, the prompt is not consistent with the provided images.
+
+3. All intuitive images in the main paper only contain five concepts. Can the authors provide more visualization results?
+
+Methods and Experiments:
+
+1. For the evaluation metrics, the authors only utilize the CLIP. I think it is crucial to utilize some human preference models like PickScore, HPSv2 to evaluate the quality of the generated images.
+
+2. The dataset constructed by the authors only contains five single concepts, consisting of 500K samples. How do the authors ensure that whether the trigger patches exist beyond these five prompts, and the detector trained in this manner can recognize other concepts? Besides, I think the construction of the dataset is resource-consuming.
+
+3. The authors claim that low entropy ISR will be higher; however, Table 6 shows that certain mid-entropy ISRs are also quite high. How should the authors explain this phenomenon?
+
+4. All of the main results are conducted only in Stable Diffusion v2. Can the authors provide the existence of the trigger patches in other diffusion models?
+
+5. Experiment results in Table 5 reveal “Random” can surpass "Initno" and "Attend-and-Excite". This phenomenon is very strange because as far as I know, both of them optimize the noise during the reverse process. They should be higher than "Random"?
+
+6. The experiment results in Generalization part raise my concerns. From the results presented in Table 8, it seems that trigger patch injection is only effective with deterministic samplers, showing no impact on stochastic samplers. The authors also default to using a deterministic sampler. Besides, based on the results in Table 10, why is the gap reduction so pronounced across other models? If the initial noise inherently carries positional bias, then its performance across different diffusion models should be similar. However, the experimental results seem to contradict the notion that the trigger patch is universal.
+
+### Questions
+See the weaknesses.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The work finds out that there are some locations in the image frame where the diffusion often gets biased. These locations are marked as trigger patches. An object detector is trained to detect these patches before the sampling process. After that, rejection sampling is used to reject samples with trigger patches to obtain the better generative image diversity.
+
+### Strengths
+Strength:
+1. The paper provides an interesting point of view on diffusion sampling
+2. The new sampling methods help to increase the diversity
+
+### Weaknesses
+Weaknesses:
+1. According to Figure 1 and Figure 3, the trigger patches are always in the right part of the image and only applicable to some objects. The whole paper also only focuses on some round objects like balls and tennis balls. This results in the question of generalization of the work. Whether or not this phenomenon will happen across datasets or with different objects?
+2. The experimental results do not show the quantitative values to measure the output quality. There could be a trade-off between diversity and quality according to the new rejection sampling. 
+3. The current diversity result is not standard in generative measures such as FID/Recall.
+
+### Questions
+1. Please test on wider range of objects as well as datasets or find the trigger patches on different locations of the dataset instead of the right edge of the photo.
+2. Please provide the qualitative values for image quality and observe the trade-off between diversity and image quality from the proposed rejection sampling method.
+3. Please provide more comprehensive evaluations on diversity such as FID/Recall.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This article introduces the intriguing concept of trigger patch noise. The authors first demonstrate the existence of trigger patch noise, then propose a method using neural networks to detect this noise in initial noise. They experimentally verify the unique characteristics of trigger patch noise and explore its potential applications, such as controlling object generation position by embedding trigger patch noise into the initial noise.
+
+### Strengths
+1.The insight is interesting.
+
+2.The authors not only demonstrated the existence of trigger patch noise through experiments but also provided a method for its detection. Additionally, the authors thoroughly explored the effects of multiple trigger patch noises, examined the relationship between trigger patch noise and prompt guidance, and investigated the generalization capabilities of trigger patch noise.
+
+3.I believe that the proposed concept of trigger patch noise has the potential to advance research and development in the field of controllable generation.
+
+### Weaknesses
+1.The experiment was conducted with only five object categories. Can trigger patch noise be generalized to a broader range of object categories? This could provide insights into its robustness across diverse applications.
+
+2.Does trigger patch noise commonly exist within any randomly sampled Gaussian noise? Additionally, if multiple trigger patch noises are present in a single Gaussian noise, what criteria should be used to select the most effective trigger patch noise?
+
+3.In the training process of the trigger patch noise detector, how do you define the ground truth for trigger patch noise? Providing a detailed algorithm or clarification of this aspect would be valuable. Furthermore, does every one of the 17,000 noises in the training set contain trigger patch noise?
+
+4.There seems to be an impact on image realism when inserting trigger patch noise. How significant is this effect, and are there any methods to mitigate it? (The BLIP-Text results in Table 11)
+
+5.The detection accuracy of trigger patch noise reported in the paper appears somewhat limited. Are there plans or suggestions for enhancing detection performance?
+
+6.What would be the effect of inserting multiple (e.g., three) trigger patch noises? Additionally, when aiming to control the size of the generated object, does resizing the trigger patch noise (e.g., scaling) still yield the intended effect?
+
+7.Can the trigger patch noise work for DiT models?
+
+### Questions
+See weakness.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+In the paper, the authors introduce the concept of "trigger patches," which are specific regions within the initial noise image that determine the positional information in object generation. They present a method for localizing these trigger patches, provide analyses to validate their correctness, and demonstrate two applications.
+
+### Strengths
+1. The concept of "trigger patches" in the diffusion generation process is both interesting and useful, as it enhances the controllability of the generation.
+2. The experiments and analyses are thorough and comprehensive.
+3. The writing is clear and intuitive.
+
+### Weaknesses
+1. In the evaluation of the second application (Sec 5.2), only "left" and "right" are considered. Is it possible to include more fine-grained positional information, such as "left-down"?
+2. In the introduction, it is mentioned that "moving/removing trigger patches can achieve certain image editing effects." Including some image editing results in the experimental section would support this claim.
+3. In cases where "trigger patches" and prompts are aligned, the positional generation accuracy is 63.5%. What factors contribute to the failures in the remaining cases? Can the authors provide a failure analysis?
+
+### Questions
+Please address the questions / concerns in the weaknesses section.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3

human_reviews/HHKboqbkec.md

@@ -0,0 +1,172 @@
+## Human Reviewer 1
+
+### Summary
+The paper introduces a novel approach to utilizing large-scale models for Bayesian Theory of Mind (ToM) reasoning. Rather than fine-tuning large models directly, the authors propose fine-tuning smaller models and leveraging the difference between the fine-tuned and naive versions of these small models to approximate the impact of fine-tuning on the larger model. This method is rigorously evaluated on the MMToM-QA dataset, demonstrating its effectiveness in handling complex ToM tasks.
+
+In the interest of transparency, I would like to note that this paper is somewhat outside my area of expertise, and as such, my confidence in assessing it is limited. However, after discussing this with the Area Chair, we agreed that I would proceed with the review and focus on evaluating aspects I can assess.
+
+### Strengths
+The proposed method is both simple and robust, demonstrating a well-grounded approach. The experiments are rigorously conducted, and the results reflect strong performance, effectively supporting the validity of the approach.
+
+### Weaknesses
+A weakness I noted is that I found the paper challenging to understand. While this is possibly due to my limited expertise in this specific area, I looked at recent related works (BIP-ALM and SimToM) and found them notably more accessible. This suggests that, although my background may play a role, the paper could benefit from clearer explanations or added context to help readers better grasp its motivation and contribution - especially those outside the immediate field.
+
+### Questions
+- The general idea of approximating the changes that fine-tuning a large model would have by fine-tuning a smaller one and approximating the changes by reweighing using a ratio seems very general. Could this exact methodology be applied to other types of problems? (Or was it already used? - the relation to the works using reweighing, mentioned in the related work, is unclear to me. )
+
+- Is there a fundamental reason why fine-tuning models for the kind of ToM tasks considered is harder / more computationally expensive than fine-tuning for other tasks?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+1
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper builds upon the MMToM-QA benchmark by introducing a method to align the general, amortized policy of a large LM (as originally used in MMToM-QA work) to the ToM tasks at hand using a post-trained expert small LM. The main contribution is a reweighting mechanism that integrates the extensive world knowledge from the large LLM with the ToM-specific policy from the small model. An extensive analysis on the improvement in performance as a result of this on the MMToM-QA benchmark. The paper thus effectively propose a way to improve upon the MMToM-QA method for better/more accurate Bayesian likelihood estimation of the goals and beliefs, that also now leverage the specialized knowledge corresponding to a particular ToM tasks in a scalable manner.
+
+### Strengths
+These are the strengths of the paper in my opinion:
+
+1. Improves upon the Bayesian Inverse Planning Accelerated by Language Models, by making them more adaptive to ToM tasks at hand using a post-training process.
+
+2. The proposed post-training method is more scalable and the integration of this expert post-trained knowledge to guide the large LM seems principled.
+
+3. Extensive analysis and ablation studied on the MMToM-QA benchmark, effectively illustrating the methods adaptability and efficiency.
+
+### Weaknesses
+These are the weakness of the paper in my opinion:
+
+1. The paper is very poorly written. Unless and until you go back and read the details from the appendix of MMToM-QA [1] method, the explanations are incomplete and hard to read. The paper would benefit from an appendix section and can be more upfront about the fact that this is a direct extension to the the MMToM-QA paper. For example, see the question 2, such implementation details are not discussed anywhere.
+
+2. The paper is fairly incrimental as far as the methods used are concerned. The ToM inference is directly an extension of previous work, and so is the post-training method (eg: takes inspiration from the post-training literature, that uses a reweighting ratio between an expert and non-expert and subsequent normalizations [2][3]).
+
+### Questions
+1. Could you please explain why such a reweighting scheme between an expert and non-expert is a principled mechanism for policy adjustments/redirection in Equation 6 ? Is it related to some well-established theories in Bayesian / Probabilistic modelling literature (for example say Importance Sampling, or minimizing KL divergence between policies etc) ?? Expanding on this theoretical foundation would enhance the paper’s contribution and clarify the underlying principles.
+
+2. How is the agent's belief distribution b(s) modelled? How is the belief evolution modelled $P\left(b_1^t \mid \hat{b}^{t-1}, s^t\right)$ ?? These details are missing. The paper would benefit from a thorough rewriting to make it more accessible to the reader.
+
+3. How much time does this fine-tuning process take (specify the compute resources used etc) and what is the size of the pool $
+\mathcal{D} = \left(s_i, b_i, g_i, a_i\right) _{i=1}^N$ in terms of N, used for optimizing eq 5? 
+
+
+References
+
+1. MMToM-QA: Multimodal theory of mind question answering https://arxiv.org/pdf/2401.08743
+2. DExperts: Decoding-time controlled text generation with experts and anti-experts. https://arxiv.org/abs/2105.03023
+3. An Emulator for Fine-Tuning Large Language Models using Small Language Models. https://arxiv.org/pdf/2310.12962
+
+
+
+--------
+### **Post Discussion Update**
+
+**I increase the score to a 5 for incorporating several suggestions, especially for a more accessible writing style post rebuttal. The effort of introducing the new theorem to justify the reweighting mechanism (under some strong initial assumptions) is also appreciated. Methodologically it's an incremental work extensively based on a recent paper MMToM-QA, but now without the need to explicitly post train the large model. Because of the limited novelty, I unfortunately cannot recommend a strong accept. However, I won't oppose if the other reviewers argue for an accept.**
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper presents a method for scaling Theory of Mind (ToM) capabilities in AI systems using a weak-to-strong Bayesian reasoning framework. The core approach combines Bayesian state inference with language models, where a smaller post-trained language model (4-8B parameters) guides the reasoning process of a larger model (up to 405B parameters) during test time1. The method works by transferring ToM-specific behaviors from the post-trained small language model to influence the latent reasoning of the larger language model. This approach enables the system to leverage both the extensive world knowledge of large language models and the ToM-specific behaviors learned through post-training, while avoiding the computational costs typically associated with post-training large models
+
+### Strengths
+- The weak-to-strong Bayesian framework provides a practical solution to scale Theory of Mind capabilities without post-training large models, addressing a key efficiency challenge in the field.
+- The technical approach cleanly integrates Bayesian state inference with language models, providing a principled and effective foundation for ToM reasoning.
+- The evaluation framework systematically tests both scalability and generalization, with clear ablations demonstrating the contribution of each component.
+- The paper is well written, with a clear motivation, methodology, experimental section, and supporting figures.
+
+### Weaknesses
+- The paper doesn't adequately address the apparent contradiction: recent research shows smaller models have fundamental reasoning gaps, yet the method relies on a small model to guide larger model reasoning. This raises questions about whether the smaller model can effectively decompose problems it may struggle to reason about itself. 
+
+- The evaluation methodology could be strengthened by adopting a variance analysis framework similar to GSM-Symbolic (Mirzadeh et al., 2024, https://arxiv.org/pdf/2410.05229), which would help quantify how the model's Theory of Mind capabilities vary across different phrasings of the same underlying task.
+
+### Questions
+- Would you classify tom beliefs as a form of reasoning? If so, what makes a smaller model capable of classifying tom beliefs?
+- Could the authors clarify if the primary mechanism of the weak-to-strong framework is to align the larger model's distribution with ToM-specific beliefs and task structure, while preserving its general reasoning capabilities, rather than directly transferring reasoning abilities from the smaller model? If so, this would suggest the smaller model acts more as a task-specific prior that guides the larger model's attention and beliefs, rather than as a direct source of reasoning. This interpretation could help explain the framework's effectiveness despite the typically limited reasoning capabilities of smaller models.
+- Figure 1 is confusing, its not clear from the figure exactly how the smaller post trained lm conrols the larger lm. Why is this model denoted  'extreme'?
+- How how is model performance impacted when the post trained lm is itself large (e.g. 70B class)?
+- As stated in weakness 2, how does performance degrade when the task phrasing is edited?
+
+
+Notes
+- there are no line numbers
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper presents a scalable approach to enhance machine-based ToM, which is crucial for understanding and predicting human-like mental states in multimodal environments. The authors propose a weak-to-strong Bayesian reasoning framework that leverages the strengths of large language models to improve ToM inference without the computational costs associated with extensive post-training.
+
+### Strengths
+1. The author's approach seems to be novel
+2. I like the experimentation, did show the scaling trend.
+
+### Weaknesses
+1. Presentation could be better. For instance, Fig 1, it contains too much information.
+2. The major concern here is that everything is done on Lora. What happen if you do full-scale training? How much performance improvement do you expect to gain?
+3. Didn't test smaller model such as phi-3b, gemma-2b etc. Maybe you can run full fine-tuning on smaller model.
+
+### Questions
+1. My major question is, how much improvement do you expect to gain if you run full-scale training?
+2. Would be nice to see if you can test this on a smaller model, GPT2, or gemma-2b and see if there is larger performance gain. 
+3. The performance difference between llama2 vs llama3 vs llama3.1 isn't that much, is it because pre-training (or the quality of the pre-trained model) matter less in this context? Do you have theory/explaination why this happens?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+2

human_reviews/IfPfUHRowT.md

@@ -0,0 +1,163 @@
+## Human Reviewer 1
+
+### Summary
+This paper presents a novel foundation model for sinogram inpainting based on the latent diffusion model (LDM). By considering certain physical characteristics of CT acquisition and sinograms, they designed an innovative physics-informed loss function that significantly enhances the autoencoder's training performance. Additionally, they introduced a sinogram blending technique that balances content and style perception to improve the synthesis quality between predicted and real sinograms. Experimental results demonstrate that this method outperforms current state-of-the-art approaches in sinogram inpainting, highlighting its potential as a foundational model for CT applications.
+
+### Strengths
+By incorporating the physical characteristics of CT acquisition and sinograms into the design of the loss function, this approach effectively combines domain knowledge with advanced techniques, enabling more efficient learning for domain-specific tasks.
+
+### Weaknesses
+- The structure of this paper is highly disorganized. For instance, in lines 171–176, after introducing various works on sinogram inpainting, the authors abruptly shift to discussing image blending, which may confuse readers. Furthermore, the authors seem unfamiliar with the proper use of `\citet` and `\citep` commands, as nearly all citations fail to meet academic writing standards, particularly in Section 2, where incorrect citation formatting severely disrupts the reading experience. In Section 4.4, almost all citations are wrong, with many references not corresponding to their intended sources. The authors should thoroughly review and correct all citations with careful attention to detail.
+- Figure 9 is severely blurred, making it difficult to discern the comparative effectiveness of the proposed method against other approaches. Additionally, the paper lacks a quantitative performance comparison with the baseline methods. Although the primary task is sinogram inpainting, in practical applications, the focus is ultimately on the quality of the final CT images reconstructed from the inpainted sinograms. This aspect is missing from the current work. The authors should consider including both quantitative and qualitative results of the CT images reconstructed from sinograms completed using the proposed method.
+
+### Questions
+- In line 206, the phrase “perceptual loss term, $L_P$ in addition” seems to have incorrect comma placement. It should be written as “perceptual loss term $L_P$, in addition,” correct?
+- In line 263, the authors state, “we also need a ramp filtering operation for noise removal in the image.” This is inaccurate; it should be clarified that the purpose of the ramp filtering operation is to remove blurring effects.
+- In line 367, the authors mention, “we fine-tune the model with fewer data,” and that “this fine-tuning approach does not require a high computational overload and can be implemented on a compute node with limited GPU resources.” However, they also mention using 25,000 real-world training samples for fine-tuning, which is not a small dataset. This raises the question of how it can be feasible “with limited GPU resources.”
+
+### Soundness
+3
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper proposed a model which integrate the Latent Diffusion Model (LDM) with physics-based domain knowledge. A set of loss functions were designed. The paper proposed a blending algorithm to improve the accuracy of inpainting task. The proposed method is work for simulated parallel projection data (not real-world data).
+
+### Strengths
+(1)	domain specific physics knowledge of CT image formation for inpainting sinograms taking into account both measurement and reconstruction domains.
+(2)	Recover sinogram with different masks and different sampling ratio. 
+(3)	Suitable for Sparse data acquisition and Missing Wedge Problem.
+
+### Weaknesses
+(1)	The loss is complex and too many parameters. Ablation of every part is necessary.
+(2)	The paper uses the simulated projection data other than real word data.
+(3)	The proposed method is only work with Parallel beam projection geometry with is xxxxx
+in real application.
+
+(4)	The downstream task of the method is image reconstruction. Comparison with reconstruction method for sparse view reconstruction and limit view reconstruction is necessary, such as dual domain reconstruction.
+[1] W. Wu, D. Hu, C. Niu, H. Yu, V. Vardhanabhuti and G. Wang, "DRONE: Dual-Domain Residual-based Optimization NEtwork for Sparse-View CT Reconstruction," in IEEE Transactions on Medical Imaging, vol. 40, no. 11, pp. 3002-3014, Nov. 2021, doi: 10.1109/TMI.2021.3078067
+[2] Ding, Q., Ji, H., Gao, H., Zhang, X. (2021). Learnable Multi-scale Fourier Interpolation for Sparse View CT Image Reconstruction.
+(5) The literature review is limited, many CT reconstruction works, such as iterative reconstruction and deep learning reconstruction (image domain, unrolling (ADMM-Net), and plug-&play method) are not given
+
+### Questions
+1 The training and inference time comparison is necessary for sampling is time-consuming.
+2 Notation is not defined clearly, e.g. equation (1) sg. and z_qn never used.
+3 Can the method extend to 2D fan-beam and 3D Cone beam reconstruction
+4 Visual comparison of reconstructed image from recovered sinogram also needed for that a minor error of sinogram may lead to streaky artifacts in image.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper deals with the problem of sparse sinograms in CT. The authors combine an autoencoder with a latent diffusion model which incorporates physics knowledge in the loss functions. This stabilizes the training process of the autoencoder. Furthermore, a blending algorithm is applied such that the output of the diffusion model also aligns with the measurement data.
+
+### Strengths
+* Including physics-domain knowledge is an essential task in machine learning.
+* Sparse sinograms are problems in practice to reduce the dose and/or the acquisition time.
+* The method was tested on a real-world dataset.
+* The paper has a good structure, explaining the individual components.
+
+### Weaknesses
+* The proposed method seems very complex. There are many different sub-modules: autoencoder, diffusion model, blending.
+* Training an autoencoder is hard.
+* There are many hyperparameters to adjust in the different loss functions.
+* The blending technique is not suitable for lower mask ratios.
+
+### Questions
+* Do we really need such a complex setting? Isn’t it possible to use a simpler model like a diffusion model and include all the losses there?
+* Is this approach end-to-end trainable?
+* What are phantom features?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes a method that integrates a Latent Diffusion Model (LDM) with physics-based loss functions to address missing data issues in CT imaging by inpainting sinograms. The paper demonstrates an approach by combining machine learning techniques with CT-specific physical properties. However, the manuscript has shortcomings regarding organization, precision, and detail.
+
+### Strengths
+This paper presents an approach by integrating a Latent Diffusion Model (LDM) with physics-based loss functions for CT sinogram inpainting, demonstrating originality in combining generative AI with domain-specific knowledge. The work introduces unique physics-driven loss functions that enhance the stability and accuracy of the model.
+
+### Weaknesses
+The manuscript is challenging to read due to disorganized logic, unprofessional expressions, and vague descriptions, making it hard to follow the motivation and methodology.
+
+### Questions
+1.     This manuscript is challenging to read, with disorganized logic and non-standard, unprofessional expressions:
+	•	The paper only lists existing methods without clearly identifying their limitations or explaining the motivation for the proposed approach.
+	•	Expressions such as “During a CT experiment” and “CT experimental time” are imprecise and lack professionalism.
+	•	The statement, “The object is rotated stepwise around a central axis,” is misleading; typically, it is the source and detector that rotate, not the object itself.
+	•	Phrases like “Combining several such projections, a specialized algorithm” are vague and difficult to understand. Please clarify.
+	•	The sentence, “it can also lead to morphological deformation in the sample due to extended radiation exposure,” seems to imply that deformation results from patient motion over extended scan times.
+	•	The phrases “In one approach,” “In another approach,” and “In a third approach” might be referring to sparse-view CT, limited-angle CT, etc. However, other forms of incomplete projection data also exist, which could be covered or referenced in a comprehensive review article [1].
+
+[1] Wang T, Xia W, Lu J, et al. A review of deep learning ct reconstruction from incomplete projection data. IEEE Transactions on Radiation and Plasma Medical Sciences, 2023.
+
+2.     The so-called “novel blending algorithm” proposed in this paper is, in fact, a common practice in CT reconstruction from incomplete measurements.
+
+3.     The term “realistic simulated data” raises ambiguity regarding the authenticity of the experimental data; it is unclear whether the data are from real experiments or simulations. Although the paper emphasizes the importance of physics-based loss functions, there is no detailed description of imaging parameters. Additionally, the use of “re-project” suggests that the projection data may not be genuine, but rather re-generated, which further questions the authenticity of the data used.
+
+4.     Since the method uses FBP with a reconstruction loss, why not present the reconstructed CT images? This would provide a more intuitive understanding of the results.
+
+5.     The resolution of Figure 9 is too low, making it difficult to read the results.
+
+6.     The paper involves many hyperparameters but lacks any analysis or justification regarding their selection. Given the importance of these parameters to the model’s performance, a sensitivity analysis or at least a discussion on hyperparameter tuning would enhance the reliability and reproducibility of the results.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+1
+
+### Confidence
+5

human_reviews/KijslFbfOL.md

@@ -0,0 +1,169 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces SIIHPC, an incomplete multi-view clustering method that bypasses traditional missing view recovery by directly imputing similarities. SIIHPC replaces single prototype construction with hybrid-group prototypes, allowing each view to capture unique features within a unified framework. This framework is optimized through a well-designed iterative algorithm with convergence guarantees, ensuring robust and effective clustering.
+
+### Strengths
+The paper presents a novel approach SIIHPC to incomplete multi-view clustering through similarity-level imputation, which diverges from conventional methods that focus on reconstructing missing samples or features. This approach serves as an efficient alternative to traditional recovery-based techniques. By addressing incomplete multi-view clustering without explicit missing view imputation, SIIHPC reduces computational overhead and allows clustering methods to function effectively without relying on fully complete data structures.
+
+A key innovation of the paper is the introduction of hybrid-group prototype construction, which replaces single prototypes with a more flexible model that enables each view to capture unique features. This enhances the model’s adaptability to heterogeneous data views, improving clustering performance across diverse datasets and scenarios.
+
+Extensive experimental evaluations on multiple datasets with varying missing ratios highlight the robustness and effectiveness of SIIHPC. The inclusion of ablation studies further enriches the assessment, providing valuable insights into the contribution of each component to the model's overall performance and validating the framework’s design choices.
+
+### Weaknesses
+The terminology introduced in the paper is unclear; for instance, terms like T-PBL, SLI, IVHGP, and MSVSG are not properly explained in the captions of Figure 1, making it difficult for readers to follow the framework without additional context.
+
+The overall process of the proposed clustering algorithm lacks clarity. Specifically, the final steps for obtaining clustering results and the initialization settings for relevant parameters should be clearly defined to improve the reader's understanding of the algorithm’s workflow.
+
+While the paper provides strong experimental results, the interpretability of the clustering outputs and the practical implications of hybrid-group prototypes could be explored further. It would be helpful for the authors to discuss how the clustering results or prototype structures might be interpreted in real-world applications, providing insights into the method's broader utility.
+
+### Questions
+Can the authors clarify the terms used in Figure 1, such as T-PBL, SLI, IVHGP, and MSVSG, and ensure that these terms are defined either in the figure caption or in the text to improve readability?
+
+Could the authors provide a more detailed explanation of the overall process of the clustering algorithm, particularly the final steps for obtaining clustering results and the initial parameter settings?
+
+Given that the method relies on similarity-level imputation rather than missing view recovery, how sensitive is the model's performance to the quality of the imputed similarities? Could the authors discuss any potential limitations in scenarios with high levels of data incompleteness?
+
+To better assess the stability and reliability of SIIHPC’s performance, could the authors provide standard deviation metrics for their experimental results? Including this information would help illustrate the consistency of the method across different runs and give a clearer view of its robustness under varying data conditions.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+In this manuscript, the authors address the challenge of incompleteness due to missing samples by restoring the missing data at the similarity level. They propose a method that employs a hybrid approach to extract data representations, utilizing multiple quantities of prototypes for each individual view rather than relying on a single quantity across all views. This approach effectively resolves the misalignment of observed samples and incorporates potentially useful information from the missing samples into the bipartition similarity. Additionally, the authors balance contributions from different views while defining overall similarity based on the intrinsic characteristics of each view. These goals are effectively realized within a cohesive learning framework. The proposed method's effectiveness is convincingly demonstrated through experimental results on six public datasets with varying missing ratios.
+
+### Strengths
+1. The manuscript presents a logically structured approach, with Figure 1 providing a clear and intuitive overview of the framework.
+
+2.  The authors utilize a variety of missing ratios and metrics in their experiments to evaluate clustering performance effectively.
+
+3. The proposed SIIHPC demonstrates favorable memory and computational efficiency.
+
+### Weaknesses
+1. There is an issue with the manuscript's organization where the title of Section 5.5 is found to be identical to that of Section G in the Appendix.
+
+2.  Experiments exhibit that some comparison methods do not function properly on VGGFACEHUND, YOUTUBEFACE and FASHMINST. However, there is a lack of theoretical explanations to substantiate these observations.
+
+3. The space complexity of Algorithm 2 is not described enough.  
+
+4. The manuscript lacks definitions or explanations for certain notations used throughout the text. For instance, the notation $()_{+}$ in (11), N/A in Table 2 and Table 3  are not clarified.
+
+### Questions
+1. Is the auxiliary function $g$ influenced by the percentage of missing samples?  Both $\mathbf{L} _{v}$ and $\mathbf{P} _{v,s}$ incorporate the missing percentage, which raises the question of how this affects the function.
+
+2. According to the dataset descriptions in Table 1, FASHMINST has more samples than YOUTUBEFACE. However, the results in Table 3 indicate that SIIHPC incurs a higher memory overhead on YOUTUBEFACE compared to FASHMINST. What factors contribute to this discrepancy?
+
+3. As indicated in Table 3, GSRIMC is limited to operating on only two small-sized datasets and exhibits higher consumption of running time and memory overhead. In comparison with SIIHPC, what is the theoretical complexity of GSRIMC?
+
+4. Beyond the comparison with graph or tensor-based methods, how does SIIHPC fare when compared to deep learning approaches?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes a new incomplete multi-view clustering method, which reconstructs similarity relationships through the original sample form to capture the intrinsic information of all views. By relaxing traditional non-negative constraints, the method enables more flexible similarity characterizations. Additionally, a hybrid prototype set is introduced for each view, facilitating view-specific feature extraction and contributing to a comprehensive consensus graph. An innovative auxiliary function with monotonic properties is designed to solve the optimization problem effectively. Experimental results on various incomplete multi-view datasets demonstrate the robust clustering performance of SIIHPC.
+
+### Strengths
+Paper Strengths
+1.The paper is well-organized and easy to follow. The intuition is clearly discussed in the Introduction.
+2.The proposed method performs exhaustive experimental analyses, comparing a wide array of clustering algorithms, thereby strengthening the credibility of the results.
+
+### Weaknesses
+Paper Weaknesses
+1.The process for constructing the consensus graph $G$ is not clearly explained, particularly in relation to $G_s$.
+2.Does the dataset include both images and text? Please specify the composition of each view within the dataset.
+3.It would be helpful to discuss how Similarity Level Imputation and Hybrid-group Prototypes specifically contribute to improving clustering performance in Ablation section.
+
+### Questions
+1.The experimental setup lacks sufficient detail, particularly regarding the range and selection criteria of hyperparameters.
+2.The Symbol Summary in the Appendix does not cover all symbols used throughout the paper, which may hinder clarity for readers. It is recommended to include all symbols and provide detailed explanations for each variable to improve readability and understanding of the methodology.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper presents an incomplete multi-view clustering (IMVC) algorithm, SIIHPC. To utilize the latent valuable information of incomplete samples, it successfully separates out the observed parts via bipartition learning transformation and relaxes traditional non-negative constraints via sample regularization. It introduces the learnable consensus graphs to provide unified structure, and based on the relationship between consensus graphs and exclusive similarities it imputes the incomplete part at the similarity level. To get rid of the limitations of single prototype quantity, it assigns a group of hybrid prototype quantities for each view, and performs spectral grouping on the resulting multi-scale graphs to generate the cluster labels. Afterwards, it gives a four-step updating scheme to minimize the formulated objective. Overall, the overall organization of this manuscript is clear and easy to follow.
+
+### Strengths
+1. The motivation is good. Incorporating incomplete samples into the construction of bipartite similarity enhances the representation of relationships between different entities.   Diverse prototype quantities facilitate the accurate characterization of different views.
+
+2. The designed auxiliary function is interesting. It equivalently solves one quadratic programming problem with orthogonal constraints and meanwhile is demonstrated to be with monotonic-increasing properties.
+
+### Weaknesses
+1. In the paper, the hyper-parameters $\lambda$ and $\beta$ are adjusted in $[10^{-7}, 10^{-6}, \cdots, 10^{-2} ]$  and $[10^{2}, 10^{3}, \cdots, 10^{7} ]$ respectively. The reasons for setting parameters are not given. 
+
+2. Similar to the analysis about Table 3, some reasons for the sub-optimal results in Table 2 should be added. 
+
+3. Descriptions about the theoretical space overhead should be more in-depth.
+
+### Questions
+1. As descripted, PSIMVC is also based on prototype, what are the differences between the proposed SIIHPC and PSIMVC? Specially, combined with Table 3, PSIMVC can consume less time and/or memory.  
+
+2. Figure 4 show that at each iteration of Algorithm 2, the objective of function $g$ is increasing while at different iterations of Algorithm 2, it doesn't seem like this, such as at the $3$-th, $4$-th, $5$-th iteration respectively. 
+
+3. After transformation, why is the complexity about $\mathbf{H}_{v,s}$ reduced from square to linear?
+
+4. When adopting the prototypes constructed by elicitation sampling, what is the performance of SIIHPC?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+5

human_reviews/LFn7s8yRUF.md

@@ -1,24 +1,5 @@
-# EXPLORING THE IMPACT OF DATA AUGMENTATION ON LOCALIZED PERSONALIZED AI TRAINING WITH LLAMA3 AND LORA
-
-- Decision: Reject
-- Scores: 3, 1, 1, 1
-
-## Abstract
-With the development of personalized AI models, particularly those emulating characters from novels, games, anime, and films, a significant challenge is the scarcity of suitable dialogue data. These works often feature distinctive styles and character dialogues that may not generalize well to everyday conversations. Data augmentation is crucial for enriching these limited datasets, ensuring sufficient data for learning the target character’s tone and linguistic habits. This paper investigates the impact of various data augmentation techniques on personalized AI models in NLP, specifically focusing on models trained using LLaMA3 through Low-Rank Adaptation (LoRA). We employ different data augmentation strategies, including random deletion, synonym replacement, swapping, random insertion, back translation, and paraphrasing. To provide a comprehensive analysis, we apply these techniques across three distinct datasets, each representing different dialogue styles and contexts. By systematically comparing these methods, we demonstrate their influence on model performance and robustness. This study provides valuable insights into the effectiveness of different data augmentation strategies for enhancing the versatility and robustness of personalized AI systems trained with LLaMA3 using LoRA.
-
-## Human Reviews
-
 ## Human Reviewer 1
 
-### Rating
-3
-
-### Rating Number
-3
-
-### Confidence
-4
-
 ### Summary
 The paper investigates the relevance of data augmentation techniques for the training of personalized dialogue agents. It gives an overview of existing text augmentation methods, of existing open-weights LLMs and of low-rank adaptation techniques. It introduces a custom dataset based on a TV series and a video game, and a task consisting of enacting a specific character within these environments.
 
@@ -72,18 +53,15 @@ Regarding the paper presentation:
 ### Contribution
 1
 
----
-
-## Human Reviewer 2
-
 ### Rating
-1
-
-### Rating Number
-1
+3
 
 ### Confidence
-5
+4
+
+---
+
+## Human Reviewer 2
 
 ### Summary
 This paper investigates the effectiveness of various data augmentation techniques for training localized, personalized AI models using LLaMA3, fine-tuned through LoRA. The study employs methods like random deletion, synonym replacement, backtranslation, and paraphrasing to enhance the quality and robustness of personalized dialogue models trained on character-specific data. Experiments are conducted on two datasets with distinct linguistic styles, aiming to provide a detailed comparison of augmentation techniques in low-resource, domain-specific training scenarios.
@@ -120,19 +98,16 @@ N/A
 ### Contribution
 2
 
----
-
-## Human Reviewer 3
-
 ### Rating
 1
 
-### Rating Number
-1
-
 ### Confidence
 5
 
+---
+
+## Human Reviewer 3
+
 ### Summary
 The paper explores different data augmentation strategies (i.e., random deletion, synonym replacement, swapping, random insertion, back translation, and paraphrasing) to train personalized AI models. The model is evaluate by using LoRA and two dataset proposed by the authors.
 
@@ -156,19 +131,16 @@ None
 ### Contribution
 1
 
----
-
-## Human Reviewer 4
-
 ### Rating
 1
 
-### Rating Number
-1
-
 ### Confidence
 5
 
+---
+
+## Human Reviewer 4
+
 ### Summary
 This paper studies the use of text based data augmentations techniques for learning models which can role play as specific characters in diverse scenarios. The paper compares 6 data augmentations strategies on two characters whose content they sample from games as the basis of training Llama 3 using LoRA.
 
@@ -198,3 +170,9 @@ For example, the section on Unsloth in section 2.4 and the inclusion of external
 
 ### Contribution
 1
+
+### Rating
+1
+
+### Confidence
+5

human_reviews/Lxst78Rrwj.md

@@ -0,0 +1,176 @@
+## Human Reviewer 1
+
+### Summary
+The paper utilizes the invariance property of $P(X \mid \text{Pa}(X))$ despite having different $P(\text{Pa}(X))$ to find the causal parents. For this purpose, they choose some source priors, generate corresponding datasets, and verify the invariance. Finally, they show their performance on a variety of synthetic and real-world setups.
+
+### Strengths
+The paper is well written. The figure is helpful for understanding the algorithm. The experiments are extensive.
+
+### Weaknesses
+## Minor weaknesses:
+* The authors mentioned “finding the maximum clique in an augmented bidirectional graph” multiple times but without a proper definition or example/visualization.
+* The source variables should be defined in a little more detail.
+* What does $P'$ in equation 2 refer to? It should be precise.
+* “The intuition is if we can re-sample $D_i$ from $D \sim P(X)$ such that $D_i \sim P_i(X)$,” This is a little unclear. How are $D \sim P(\mathbf{X})$ and $D_i \sim P_i(\mathbf{X})$ different?
+* It is unclear how $D_1, D_2, \dots, D_M$ are sampled. How are the $m$ source priors ($P_i(\mathbf{B})$) obtained? Although these are discussed later, some hints/intuitive discussion should be provided earlier in the paper.
+* “We cannot compute $P_i(X)$, … we can re-sample $D_i$ from $D$ so that $D_i \sim P_i(X)$” – based on my understanding, the first case is computing the numerical probability table, and the second case is sampling without any such table. This difference should be made clear.
+* More details on "downsampling without replacement" are needed.
+* An intuitive explanation of the “minimal downsampled rate” is required.
+
+
+## Major weaknesses
+* Suppose $Z$ is not a parent but an ancestor. Shouldn’t we also get variance = 0 (equation 1) in such cases? Does a change in $P(\text{Ancestor})$ affect $P(\text{descendant} \mid \text{ancestor})$?
+* Many important concepts are delayed until section 4.2. The authors should consider introducing them earlier in the paper.
+
+I will consider increasing the score after seeing author's response and reviewer discussion.
+
+### Questions
+## Questions:
+* How are the authors resampling the datasets?
+* Do you have to perform this invariance test for all possible parent sets?
+* Why is $Pa[B] = \emptyset$ in the definition of set $\mathbf{B}$ (section 4.1)? What does that imply?
+* In practice with real-world data, is the variance always zero for all true parents (equation 1)? Why or why not? Should a threshold be used?
+* How expensive is it to compute $\phi(X)$? Do we have to iterate over all $X$? And do it again after performing step ii in Theorem 3?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 2
+
+### Summary
+Based on the fact that the distribution of an effect, conditioned on its causes, remains invariant to changes in the prior distribution of those causes, this paper proposes a causal discovery algorithm for large-scale datasets. Specifically, it designs an invariance test, which is achieved by a downsampling scheme. It also makes the best of searching Markov blankets of all variables, to reduce the time complexity. Experiments on synthetic datasets and real-world networks show better scalability.
+
+### Strengths
+- This paper is written well, with clear descriptions and motivations. 
+
+- The authors propose practical algorithms for causal discovery, with some interesting theoretical findings, e.g., the basis of a DAG, the minimal downsampling rate, etc. 
+
+- The experiments under synthetic datasets and real-world networks are extensive, which verified the advantages in large-scale datasets.
+
+### Weaknesses
+- Some details seem to be missing in the paper.
+For example, 
+
+i) Footnote 2 and Theorem 1 tell how to find non-parent sets, whereas how to set the threshold for the variance is not clear. Please give the details in the paper.
+
+ii) How to learn the different priors $P_i(X)$, with the estimated $m$? Did the authors assume some distributions?
+
+- Theorem 2 provides a necessary condition to test whether a subset $Z$ is the parent set of X. However, it is not a sufficient condition. Although the authors stated, “When m is infinitely large, the implication in Eq. (1) becomes bi-directional and $V[P+(X | Z)] = 0$ definitively implies $Z = Pa[X]$”. It is not that clear why this implication we can get. Please elaborate on it more.
+
+- It is better to perform some real-world datasets for validation. This is because bnlearn provides real networks and it generates the data based on the networks. These datasets look like semi-synthetic. BTW, in Table 2, when dealing with small-scale datasets (or even a large-scale dataset Munin), the runtime all seem not to be satisfactory. Please explain it.
+
+### Questions
+- Does the proposed causal discovery algorithm work for time-series data? What are the challenges?
+
+I would be very glad to increase my score if the authors could resolve my concerns.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper introduces a novel approach to causal learning through a new invariance test for causality, which underpins a reliable and scalable algorithm for reconstructing causal graphs from observational data. This method leverages a core insight that the conditional distribution of the effect given the cause remains invariant under changes in the prior distribution of the cause. This insight enables a parent-identification process for each variable using synthetic data augmentation. This process is integrated with an efficient search algorithm that utilizes prior knowledge of each effect variable’s Markov blanket, along with the empirically observed sparsity of causal graphs, to significantly reduce computational complexity.
+
+### Strengths
+1. The proposed method is rather novel.   
+2. Overall, the paper is well-structured and clearly written.    
+3. The experiments are extensive, covering 3 types of functional causal models, 6 causal discovery baseline methods, and varying graph sizes.
+
+### Weaknesses
+Any thoughts on how to extend your method to handle heterogeneous or time-series datasets?
+
+### Questions
+(See above)
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes a new framework that leverages the invariance of effect conditioned on its causes for causal discovery from observational data. The main idea is it try to disturb the p(cause) distribution and see whether the p(effect|cause) would change after the disturbance.
+
+### Strengths
+- This work leverages the invariance of conditional distribution and then proposes a downsampling method, combining them to find the parent set.
+
+### Weaknesses
+- The main issue is that since the real intervention is not applicable to the observational data, it provides a downsampled technique to approximate the p(effect|cause) after the disturbance, which, however, has no theoretical guarantee. This is, how to guarantee such a downsampled correctly corresponds to the real distribution after the disturbance?
+- Since the basis variables would include the leaf vertices, in such a case, changing the prior basis variables will not affect the distribution of their ancestors, and thus it may not have a similar effect to changing prior over source variables.
+
+Typos:
+- "Theorem 2" -> "Theorem 4" in Theorem 5
+
+-----
+After rebuttal:
+
+After multiple rounds of discussion with the authors, my fundamental concerns remain unresolved. The core issue persists: the proposed approach fails to adequately address the challenge of obtaining interventional distributions from observational data.
+
+As highlighted in seminal works by Pearl [1] and [2], answering interventional questions such as "What would be the impact on the system if this variable were changed from value x to y?" requires explicit causal knowledge. This limitation is well-established in the causal inference literature, as also discussed comprehensively by [3].
+The authors' repeated attempts to address my concerns through mathematical manipulations and resampling techniques which do not overcome the fundamental identification problem in causality.
+
+[1] Pearl, J. (2009). Causality: Models, Reasoning, and Inference. Cambridge University Press.
+
+[2] Brouillard, Philippe, et al. "Differentiable causal discovery from interventional data." Advances in Neural Information Processing Systems 33 (2020): 21865-21877.
+
+[3] Bareinboim, Elias, et al. "On Pearl’s hierarchy and the foundations of causal inference." Probabilistic and causal inference: the works of judea pearl. 2022. 507-556.
+
+### Questions
+See the weakness above.
+
+### Soundness
+1
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+5

human_reviews/M9U49u9GA7.md

@@ -0,0 +1,196 @@
+## Human Reviewer 1
+
+### Summary
+This paper presents a work on learning with noisy labels in datasets annotated by LLMs. The proposed framework, SiDyP, is based on a noisy-supervised model called PLC and a simplex diffusion model. In this framework, the PLC model, fine-tuned with the noisy labeled data, is used to obtain dynamic training trajectories and induce a prior distribution, while the diffusion model serves as a posterior to iteratively calibrate PLC’s prior. Specifically, this paper introduces a true label candidates retrieval algorithm and a prior dynamic distillation
+algorithm to mitigate the negative impact brought by noisy labels. Extensive experiments demonstrate that the proposed framework achieves competitive classification performance on multiple tasks.
+
+### Strengths
+1. Denoising LLM-generated labels may be a promising direction and is worth further investigation.
+2. Experimental results are good, which shows that the proposed method is a state-of-the-art one.
+
+### Weaknesses
+1.The novelty of this paper is limited, capturing the label prior distribution based on training dynamics and using the simplex model for iterative refinement are the technical contributions of DyGen and TESS, respectively. The contributions of this paper seem like engineering innocations.
+
+2.The proposed method introduces a few more hyperparameters. In practice, when we don't have a clean validation set, these hyperparameters can be difficult to set.
+
+3.Writing needs to be improved. the dynamic prior needs to be clearly explained in the abstract and introduction, which can easily be confused with training dynamics. Additionally, the organization of Section 3 is not reasonable.
+
+4.A more detailed ablation study would have been desirable, such as the number of candidate labels in Algorithm 1 and the number of iterations in Algorithm 2.
+
+### Questions
+1.Compared to DyGen, what are the main contributions of SiDyP? It seems that SiDyP just replaces the VAE part in DyGEN with a simplex diffusion model. There are no insightful innovations.
+
+2.Line 165, why do we only consider the two candidate categories with the highest probabilities? In fact, the types and distributions of labels vary significantly across different tasks.
+
+3.Some descriptions and settings of hyperparameters can be confusing. Does the estimated error rate refer to the noise ratio? What are the bases for setting these hyperparameters?
+
+4.The symbols in Figure 1 are inconsistent with those in the main text.
+
+5.Section 4.5, the effectiveness of the proposed label candidate retrieval algorithm has not been validated.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes a method called SiDyP to address the learning-from-noisy-labels problem.  
+In the experiments, the authors validated the effectiveness of the proposed method using noisy data generated from the zero-shot and few-shot inference of Llama-3.
+
+### Strengths
+The authors conducted experiments using noisy labels generated by LLM Llama and demonstrated the effectiveness of the proposed method, which is one of the contributions of this paper.
+
+### Weaknesses
+1)	On the novelty of introducing LLMs for generating noisy data for learn-from-noisy-labels methods.  
+The authors claimed in the Introduction that, this is the first time that LLMs-generated datasets have been introduced in the realm of learning from noisy labels.
+Actually, this statement is not rigorous. 
+There are already some works that focus on learning from noisy labels generated by LLMs [1,2,3].
+These works concern the topic of Weak Supervision, which can be seen as a subset of the learning-from-noisy-labels problem; even though Weak Supervision considers the context of each one sample with multiple noisy labels, the learning-from-noisy-labels problem often refers to the context of one sample with one label.
+Anyway, the authors should take these prior works into account while writing.
+
+[1] Alfred: A System for Prompted Weak Supervision. ACL-2023 (demo).  
+[2] Combining prompt-based language models and weak supervision for labeling named entity recognition on legal documents. Artificial Intelligence and Law, 2024.  
+[3] https://snorkel.ai/blog/few-shot-learning-large-language-models/. Blog from Snorkel.
+
+
+2) On the core research motivation and contributions.  
+Related to the first issue mentioned above, and more importantly, the core research motivation and contributions of this paper are not sufficiently clear:  
+
+     2.1) If the designed method is more specifically tailored to noisy labels generated by LLMs (rather than noisy labels in the general learning-from-noisy-labels works), or in other words, if the proposed method is specifically designed for noisy labels generated by LLMs, then:
+    The authors should analyze from theoretical or experimental perspectives whether noisy labels generated by LLMs have different characteristics compared to noisy labels in the traditional sense, and why the designed method is well-suited to handle these characteristics.
+
+     2.2) If the designed method is not specifically aimed at noisy labels generated by LLMs, then this paper essentially presents a new typical learning-from-noisy-labels method.    
+
+     2.3) In summary, the authors do not clarify in the paper whether noisy labels generated by LLMs have different characteristics compared to traditional noisy labels, nor do they specify whether their contributions are specifically designed for noisy labels generated by LLMs.
+
+### Questions
+Please refer to the above Weaknesses.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper presents a novel approach using a simplex diffusion model for denoising labels generated by large language models (LLMs). The method aims to address the inherent noise in LLM-generated labels by leveraging an iterative denoising process within a continuous probability simplex space. The approach is motivated by the high cost and time demands of manual labeling, coupled with the practicality of LLMs for fast, albeit noisy, label generation. The paper introduces several key assumptions, including the local consistency of embedding spaces and the ability of diffusion models to gradually refine noisy labels into accurate representations. Experiments across different noise types, including symmetric, asymmetric, and instance-dependent noise, demonstrate the model's robustness. The results suggest that the method effectively reduces noise without requiring additional labeled data, making it a potentially valuable contribution to applications reliant on large-scale automatically generated labels.
+
+### Strengths
+The paper has several notable strengths across originality, quality, clarity, and significance:
+
+Originality: The approach is innovative in applying a simplex diffusion model to iteratively denoise noisy LLM-generated labels. Using the probability simplex space to handle label noise is a creative application, enabling a gradual refinement of noisy labels, which is distinct from traditional denoising methods. This method also addresses the growing need for handling noise in labels generated by LLMs, making the approach highly relevant and timely.
+
+Quality: The paper presents a structured approach with a combination of techniques, such as embedding-based candidate selection, to effectively manage noise. The experiment setup is comprehensive, covering multiple noise types (symmetric, asymmetric, instance-dependent), which reflects the robustness of the method. By comparing the proposed model to various baselines, the paper demonstrates a clear performance advantage in multiple scenarios, supporting the claim that the simplex diffusion model can enhance accuracy in noisy data environments.
+
+Clarity: The overall structure of the paper is clear, with a logical flow from the problem formulation to the model design and experimental results. The motivation for the proposed method is well-presented, with a clear description of how each component of the simplex diffusion process contributes to the final denoised output.
+
+Significance: The work is significant for applications that rely on LLM-generated labels, which often contain noise. By introducing a method that reduces noise without needing additional labeled data, the paper addresses a practical challenge, potentially improving the scalability and quality of labeling in real-world scenarios. This could be valuable across fields where large-scale data labeling is a bottleneck, such as natural language processing, computer vision, and medical imaging.
+
+### Weaknesses
+While the paper is promising, several areas could benefit from further improvement:
+
+The paper lacks in-depth theoretical discussion regarding why the simplex diffusion model is particularly well-suited to the task of LLM label denoising. For example, the rationale behind using iterative denoising in the simplex space could be strengthened by exploring why this approach may perform better than existing denoising methods in managing specific types of noise common to LLM-generated labels.
+
+Key Assumptions Verification: The paper is based on several assumptions, such as the local consistency of embeddings and the relationship between training dynamics and sample quality. However, these assumptions are not directly validated in the experiments. For instance, the assumption that noisy labels will have larger mean and standard deviation distances in the embedding space could be tested and quantified to provide stronger empirical support.
+
+The paper could provide a more detailed analysis of the noise characteristics specific to LLM-generated labels. The motivation would be more convincing if it explained how these labels differ from other types of noise and why standard denoising methods are insufficient. Understanding these nuances would clarify why the simplex diffusion model is a more suitable choice.
+
+Although the experiments include different noise types, they do not explore the method’s limitations in scenarios where noise is highly complex or includes extreme cases. Additionally, there is a lack of quantitative analysis regarding the training dynamics feature assumption, which would help validate the model’s robustness and provide further insights.
+
+The paper lacks a failure analysis that would highlight the model's limitations. Discussing scenarios where the model underperforms could provide readers with a more realistic understanding of its applicability and suggest areas for future improvement.
+
+### Questions
+Noise Characteristics in LLM Labels: Could the authors provide more insights into the specific noise patterns in LLM-generated labels that this model targets? Understanding these patterns would clarify why traditional denoising methods might struggle with this type of noise.
+
+Embedding Space Consistency Assumption: Can the authors validate the assumption that the local consistency of embeddings allows clean and noisy labels to be differentiated? This could be demonstrated by visualizing the embedding space to show the distribution differences between noisy and clean labels.
+
+Training Dynamics Validation: Could the authors conduct a more detailed analysis of the training dynamics related to noisy samples? Quantifying the mean and standard deviation differences in distances for noisy versus clean samples would strengthen the paper’s empirical foundation.
+
+Application Scenarios and Limitations: In which specific scenarios would the simplex diffusion model be most applicable or potentially limited? Additional examples would help readers understand the practical applicability of the method.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes applying simplex diffusion to denoise annotations generated by large language models (LLMs). The core idea is to first obtain noisy annotations from LLMs and then iteratively refine these labels with a simplex diffusion model, which maps labels to a continuous probability space. The main experiments are conducted across four common text classification datasets using Llama 3.1-70b. Overall the results show that the proposed method achieves competitive performance in both zero-shot and few-shot settings. Additionally, the authors explore the method’s robustness by testing with different 4 LLM backbones and different types of label noises.
+
+### Strengths
+Overall, I think this paper is well-written and relatively easy to follow. The results are very competitive and clearly show the advantage of the proposed method. I think the construction of experiments is correct.  Additionally, the authors clearly communicate the problem and the potential impact of their approach on improving LLM-generated annotations, which would be a valuable contribution to the field.
+
+### Weaknesses
+= The comparisons with Mixtral models might not be entirely fair. If the authors used the base Mixtral-8x22b model, which is not instruction/chat-finetuned, it may not perform comparably to other models that have been finetuned on instruction data. For a fairer comparison, using Mixtral-8x22b-Instruct-v* or similar instruction-tuned models would better align with the capabilities of the other models tested.
+
+= Insufficient literature coverage: Overall, this paper aims to solve the problem of LLM for data annotation and reduce annotation noise. I think it would be justifiable to have a better coverage of related work especially on LLM for data annotations. Give the limitation in space for the initial submission, I expect the authors to conduct a more rigorous literature survey of related work.
+
+= Although the approach is novel for this specific problem of denoising llm annotaions, the method largely adapts simplex diffusion for signal denoising directly to noisy label refinement. This may raise questions regarding the technical novelty, as the adaptation itself does not appear to involve significant theoretical modifications.
+
+### Questions
+It would be helpful if the authors could clarify the following:
+
+= For missing labels, why was a random label assignment chosen? Given the low failure rate, might it be preferable to disregard these cases instead?
+
+= Were any specific system prompts or chat templates used when applying the prompts for LLM label generation? If so, could the authors provide details?
+
+= In the few-shot experiments, the authors selected one example per label. Were these demonstration examples sampled from the training set or the validation set?
+
+= For the 20News dataset, is there a reason why no results are reported for the few-shot experiments?
+
+= Since the validation set is used for model selection, as an exploratory experiment, have the authors considered prompting the LLM with a larger few-shot sample (e.g., 50+ examples) from the validation set, given the support for long contexts in the LLMs used? It would be interesting to see how the efficiency of this approach compares with the proposed diffusion-based denoising pipeline.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3

human_reviews/MMHqnUOnl0.md

@@ -0,0 +1,144 @@
+## Human Reviewer 1
+
+### Summary
+The paper introduces a pre-training strategy, **Hierarchical Encoding for mRNA Language Modelling (HELM)**, that incorporates the hierarchical structure of mRNA codons into language model training. By modulating the model's loss based on codon synonymity, HELM aligns learning with the biological structure of mRNA. Results show HELM outperforms non-hierarchical models by around 8% on various mRNA tasks, including antibody annotation and property prediction, while also enhancing generative capabilities for mRNA sequences.
+
+### Strengths
+- HELM integrates codon hierarchy into model training, capturing the inherent structure of mRNA data.
+- Demonstrates a good performance improvement over existing models on multiple downstream tasks.
+- The model can generate more biologically plausible and diverse mRNA sequences than existing approaches.
+
+### Weaknesses
+- The authors curate a dataset from OAS specifically to train their models. This may explain much of the performance gains observed in Table-1, as the model is trained and tested on data with similar statistical characteristics. An ablation study isolating the effect of the curated dataset would clarify the extent to which performance gains are due to the data source versus tokenization and architecture design.
+- As the authors noted it, representing hierarchical relationships in Euclidean space might limit HELM’s ability to capture the full complexity of these structures, potentially affecting performance.
+- Although the results are encouraging, they are primarily based on antibody and property prediction tasks, which could restrict the model's applicability to other mRNA-related contexts.
+- Additionally, the strategy for splitting the data into training, validation, and test sets is crucial and should be detailed in the main paper. In sequence-based datasets, it is common practice to split data based on clusters after clustering sequences, which minimizes data leakage. The reported performance gains may be overstated if there is any leakage between the splits.
+
+
+I am open to revisiting my score; however, I will wait for: (i) the authors' response to my comments, and (ii) feedback from other reviewers.
+
+### Questions
+- Are you planning to open-source the curated dataset ?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper introduces HELM, a hierarchical encoding approach tailored for mRNA language modeling that leverages the codon-level structure of mRNA sequences. Unlike traditional models, HELM incorporates a hierarchical cross-entropy loss function to align with the biological structure of mRNA, particularly focusing on synonymous codon usage and its functional implications. The model is evaluated on multiple mRNA datasets, demonstrating improvements in downstream prediction tasks, generative diversity, and sequence annotation accuracy over non-hierarchical baselines. Overall, HELM represents a biologically-informed advancement in mRNA modeling, showing enhanced performance on tasks relevant to protein synthesis and gene expression.
+
+### Strengths
+1. Biological Prior Integration: The hierarchical encoding strategy effectively incorporates biological knowledge of mRNA structure, particularly the synonymous codon usage, which enhances the model’s performance on property prediction and generative tasks.
+2. Diverse Evaluations: HELM's performance is thoroughly evaluated across multiple tasks, including property prediction, generative sequence design, and antibody sequence region annotation, showcasing the model’s versatility and relevance to bioinformatics applications.
+3. Improved Representational Quality: By aligning model training with the codon hierarchy, HELM captures the underlying biological structure more effectively, achieving better clustering of synonymous sequences and improved predictive accuracy in key bioinformatics tasks.
+
+### Weaknesses
+1. Lack of Scaling Experiments: The paper primarily employs models with 50M parameters, which is relatively small compared to large language models (LLMs) in NLP. This limitation raises concerns about the necessity of HELM’s hierarchical encoding, as larger models might naturally learn these hierarchical relationships without explicit design. Experiments with larger models could help clarify if the hierarchical loss function offers unique advantages or if it becomes redundant with increased scale.
+2. Limited technical contribution:  The primary contribution of this work lies in adapting hierarchical cross-entropy (HXE) to the mRNA structure, which, while valuable for bioinformatics, represents a limited advancement from a machine learning perspective. The adaptation of HXE to this domain may not provide significant innovation to the broader ML community.
+3. Baseline suggestion: A simpler way to capture the hierarchical structure could be to introduce separate embeddings for different types of codons and combine these with the token embeddings as input. For instance, embeddings such as [start], [stop], [S], and [Q] (as shown in Figure 6) could be used to represent codon types without modifying the loss function. This approach would provide a straightforward baseline for evaluating the benefits of the proposed HXE solution.
+
+### Questions
+See W3.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces Hierarchical Encoding for mRNA Language Modeling (HELM), a novel pretraining strategy that incorporates the hierarchical codon structure of mRNA into language model training. It addresses the limitations of existing models that overlook codon synonymity, which can lead to suboptimal performance in mRNA tasks. HELM modifies loss functions to prioritize errors between different amino acids over synonymous codons, enhancing the model's alignment with biological realities. Evaluations show that HELM outperforms standard language model pretraining and other state-of-the-art RNA models by approximately 8% across six diverse downstream tasks, including property prediction and antibody region annotation, while using fewer model parameters. Additionally, HELM demonstrates improved generative capabilities, producing mRNA sequences that better align with true data distributions. Overall, HELM effectively captures the hierarchical nature of mRNA, leading to enhanced performance in both predictive and generative tasks.
+
+### Strengths
+1. This study proposed a novel approach for embedding biological hierarchical structures into language models to enhance interpretability and accuracy in biological sequence analysis.
+2. A comprehensive technical evaluation spanning multiple model architectures was presented, benchmarked across diverse datasets and use cases.
+3. Clear and precise presentation of technical methods and experimental setup were provided to facilitate reproducibility and further research.
+
+### Weaknesses
+1. The performance improvement seems more attributable to data selection than methodology: - HELM uses antibody mRNA while baselines use different data types (ncRNA, pre-mRNA, diverse organisms mRNA).
+2.  The paper fails to justify why antibody mRNA pre-training would generalize to: - Viral RNA sequences - Riboswitch sequences.
+3.  The evaluation methodology is insufficient: a. Over-reliance on single metric (Spearman correlation) b. No analysis across sequence lengths c. No biological interpretation of prediction errors d. The evaluation lacks comprehensive stress testing on edge cases (e.g., extreme sequence lengths, unusual GC contents, rare codon clusters), raising concerns about the model's reliability in challenging real-world scenarios.
+4. Methodological issues: (1) α parameter (0.2-0.6) lacks biological basis (2) Oversimplified codon bias analysis (3) Insufficient validation of hierarchical structure's biological relevance.
+5. Limited practical applications: (1) Too narrow focus on antibody sequences and expression (2) Critical applications remain untested: a. RNA vaccine design applications b. Abnormal or mutated sequences c. Real-world biological scenarios.
+
+### Questions
+1. What is the biological justification for the α parameter range selection?
+2. How does the model handle sequences with unusual codon usage patterns? Such evaluation is crucial because unusual codon usage patterns could significantly impact translation efficiency and protein expression levels, yet these scenarios are not addressed in the current validation.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper proposes an elegant and novel solution for incorporating codon hierarchical structure into language model training, which is hierarchical cross-entropy loss. Although the method is simple, the biological insight is deep and the performance is convincing. The authors also curate the hierarchical information and a pre-train dataset, which may benefit future research. Overall, I feel this would be an impactful work for the community.
+
+### Strengths
+1. The authors combine biological insight with language modeling.
+2. The proposed HELM is solid and performs well for mRNA-related tasks. It achieves SOTA performance on several practical and impactful tasks, e.g., mRNA sequence design, and sequence region annotation.
+3. The authors provide a curated dataset, and curated domain knowledge, which may benefit future computation-oriented research.
+4. The authors provide solid benchmark experiments of various backbone architectures and tokenization methods.
+
+### Weaknesses
+1. The methodology itself is relatively simple.
+2. It is not clear to what extent the model, the codes, and the data will be made public.
+
+### Questions
+I do have a strong concern for the reproducibility of this paper. Despite the fact that I think highly of this paper, I don't think the paper should be accepted without further declaration on reproducibility. Specifically, will the model checkpoint, source codes, and datasets be released?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+4
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/N83O2FcqzN.md

@@ -0,0 +1,238 @@
+## Human Reviewer 1
+
+### Summary
+A latent variable model is described and used to reanalyze a publicly available dataset of population recordings in mouse visual cortex.
+In particular, this work proposes a time series analysis of the neural dynamics in response to videos.
+This method combines a variational auto-encoder loss and a self-supervised contrastive learning objective in the latent space.
+The latent space is split into a static state and a dynamic state, and both evolve according to learned recurrent dynamics.
+
+### Strengths
+As large population recordings neural datasets are being measured, developing correspondingly flexible latent time series models is an important research direction.
+This work proposes an architecture and unsupervised objective for discovering structure in such neural data.
+The focus on neural dynamics in response to video input is timely and relevant.
+The presentation begins with simple toy synthetic experiements and then builds to the analysis of a rich neural dataset.
+
+### Weaknesses
+The proposed architecture is quite complex and no clear justification is provided for the design of factorized latent states. Moreover the latent states are also swapped for the computation of the objective function. I see that these design choices are inherited from a previous publication, but I think that it would help the reader if you could explain clearly the reason for these choices. If a conceptual or mathematical explanation is not possible, ablation experiments could also help (e.g. with vs. without latent state factorization).
+The current explanation is vague, e.g. line 231: "To enhance the effect of the positive sample, we adopt the practice of swapping content latent variables between the positive pairs while maintaining style latent variables."
+
+The proposed objective function combines three terms: a VAE style ELBO, a weight decay regularizer, and a contrastive term.
+Your reader will find it difficult to think about the overall objective: what is being optimized exactly?
+Is this multi-objective optimization still amenable to an interpretation in terms of likelihood?
+It seems like you are combining a generative and a disctiminative objective, I wonder if it may be possible to express all that in a nice unified and coherent Bayesian framework, rather than simply saying that you added many terms.
+This is also important for the interpretation of the results, it would help clarify what each of these part is contributing and may also help selecting natural hyperparameters.
+The writing is vague and should be clarified, for example:
+line 74: "we apply self-supervised contrastive learning to enhance the time constraint and to shape latent variables"
+line 134: "The output ... is an estimate of firing rates of the input."
+line 213: "Besides, we added L2 regularization to the expectation and log-variance of the prior distribution to stabilize model training."
+
+The results are difficult to interpret because you compare methods that are all quite sophisticated.
+It would help if, for each task, you reminded your reader where chance performance is.
+Another way to do that would be to provide some simple linear method baseline.
+Some of the writing is unclear, e.g. line 298: "The dataset contains 32 sessions, each for one mouse. Since the class of neurons responsive to natural visual stimuli is found in six visual regions, in this work we choose to analyze the neural activity of five mice that have as many neurons as possible (about 300, see Appendix D for exact numbers) with them evenly distributed across all regions (the coefficient of variation for the number of neurons across six brain regions is below 0.5)."
+
+Most of the evaluations are concerned with properties of the learned representations.
+Adding a discussion of the performance of the autoencoding performance would help.
+How well are the spikes reconstructed by the autoencoder, can you show examples of successes and failures?
+
+### Questions
+Have you considered using only the contrastive objective, i.e. without the decoder and VAE loss? I am asking this because all your benchmarks quantify decodability of image identity from the learned representation.
+ 
+You write line 303 "We select five scenes that elicit the strongest average responses".
+Could you tell us more about the images you selected, and maybe show examples?
+How does you method behave on the other images, is the approach robust enough to apply to lower signal to noise examples?
+
+Again for the natural videos, it would help if you could give a precise description of the stimulus.
+In particular, how much motion is there in these videos?
+This is important because you define a very liberal criterion for correct decoding, line 447: "we take the accuracy measured by considering the error between a predicted frame and the true frame within 1s (default size of time window constraint) as a correct prediction."
+
+In line 409 and following, you interpret the shape of the tSNE visualizations.
+Is it safe to interpret such a nonlinear embedding? How much of your interpretation depends on hyperparameter choices?
+
+There is a very large performance difference between Mouse 1 and others in Table 3 and 4, what do you make of the variability and can you help us understand this outlier? You already removed one mouse from your analysis, what was the reasoning behind that?
+
+The original paper from which you get the neural data was focused on describing the mouse visual hierarchy. Can your method reveal interesting structure in each visual area and can it reveal a gradient in the visual representation along the cortical hierarchy?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes TiDeSPL-VAE, a sequential latent variable model that aims to jointly model neural and visual inputs. The model uses separate embeddings to represent visual features and internal states. However, there are significant issues in motivation and clarity that hinder evaluating its contributions in its current form.
+
+**Strengths:**
+- The idea of separating visual features from internal states could be beneficial for neuroscience and machine learning, with potential applications in visual decoding.
+  
+**Concerns:**
+1. **Motivation and Related Work**: The study’s motivation is weak and lacks clarity. Statements on motor encoding and temporal relationships lack precision, and relevant prior work on task-relevant/irrelevant models is omitted. Checking for grammar would also be good to do.
+2. **Benchmarking and Fairness**: No clear details on benchmarking models, hyperparameter choices, or treatment of time dependencies are provided. Models like CEBRA and LFADS are compared inadequately due to differences in architecture and training parameters. Additionally, benchmarks on common datasets like the Allen Brain Natural Movie 1 or Sensorium Challenge data are missing, limiting reproducibility/interpretability.
+3. **Scientific  Impact**: It remains unclear if the model’s components contribute unique insights, as it largely combines established techniques (VAEs, contrastive learning) without a compelling scientific rationale. The model complexity, evidenced in ablations, appears to drive performance rather than underlying innovations in latent space separation.
+
+**Recommendation:**
+While TiDeSPL-VAE may hold promise, the paper’s current form lacks clarity and adequate benchmarking, raising concerns about model novelty and effectiveness. Addressing these issues is essential for it to be a meaningful addition to the field.
+
+### Strengths
+Anonymous et al. present a method for joint modeling of neural and visual inputs called TiDeSPL-VAE. The proposed method leverages a sequential latent variable model (LVM) to learn two separate embeddings - one to represent the visual latent features and one to capture the animals' “internal state.” As such, building robust and interpretable latent variable models is of great interest to fundamental neuroscience, translational work (restoring vision), and in machine learning. However, due to missing details and critical baselines (see below), it is not clear the originality or value of the solution they present.
+
+### Weaknesses
+I find the study poorly motivated with statements such as *“Most work has focused on analyzing motor neural activity that controls clear behavioral traces and has modeled neural temporal relationships in a way that does not conform to natural reality.”* - what is natural reality? 
+Also, note that actual realistic motor behavior encoding-decoding is an unsolved problem. 
+
+Or equally worrisome: *“even though how the visual system encoded input to recognize objects is a primary topic(DiCarlo et al.,2012), and decoding visual neural activity visual stimuli is a challenging research highlight in the neuroscience community (Kay et al.,2008; Wen et al.,2018; Duetal.,2023). Furthermore, existing LVM treat temporal relationships unnaturally (Pandarinathetal., 2018; Schneider et al., 2023) or even don't consider time dependence (Zhou & Wei, 2020; Palmerston & Chan, 2021). Given that visual neural activity has strict antecedent time dependence, these models may struggle to build high-quality latent representations.”* - as a trained neuroscientist and ML expert, I can confidently say this is not neurally accurate. Also, how does LFADS or CEBRA treat “temporal relationships” unnaturally? LFADS uses a dynamical system and CEBRA can leverage the natural time inherent in the neural code, which is a tunable hyperparameter. 
+
+**Related Works/Motivation**
+
+- Why not motivate this study by just stating what you do – you want to see if separating visual features from internal states could lead to better LVMs for understanding (decoding) visual systems. 
+
+- You also miss the related work on task-relevant and task-irrelevant models, such as [1](https://scholar.google.com/citations?view_op=view_citation&hl=en&user=csGAeKgAAAAJ&sortby=pubdate&citation_for_view=csGAeKgAAAAJ:WbkHhVStYXYC) and [2](https://pubmed.ncbi.nlm.nih.gov/37398400/), which should be benchmarked. 
+
+- You also claim, of the models you benchmark: *“None of them build latent representations progressively along the chronological order.”* (L253) - what does this mean? 
+
+**Benchmarking Issues**
+
+- In the main text or Appendix, there are NO details on how you benchmarked the other models, aside from *“In this experiment, we set all models to have 128-dimensional latent variables and train them for 5,000 iterations”*. How is this a fair comparison? Did you not try other dimensions and training losses until convergence? How did you treat time in CEBRA or LFADS, as you can set different offsets - you only mentioned once: *“5 time points for natural scenes and for natural movie”*, but this does not tell me about the offset time in the model? 
+
+- Moreover, the complexity of the different models - B-VAE, LFADS, CEBRA are extremely different. For example, the work from Schneier et al. 2023 (CEBRA)  uses only a 5-layer MLP/GeLU, as the innovation is in the generalized contrastive loss - the point was you can get great performance with a simple model; to fairly benchmark this, then you should use the same RNN as the base model (vs. their MLP). All models you test have many hyperparameters and a full analysis of the parameters you tested is absolutely critical to show. 
+- You additionally need to compare model complexity and ideally, run-time estimations.
+
+- You use the Allen Brain Observatory Dataset because you allude that this is the data that Schneider et al. 2023 use, which, I agree, would be ideal. However, you do not benchmark on Natural Movie 1 in the same way; rather, your main benchmark is Natural Images – notably, devoid of time in the stimulus space. It is strange to me that you don’t benchmark on any data - synthetic or otherwise - that anyone else has used in ML to benchmark on. It raises the concern that you are unfairly enhancing the performance of your approach compared to letting the original authors run their code on common data. See my points above.
+
+- Thus, my strong recommendation is that you should also test your approach on “motor tasks” on standardized benchmarks, such as the Neural Latent Benchmark, as well as established vision benchmarks, such as the [Sensorium Challenge](https://www.sensorium-competition.net/) for natural images. At the very least, the Allen Brain Natural Movie 1 data as it was presented in CEBRA where you can add their numbers to a Table to directly compare; here you don’t reach the same decoding performance and you seemingly pick the neurons differently (they report close to 90%, which here your overall is much lower (65%), again making it hard to compare.
+
+- Scientifically, I am also missing how the separate embeddings perform in tasks; by separating vision from task-irrelevant information, do you see a more interpretable latent space? 
+
+- Moreover, it is unclear to me how this is not just an A+B solution; you leverage VAEs, and swaps, and contrastive learning, I don't inherently see the novelty (but this does not affect my score, novelty is a construct anyhow), but I would like to better understand the scientific rationale for the design choice.
+
+**Summary**
+
+I believe your approach could be useful for the field, but in the paper's current form:
+
+1. Lacks the proper motion and clarity in grammar/thought.
+2. It is impossible to know if you did a fair comparison due to lacking details, and common benchmarks, and the missing required benchmarking of alternative methods. 
+3. Missing task-relevant and task-irrelevant baselines.
+4. Your method is more complicated than those you benchmark, and your ablation studies hint at this as well. Comparing Mouse 2 in Table 3 you reach 65.38%, while the second best (much simpler) method reaches 52.76% decoding performance. Now, in Table 4, if you remove anything, you are well below the 52.76%, suggesting that the complexity of your model is driving performance. Just as another test, if you drop in an MLP (as you did from GRU→ LSTM), do you still outperform other models? I would suspect not.
+
+### Questions
+Please see above weaknesses.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces a new model, the TiDeSPL-VAE, to decode mice's neural activity during passive viewing of natural images and movies. 
+The TiDeSPL-VAE is a recurrent auto-encoder-like model composed of two latent spaces, each updated based on the current input and the recurrent state of the previous input. The model is also trained using a combination of losses: reconstruction, KL, and contrastive.
+All in all, the authors demonstrate that the TiDeSPL-VAE outperforms numerous baselines in reconstructing spiking activity on synthetic and naturalistic datasets.
+
+### Strengths
+The authors have been rigorous in the research methodology, comparing their model against a wide variety of baselines and testing the ability of these models on different datasets, from synthetic to more naturalistic settings -- and a complementary ablation detailing the critical components of the model.
+
+The authors also attempt to solve a novel problem by decoding visual data from mice instead of the widely studied mice's motor system.
+
+### Weaknesses
+**Major Weakness**: The authors insist on the novelty of their work as it deals with visual data instead of motor data from mice. However, the format of the data -- spiking activity -- is treated similarly, whether it is from the motor or the visual cortex. The model still takes the spiking activity as input and outputs the reconstructed spiking activity. To take advantage of the visual modality, it would be interesting to evaluate whether it is possible to decode the input stimuli (image or movie) or to relate the model better with neuroscientific findings about the mouse visual cortex. As it is presented now, the model could also be employed on motor data and perform similarly well. The use of visual data should be better exploited to highlight the scientific advances of this article more than the methodological part. 
+
+**Minor Weaknesses**:
+- The writing should be slightly improved to make the paper easier to read. A lot of sentences are not very clear, and others have grammatical flaws. For example, L269 "... and others are even difficult to split four clusters ..."; L482 "contrast learning"; ...
+
+### Questions
+- Why are the reconstruction and KL losses computed on the positive samples as well as the other samples? Isn't there an incentive already with the contrastive loss to emphasize positive samples? Can the authors include in the ablation study the training with the losses on all samples only to evaluate the impact of these terms?
+
+- L259, the authors mention that the two synthetic datasets consider "different properties of visual neural activity". What are these properties?
+
+- It looks like the authors train a model for each mouse and treat each mouse as a single dataset. How different would the results be if the authors tried training on all the mice? Or training on some of them and testing on the left-out individuals?
+
+- Can the authors include in appendix the number of samples (train, val and test) for each dataset? It looks like, with all the exclusion criteria, the dataset size ends up being very small. Can the authors confirm that the models are not overfitting on data of a single animal?
+
+- Can the authors also include a table indicating the number of parameters of each model in the Appendix to confirm that the comparisons are fair?
+
+- The ablation study highlights the importance of the contrastive loss and swap operation. Without these objectives, the model does not outperform most of the baselines. But the model has other characteristics that the baselines do not have such as the recurrence due to the two RNNs and two latent spaces. To ensure that these additions are important, can the authors include in the ablation study a non-recurrent version of the model (i.e. setting the number of timesteps of the RNNs to 1), and a version without each of the latent spaces?
+Additionally, would it be possible to train one of the best baselines with the contrastive loss and the swap operation?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The authors introduce the latent model, which separates the visual activity in the content and context representations, also taking the time dependency into the account. This seems to help to have more disentangles representations of mice visual cortex activities
+
+### Strengths
+* Original work, applying a new idea of splitting context and content of neural activity in the visual cortex
+* The necessary baselines such as LFADS, pi-VAE, and CEBRA are represented in the paper
+* Careful evaluation on both synthetic and real-world datasets, also following classic for the field benchmarks, like Lorenz systems
+* Sanity checks experiments, such as shuffling time dependent dateset help to disentangle the impact of time-dependency for the model 
+* The authors are very transparent about the limitations of the model
+
+### Weaknesses
+* Reproducubility might be an issue. The code is not provided and the implementation details in the appendix A are not sufficient. Details like how many RNN layers were used, how exactly the RNN was implemented, etc are missing. GRU is mentioned later in the paper but I think an explicit model architecture picture and more parameter details in appendix A would help to improve the paper.
+* Generalisability and robustness of the model is an issue as scores per mouse differ a lot, also it is not clear how the model would behave if the animal was shown different stimuli (ie gratings instead of the natural images).
+* Some references like [1,2] are missing, which could be nice to position the TiDeSPL-VAE model in the space of LVM as they are the representatives of modern non-VAE based LVM (while they are different from the current work)
+
+[1] Bashiri et al 2021 https://proceedings.neurips.cc/paper/2021/hash/84a529a92de322be42dd3365afd54f91-Abstract.html  
+[2] Kapoor, Schulz et al 2024 https://arxiv.org/pdf/2407.08751
+
+### Questions
+* RNNs are often tricky to train. Were there any issues that you phased during the model training or some details, which are crucial for a successful learning. 
+* Did you train a new separate model per mouse? Or was it a shared model across mice?
+* In figure 3 for TiDeSPL-VAE, LFADS and especially CEBRA there seems to be a clear train/test bias (points are separated), which is stronger than the cluster separartion based on the stimuli. How do you interpret this?
+* how exactly the train/test splits are done with respect to the time? As the train loss takes a sequence of responses from the near future with some $\vartriangle t$ - do you ensure that none of the validation/test responses were shown during training?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3

human_reviews/OpNMWVDdKS.md

@@ -0,0 +1,220 @@
+## Human Reviewer 1
+
+### Summary
+The paper proposes Associative Latent DisentAnglement (ALDA), a novel method for zero-shot generalization in vision-based reinforcement learning (RL) without data augmentation. ALDA combines disentangled latent space representations with an associative memory mechanism inspired by biological systems like the hippocampus. This enables RL agents to generalize to out-of-distribution (OOD) environments by recalling previously seen factors of variation, without need of augmentation in training. ALDA outperforms several state-of-the-art methods in tasks involving distribution shifts, demonstrating zero-shot generalization without using data augmentation.
+
+### Strengths
+Originality:
+The idea of using disentanglement and associative memory for RL to achieve zero-shot generalization is relatively less explored. It is nice formulation to combine both for RL. Unlike existing methods that rely heavily on data augmentation, ALDA achieves generalization without it.
+
+Quality:
+The paper is well-researched. It shows data augmentation is one form of disentanglement. It proposed new methods to combine disentanglement and association. The experimentation is carried with ablation studies on different component of the method.
+
+Clarity:
+The paper is generally clear, with a well-structured narrative that explains the problem, the motivation, and the solution. It draws inspiration from biological system which is well articulated
+
+Significance:
+Its significance lies in that it presents a method for generalization without augmentation. It would be very useful under the case that we cannot anticipate what shift will happen.
+
+### Weaknesses
+The papers lacks a more thorough discussion of its failure cases.
+
+Suggestion: 
+1. What if you vary distracting intensity, how would that change the relative results, especially compared to SVEA. 
+2. I also notice it perform worse than SVEA in the training environment, can you add interpretations on that?
+
+### Questions
+1. In figure 6, top half, the robot starts from different pose when traversing the latent code. what if it starts from the same as in the bottom half
+
+2. What if we compare the proposed method to data augmentation method that uses simpler augmentation, such as color jitter or random crop? For example 
+
+Kostrikov, Ilya, Denis Yarats, and Rob Fergus. "Image augmentation is all you need: Regularizing deep reinforcement learning from pixels." arXiv preprint arXiv:2004.13649 (2020).
+
+Hansen, Nicklas, and Xiaolong Wang. “Generalization in Reinforcement Learning by Soft Data Augmentation.” 2021 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2021. Crossref, https://doi.org/10.1109/icra48506.2021.9561103.
+
+You can also test of SVEA only using simpler augmentation.
+
+I understand that the main contribution of this paper is generalization without augmentation, it is okay that it is not performing as good.
+
+3. Can you combine the proposed approach with data augmentation? How would you do that?
+
+4. Can you combine disentanglement / association with dynamics? How would do you that?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+A common way to improve the generalization ability of an agent in reinforcement learning is through data augmentation. However, collecting data can be costly with increasingly varied tasks.
+This paper achieves zero-shot generalization by learning disentangled representations, where each latent feature dimension corresponds to an unique source of variation in the input image.
+Inspired by neuroscience, the authors incorporate associative memory into the original representation learning loss in QLAE, and build a novel framework called Associative Latent DisentAnglement (ALDA) for RL.
+ALDA is evaluated on various control tasks with two types of variations (color-hard and distracting backgrounds), and compared to state-of-the-art methods.
+While its performance lags behind data-augmentation-based methods for generalization, it outperforms previous approaches of learning disentangled representations.
+
+### Strengths
+- Improving generalization by learning better representations is not a new concept.
+For instance, Zhang et al. [2021] aim to encode only task-relevant information from input images.
+However, the idea of training disentangled representations specifically for generalization, particularly without data augmentation, is a novel approach.
+Recently, numerous studies have explored the effects of scaling laws on enhancing generalization, yet methods for achieving generalization without exposure to out-of-distribution data remain relatively under-explored.
+
+- The idea of considering memory for generalization is intuitive and the author build an interesting connection between the disentangled representation learning framework and associative memory network.
+
+A. Zhang, R. McAllister, R. Calandra, Y. Gal, and S. Levine. Learning invariant representations for reinforcement learning without reconstruction. ICLR 2021
+
+### Weaknesses
+- According to the paper, the strong disentanglement requires that each dimension in the latent representation corresponds to a unique source of the variation in the image.
+As for the weak disentanglement, a more clear and precise definition (line 216 to line 219) might be helpful for the reader to understand. 
+For example, one natural question to ask is if several dimensions in the latent representation $\textbf{z}$ correspond to one unique source, is this a weak disentanglement?
+I would appreciate it if you could provide concrete examples for these two types of disentanglement.
+This would help readers better understand the key concepts and their implications for the method.
+- And the relationship between disentanglement and generalization ability is not so clear to me.
+According to the line 233 to line 235, if I understand correctly, weak disentanglement can also lead to generalization.
+Is strong disentanglement a necessary condition for generalization or is weak disentanglement sufficient?
+If not, please explain more about the benefits and drawbacks of using strong disentanglement vs. weak disentanglement? 
+- In the section of experimental results, an explanation is provided for the performance gap between ALDA and SVEA.
+However, as mentioned in the paper, the proposed method can definitely be combined with data augmentation.
+So, it will strength the paper if the results of combining ALDA and data augmentation are presented.
+- minor: missing punctuation on line 305.
+In figure 2, $n_s$ is not defined in the paper. 
+I guess it is $n_z$ on line 290.
+
+### Questions
+- In Figure 2, the encoder is only updated by the ALDA loss, as indicated by the green arrow and the latent model is updated by both the ALDA loss and the critic loss, as indicated by the red arrow.
+Is there a reason for this? Any ablation study for this design choice?
+- On line 345, according to "the scalar codebooks Z, which can be interpreted as predetermined memories", are the codebooks trained or pre-defined?
+If trainable, are they trianed with equation 8 such that $V$ are actually trained parameters?
+If pre-defined, how the values are chosen?
+- Is there an intuitive connection between equation 7 and Figure 1?
+
+### Soundness
+4
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This work introduces Associative Latent Disentanglement (ALDA), a zero-shot generalization strategy for vision-based RL free of data augmentation. Inspiring by neurobiology, ALDA enables agents to generalize to out-of-distribution (OOD) activities by combining associative memory with disentangled latent representations. The results of the experiments show that ALDA is just as good at generalization as or better than current augmentation-based methods, especially in difficult OOD settings.
+
+### Strengths
+1. The paper takes ideas from neuroscience and combines latent disentanglement with associative memory to present a new, augmentation-free method for zero-shot generalization in RL.
+
+2. The methodology is sound, with experiments across challenging tasks that effectively support the primary claims of the paper.
+
+3. The paper is generally clear and well organized, though some technical sections could be further clarified for readability.
+
+4. This work has potential impact by presenting an efficient alternative to data augmentation, likely inspiring further research on generalization in RL without large datasets.
+
+### Weaknesses
+1. More proof of ALDA's generalization ability would be obtained by extending the studies to more varied settings, such as Procgen.
+Procgen's randomly produced levels and diverse game environments present distinct challenges, unlike those found in the DMControl suite. These settings may demonstrate the efficacy of ALDA's generalization across diverse activities, offering a comprehensive assessment of its adaptation to novel and intricate situations.
+
+2. More illustrations or pseudocode could help to clarify the function of associative memory in processing OOD samples. Particularly, pseudocode illustrating ALDA's processing of an OOD sample—from initial encoding to final output—would effectively elucidate the function of associative memory in the generalization process.
+
+3. An assessment of ALDA's effectiveness in real-world settings would demonstrate its applicability and scalability. For example, testing on a robotic manipulation job requiring the robot to generalize to objects of diverse forms, sizes, or textures not seen during training, could effectively showcase its generalization capabilities in practical applications.
+
+### Questions
+1. In order to show that ALDA is generalizable to a wider variety of OOD contexts, could you test it on other benchmarks, such as Procgen?
+
+2. Could you elaborate on the associative memory dynamics in ALDA, perhaps using a pseudocode or flow diagram, to show how it specifically supports zero-shot generalization?
+
+3. Is the method applicable to an offline reinforcement learning setting, given that online reinforcement learning can be costly and risky in real-world environments? It would be beneficial to address the adjustments necessary for adapting ALDA for offline reinforcement learning, along with the potential obstacles the authors expect in this adaptation.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper proposes an algorithm based on SAC, that performs latent disentanglement representation and combines it with an associative memory model to achieve zero-shot generalization in vision-based RL. The paper shows that the proposed approach improves upon several baselines when tested in two evaluation environments (color hard and distracting cs). In addition, the authors theoretically prove that data augmentation methods for vision-based RL implicitly perform "weak" disentanglement, and therefore lead to a noticeable zero-shot generalization.
+
+### Strengths
+* The paper tackles the important challenge of generalization in vision-based RL.
+
+* The paper is well-written and easy to follow.
+
+* The background section clearly explains the different methods for disentanglement representation and associative memory. Also, it provides motivation for using these methods for zero-shot generalization.
+
+* There is a good discussion about the relevant work of zero-shot generalization in vision-based RL.
+
+* The paper visually demonstrates (Figure 6) that the resulting disentanglement representation successfully encodes different data attributes.
+
+### Weaknesses
+* The method was tested in a narrow set of test environments (color hard and distracting cs). The paper could be improved if the authors would show the benefit of the proposed approach on standard (and more challenging) procedurally generated vision-based datasets for zero-shot generalization in RL, such as Procgen [1] or Crafter [2].
+
+* In lines 180-184, the authors mention that one drawback of previous approaches is that the latent representation excludes all information not relevant to the training tasks, which can hinder adaptation to new tasks that involve information that was previously considered irrelevant. The paper would benefit from a concrete experiment that showcases this failure and demonstrates how the proposed approach handles it. 
+
+* Some baselines are missing from the experiment section. For example, a comparison to methods that use the information bottleneck to exclude irrelevant information from the state representation in vision-based RL, such as [3]. 
+
+* An ablation study of |zd| (the dimension of the latent space) is missing. 
+ 
+[1] Cobbe K, Hesse C, Hilton J, Schulman J. Leveraging procedural generation to benchmark reinforcement learning. In International conference on machine learning 2020 Nov 21 (pp. 2048-2056). PMLR.
+
+[2] Hafner D. Benchmarking the spectrum of agent capabilities. arXiv preprint arXiv:2109.06780. 2021 Sep 14.
+
+[3]  Maximilian Igl, Kamil Ciosek, Yingzhen Li, Sebastian Tschiatschek, Cheng Zhang, Sam Devlin, and 492 Katja Hofmann. Generalization in reinforcement learning with selective noise injection and information 493 bottleneck. In Advances in Neural Information Processing Systems 32, pages 13956–13968, 2019.
+
+### Questions
+1. Please address the aforementioned weaknesses.
+
+2. Is it possible to utilize an RNN unit instead of frame stacking to incorporate temporal information into the latent code?
+
+3. Section A.2 and Figure 7 are not clear to me. Would you please explain the experimental setting there and the results? 
+
+4. How was the hyperparameter search done for all the baselines?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4

human_reviews/OqEsj4S240.md

@@ -0,0 +1,236 @@
+## Human Reviewer 1
+
+### Summary
+The paper proposes a new framework based on zigzag persistence to analyze internal
+representations in LLMs by characterizing the birth and death of topological features within the
+model’s layers.   it provides a fine-grained geometric analysis of the internal representations through TDA.  Distinct from traditional methods that solely compare representations at individual layers, the proposed method  captures their entire evolutionary path, providing a richer understanding of how these features evolve and contribute to the model’s decision-making processes
+
+### Strengths
+Originality:
+The idea of applying topological data analysis to LLM  is novel and interesting
+Clarity:
+The paper is well-written and easy to follow.
+Significance:
+The proposed Persistence Similarity and zigzag algorithm have rigorous mathematical formulation and are easy to compute. The experiment in Section 4.3 is promising. Author compares the performance of layer pruning Persistence Similarity and other common llm layer pruning methods and the proposed method performs well in most cases under 10% pruning rate.
+
+### Weaknesses
+Major concerns:
+1. The discussion in Section 4 did not clearly explain the sensitivity of the proposed method when pruning LLM layers.  The result shown in Fig 5 indicates that, across different models, our metric consistently identifies the deep layers as “redundant”. Additionally, in the results from Table 1, the proposed metric does not show an advantage at a 20% sparsity level, which further heightens my concerns. 
+
+2. The authors mention, "Due to the autoregressive nature of these models, the representation of the last token in a sequence captures information about the entire sequence and is the only token used for predicting the next." I disagree with this assumption; predicting the next token is clearly based on all preceding tokens. Therefore, I would like to see an analysis of certain special tokens, such as the EOS and BOS tokens.
+
+3.  Can Persistence Similarity and the zigzag algorithm help explain the varied responses of LLMs across different abilities or areas of knowledge? The pruning experiments presented in the paper are limited to base models. Could you provide an analysis of the instruction-following abilities of chat models or their capabilities in math and code?
+
+### Questions
+Overall, I believe this paper is likely to present a novel perspective on understanding LLMs. 
+However, the experimental section is overly limited. I look forward to discussing these questions further in the weaknesses section. Addressing them may help improve the overall score.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes a novel TDA-based method for measuring LLM similarity.  Zigzag-persistence is introduced to take account of the properties of the LLM graph structure. As part of the evaluation, it has been applied to Pruning to check its effectiveness.
+
+### Strengths
+- This paper introduces a new idea to introduce Zigzag-persistentce by considering the layer structure.
+- It has been applied to pruining to check its effectiveness.
+
+### Weaknesses
+- Methods that regard Neural Networks as graphs and apply TDA have been proposed in [1], for example. This is the usual persistent homology, but the use of zigzag persistence for graph structures has been verified in [2],[3] and elsewhere. The structure focusing on the LAYER structure has some aspects of novelty, but it seems to be a simple conventional combination idea and the degree of novelty is weak.
+
+[1] T. Lacombe et al., Topological Uncertainty: Monitoring trained neural networks through persistence of activation graphs, IJCAI2021
+
+[2] A. Myerset al., Temporal network analysis using zigzag persistence. EPJ Data Science, 12(1):6, 2023.
+
+[3] Y. Chen et al., Time-Conditioned Dances with Simplicial Complexes: Zigzag Filtration Curve based Supra-Hodge Convolution Networks for Time-series Forecasting NeuRIPS2022
+
+- It is difficult to see what the similarity of the models is intended to achieve. The proposed method is similarity of the graph structure and may be similar for different models. The paper evaluates the similarity in application to pruning to demonstrate its validity. However, it is difficult to accept from the experimental results alone that the proposed method is a superior similarity simply because the implications that the similarity indicates are not clear. As a method for measuring the degree of change, such as model modifications, such as pruning, it is intuitively effective. However, as no uses have been indicated, the current situation is that there is only evidence for it as a pruning method. If the claim is that it is a superior pruning method, it should be clearly stated as such. In addition, if the pruning method is slaughtered and claimed, it should be compared with a generally sufficient advanced puruning method, for example [4],[5].
+
+[4] S. Ashkboos et al., SliceGPT: Compress Large Language Models by Deleting Rows and Columns, ICLR2024
+
+[5] X. Ma et al.,LLM-Pruner: On the Structural Pruning of Large Language Models, NeuRIPS2023
+
+### Questions
+- Clarify the claim point(s) for which the novelty of the contribution in this paper is sufficient.
+- Please clearly indicate in what sense the similarity of the proposal is valid, what use scenarios are covered (other than pruning) and whether it is sufficient evidence for this.
+
+
+Why are the titles on the system different from the titles in the PDF?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper uses zigzag persistence to empirically study LLMs.
+
+### Strengths
+Section 3 rigorously introduced the zigzag persistence.
+
+### Weaknesses
+The major problem of this paper is the lack of a problem. In fact, the word "problem" was not found in the whole paper at all. 
+
+Lines 520-521: "Our approach aims to provide a high-level geometrical and topological description of positional and relational changes across layers".
+
+If this phrase in the conclusions should be considered a problem statement, then almost any empirical study using geometry and topology can fit this description. 
+
+Section 3 "Method" describes over 3 pages the known facts about zigzag persistence without even mentioning LLMs from the title of the paper.
+
+The world "layer" appears in the paper 50+ times without a proper definition. Initially, this word is used in the context of neural networks but section 3 seems to assume that a layer is a cloud of unordered points in R^d, which is a standard input for computing persistence in TDA.
+
+The "persistence similarity" introduced in (5) raises many concerns. First, the concept of "persistent similarity" appeared in the literature, e.g. https://www.intlpress.com/site/pub/pages/journals/items/cis/content/vols/0018/0004/a004/, but no past work on such similarity was cited.
+
+More importantly, the denominator in (5) can vanish, which makes the persistence similarity S_p undefined. In this definition, it is still unclear if l_1,l_2 denote layers or point clouds because min/max of l_1,l_2 seem to be real numbers. If a layer is indeed a point cloud, the first metric axiom surely fails for l_1 and its translated image l_2, which have the same persistence.
+
+Even if we consider point clouds under isometry, because persistence is an isometry invariant of a point cloud for standard filtrations of complexes, the first metric axiom also fails for infinitely many non-isometric point clouds, see J Applied Comp Topology 2024. Computational geometry has known much stronger isometry invariants of unordered point clouds for more than 20 years, see Boutin and Kemper, 2004.   
+ 
+Lines 402-403: "Note that the plot is not symmetric by definition (cfr. equation 5)". 
+
+Does it mean that the symmetry axiom fails? 
+
+The triangle axiom also seems unlikely to hold, so a proof is needed. If a distance fails the triangle axiom with any positive error, then results of clustering are not trustworthy as proved in https://ieeexplore.ieee.org/abstract/document/10574843? 
+
+The form on "soundness" asked to evaluate "the technical claims, experimental and research methodology and on whether the central claims of the paper are adequately supported with evidence".
+
+No technical claims were found, no words "claim", "proposition", "theorem" in the paper. Here are the comments on informal claims about contributions.
+
+Lines 74-76: "We propose a new metric to measure which topological features persist across the layers of an LLM."
+
+If persistence similarity is called a metric, proofs of metric axioms are expected. The original persistence already measures "which topological features persist".
+
+Lines 77-79: "By identifying layers with high persistence similarity, we prune redundant layers without significantly degrading performance".
+
+The adjective "redundant" appears in the paper without explanation. If performance refers to percentages in Table 1, they degrade by 10%. Is it not significant?
+
+Lines 80-82: "Our findings indicate that the behavior of persistent topological features and their similarities are consistent across different models, layers, and choices of hyperparameters of the framework".
+
+To make this claim meaningful, the words "consistent" and "consistency" should be properly defined.
+
+### Questions
+Lines 523-524: "This approach allows for effective model pruning by identifying and removing redundant layers without significantly compromising performance" 
+
+All accuracies in Table 1 are between about 40% and 70%. Are these numbers enough to guarantee safety in real applications?    
+
+How are outputs of LLMs converted into point clouds in the experiments? What is the meaning of persistence in terms of word tokens or other language-related concepts?
+
+Should the review mention the previous work explaining LLMs as stochastic parrots at https://dl.acm.org/doi/abs/10.1145/3442188.3445922 and in stronger terms at https://link.springer.com/article/10.1007/s10676-024-09775-5?
+
+### Soundness
+1
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper uses zigzag persistence, an approach from TDA, to analyze the hidden representations of LLMs. With these tools, they track the evolution of topological properties of LLM hidden representations across layers. They make numerical observations such as increased persistence of topological features formed in later layers, and high similarity between representations in later layers. As an application, they prune layers based on persistence similarity.
+
+### Strengths
+1. Generally good and clear explanation of the requisite background and method.
+2. Experiments show some insights into several open-weight language models of interest, across datasets that people care about (both for computing the metrics, and for testing downstream accuracy after pruning).
+
+### Weaknesses
+1. Besides the layer pruning application, the qualitative and quantitative insights from these topological features are not particularly interpretable or grounded. We can measure how the number of $k$-cycles evolves, but there are no interpretable quantities that come out of this, besides similarities between layers. To alleviate this, perhaps further discussion of motivation and related work (why should people use TDA tools to analyze internal representations), or interpretability work (what kind of 1-cycles form, and on what kind of data) could be interesting.
+2. Utility in one downstream application (layer pruning) is questionable (the simpler methods from prior work do solidly and are not consistently outperformed).
+3. The use of zigzag persistence in layer pruning only depends on the similarity matrix formed by the method. Other simple baselines could be considered for making similarity matrices.
+
+### Questions
+1. Have you looked at how topological features vary across different data samples / domains? For instance, are there any interesting data features that form between representations of tokens from certain programming languages?
+2. Could you provide a bit more information on where the embeddings are extracted from (this is currently described in Section 4.1 as "each prompt is processed so that the last token is extracted at each normalization layer and the final normalization applied to the output layer.") Do you take the embedding before or after the normalization layer? Also, does this mean you take two embeddings from each block (one from the attention normalization, one from the MLP normalization)? Also, some discussion on how this compares to how other works extract LLM hidden representations  would be nice.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 5
+
+### Summary
+This paper proposes a novel topological data analysis method based on zigzag persistence, which describes data undergoing dynamic transformations across layers within the large language models. The proposed persistence similarity is used to quantify the persistence and transformation of topological features throughout the model layers and provide insights to prune redundant architecture modules without significantly degrading performance. Experimental results indicate that the behavior of persistent topological features and their similarities are consistent across different models, layers, and choices of hyperparameters of the framework, suggesting a certain degree of universality in the topological structure of LLM representations.
+
+### Strengths
+- The paper is clearly written, making complex concepts accessible to the reader.
+- The application of zigzag filtration to the internal representation of LLM is novel and insightful. 
+- The approach to zigzag persistence is firmly rooted in theoretical foundations, providing a robust framework that enhances its applicability and relevance in the field.
+
+### Weaknesses
+- The zigzag filtration, and persistence images are previously well-defined and existing. This raises concerns about the originality and depth of the theoretical and technical contributions presented in the paper.
+- The performance gains over existing pruning methods, as shown in Table 1, appear marginal. Including a significance test would enhance the robustness of these findings.
+- While the paper provides a geometric interpretation of the topological features, it is not clear how these features can be directly interpreted in the context of language models. The paper could benefit from a more in-depth discussion on the interpretability and implications of the observed topological properties.
+- The zigzag algorithm is computationally expensive, especially for large datasets and high-dimensional representations. The paper should discuss strategies to optimize the algorithm's performance and make it more scalable.
+
+### Questions
+- Figure 5 is somewhat unclear. Are the 10% (orange) segments included within the 20% (yellow) segments in most cases? The authors might consider a more effective visualization method to enhance clarity.
+- Since other similarity measures can also characterize layer-wise behavior, what specific advantages does the proposed persistence similarity offer over these existing methods?
+- What is the computational complexity of the proposed persistence similarity in comparison to other existing methods? 
+- Can the authors provide insights or a high-level intuition regarding the consistent patterns observed in Figure 4?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+3

human_reviews/S1GTzTFKxb.md

@@ -0,0 +1,149 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes PromptTrack, a video-level RGB-T tracking paradigm via prompt learning (streaming temporal prompt and multimodal spatial prompt), achieving SOTA tracking performance and exhibits great scalability by extending to RGB-D and RGB-E tasks.
+
+### Strengths
+1. The reasonable usage of letters and formulas makes the entire text very readable and smooth.
+2. Benefiting from the effective spatial-temporal associations during multimodal interaction, PromptTrack learns target changes and motion trajectory from dense historical frame and behaves better in complex environments.
+3. In RGB-T/RGB-D/RGB-E benchmarks, the proposed method gains the advanced performance, verifying its effectiveness and versatility.
+
+### Weaknesses
+1. The ablation study about number of search images during training is missing. Does it increase would improve the temporal modeling and association capacity of the PromptTrack?
+2. It's better to have an ablation study about number of MSP blocks and trade-off between performance and MACs/Parameters.
+3. the "k" used in "comprising 40k samples" has a repeat appearance in the latter description like "k " template images. The authors should notice and improve this.
+
+### Questions
+1. The performace of OSTrack (RGB-T) in Table 6 of appendix.C is quite low, do you test it with fine-tuning or without any re-training?
+2. The performance on three RGB-T datasets is based on single-set model parameters or divided sets? And will codes and pre-trained models soon release?
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes a video-level RGB-T tracking paradigm called PromptTrack through prompt learning. It leverages temporal information from consecutive video frames to obtain temporal prompts, and learns multimodal spatial prompts conditioned on these temporal prompts, effectively utilizing complementary information from multiple modalities. A multimodal spatial prompt (SPG) module is inserted into the one-stream backbone network to enhance inter-modal interaction, while specific prompts are generated for each modality to capture modality-specific features. The current prompt tokens are stored in template memory for future use to enhance temporal information. The method has been evaluated on several benchmarks and achieves state-of-the-art performance.
+
+### Strengths
+1. Existing template search matching methods often utilize only spatial information for matching or merely introduce dynamic templates, neglecting the rich temporal cues present in consecutive video frames. The proposed PromptTrack establishes effective spatiotemporal associations by merging spatiotemporal flow prompts in the multimodal interaction process, enhancing robustness.  
+
+2. The motivation of the paper is clear, the experiments are comprehensive, and the results are convincing.
+
+### Weaknesses
+1. Various operations in Figure 3 should have annotations; otherwise, they may lead to ambiguity (e.g., it is not immediately clear whether the plus sign indicates concatenation or addition).  
+
+2. The authors mention one limitation of the method is the need for substantial storage resources to retain all historical templates, but they do not provide detailed information about the computational and memory costs during training and inference. More details about the computational and memory requirements during training and inference compared to baseline methods should be added.  
+
+3. The implementation details of the template memory are not sufficiently clear; it states that the maximum number of templates is four, but does not describe the size of the template memory or how the templates are updated. The paper should provide more detailed information about the template memory.  
+
+4. The concept of learnable prompt tokens is similar to the temporal token concept in previous works, such as ODTrack: Online Dense Temporal Token Learning for Visual Tracking, and what are the main differences between them?  
+
+5. In the experiments on RGBE and RGBD, the comparison trackers are limited to a few recent ones. To demonstrate the advance of the proposed method, I suggest more latest trackers should be added in comparison.
+
+### Questions
+Some details should be improved for clarity, and experimental comparison should be enhanced.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper proposes a spatial-temporal feature based RGBT tracking framework using prompt learning. The authors argue that current RGBT tracking methods do not make good use of temporal cues in videos, so they propose a streaming framework to continuously learn and utilize these features. The proposed prompt generation module can implicitly obtain the motion information in consecutive frames and propagate it to the next frame.
+
+### Strengths
+[1] The proposed prompt learning based temporal information learning model is an effective exploration of motion information integration for tracking.
+[2] The framework is simple and effective, and experimental results on several benchmarks show that it achieves state-of-the-art performacne.
+
+### Weaknesses
+[1] There is need to explain carefully why the propsoed prompt is suitable to learn motion features.
+[2] It is better to give a more detail comparison and analysis of how to intergrate the propsoed module in OSTrack.
+
+### Questions
+[1] How to get p_{0}  in Fig.2?
+[2] Why we need to use a subtraction to get TIR spatial prompt in Eq.9 ? Is this operation improtant?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper introduces PromptTrack, a novel video-level RGB-T tracking paradigm leveraging prompt learning to model spatiotemporal relationships in multimodal contexts. The authors argue that many existing methods inadequately utilize temporal information, often focusing on spatial aspects or introducing sparse temporal cues. In contrast, PromptTrack employs spatiotemporal prompts, allowing it to track target appearance changes and motion trajectories more effectively across video frames.
+
+By learning temporal prompts for each modality and then incorporating multimodal spatial prompts conditioned on these temporal prompts, the proposed method enhances the complementary use of multimodal data. 
+
+Experimental results show that PromptTrack achieves state-of-the-art performance on several benchmark datasets, including a Precision score of 76.2% and a Success score of 60.7% on the LasHeR dataset, while maintaining a real-time speed of 35 FPS.
+
+### Strengths
+1.PromptTrack captures target appearance changes and motion trajectories by incorporating streaming spatiotemporal prompts, resulting in more accurate and robust tracking across video frames.
+
+2.It achieves state-of-the-art tracking results while maintaining real-time speeds (35 FPS), demonstrating both high accuracy and practicality for real-world applications.
+
+### Weaknesses
+1.I believe the contribution of this paper is quite weak. It essentially adds an extra modality to ODTrack [1]. Regarding the claims made in its motivation, I even think it is almost indistinguishable from ODTrack. Moreover, like ODTrack, it also employs a template sampling strategy.
+
+2.The paper employs a memory mechanism. For a fairer comparison, I believe the authors should clearly indicate which papers use a memory mechanism and which do not. From what I know, the use of a memory mechanism can greatly enhance tracking performance, and ODTrack also benefits from its memory mechanism. However, adding a memory mechanism typically reduces FPS, so it is necessary to clearly list this information.
+
+[1] Zheng Y, Zhong B, Liang Q, et al. Odtrack: Online dense temporal token learning for visual tracking[C]//Proceedings of the AAAI Conference on Artificial Intelligence. 2024, 38(7): 7588-7596.
+
+### Questions
+1.The entire paper is built upon ODTrack, meaning the motivation proposed by the authors has already been addressed by others. What, then, is the authors' contribution?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+5

human_reviews/SThJXvucjQ.md

@@ -0,0 +1,133 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes algorithms for the conservative contextual bandit problem under a non-linear reward model.  The proposed algorithms apply IGW exploration based on an online regression oracle, enabling the selection of either an exploratory or a conservative policy based on a safety condition.  The proposed algorithms achieve sub-linear regret with respect to the total time step $T$, and its performance is supported by numerical experiments.
+
+### Strengths
+- The proposed algorithm introduces the first conservative contextual bandit algorithm for a general reward model by adapting the assumption of access to a regression oracle and leveraging the IGW algorithm, which previously applied only to linear reward models. The authors also present regret analysis for the algorithm. Although I was unable to rigorously verify all proofs, the results appear consistent, achieving a regret bound comparable to that of the linear case.  
+- The methodology is illustrated with examples using an online gradient descent regression oracle and feed-forward neural networks. Real-world data experiments further support the algorithm’s performance.
+
+### Weaknesses
+1. This paper introduces multiple algorithms and theoretical results for the conservative contextual bandit problem with a general non-linear cost function. Consequently, much of the main content is focused on the operation of the algorithms, assumptions required for the theoretical results, and descriptions of the outcomes, with limited discussion of the technical challenges arising from handling non-linear cost functions in CCB and how these challenges were addressed. It seems that Algorithm 1 and Algorithm 2 are almost identical aside from the use of different regression oracles. Moreover, if, as the authors state, Algorithm 2 has a tighter regret guarantee than Algorithm 1, it may be beneficial to focus more on Algorithm 2, detailing the challenges and technical novelties associated with it.
+
+2. The main theorem statements lack completeness. For example, in the statement of Theorem 3.1, there is no indication of how to set the input parameter $\gamma$ for the algorithm to achieve the stated regret bound. Similarly, Theorem 5.1 could benefit from more precise language rather than vague expressions like “appropriate choice of parameter”. Additionally, based on my understanding, the exploration parameter $\gamma_i$ in Algorithm 2 appears to depend on $\eta_i$ (line 1218). However, $\eta_i$ in turn depends on $L_i^*$, which is the true cost value of the optimal action and unknown to the agent. Further clarification is needed on how $\gamma_i$ is determined.
+
+3. The description for experiment reproducibility is lacking. Additional information, such as how the hyperparameters for each algorithm are set, would be helpful.
+
+### Questions
+1. In Algorithm 1, if the safety condition is not met, the baseline policy ($b_t$) is used, and its noiseless cost $h(x_{t,b_t})$ is observed. However, this data is not used in the oracle. Since the baseline policy provides high-quality data with zero noise, why is it not utilized? Would using this data yield a better estimator? Additionally, in the linear cost case (Kazerouni et al., 2017), a noisy reward is observed for actions chosen by the baseline policy (e.g., $y_{t,b_t}$). What would happen if, in this paper’s setting, the algorithm observed $y_{t,b_t}$ instead of $h(x_{t,b_t})$?
+
+2. The algorithm introduced in Section 4 is referred to as having a “first-order regret bound.” Why is it called “first-order”? How would second-order and third-order bounds be defined?
+
+3. When defining the Neural Tangent Kernel (NTK) matrix, is $x_t$ the feature chosen by the algorithm at time $t$? If so, it seems unclear how the NTK can be defined given that the chosen features are unknown prior to starting the algorithm. It may be better to define the NTK matrix based on the context of all actions at all times, as in Zhou et al. (2020). Additionally, the regret bounds in Theorems 5.1 and 5.2 are independent on $\lambda_0$. Why is this assumption necessary, and does it not affect the regret bound?
+
+4. Most literature on Neural Tangent Kernel-based neural bandits shows a dependency of regret on the effective dimension of the NTK. In contrast, the proposed paper presents a dependency on $K$ instead of the effective dimension. Could you explain what enables this distinction?
+
+5. In the Neural Conservative Bandit, the activation function in the first-order bound changed from a smooth & Lipschitz function $\phi$ to a sigmoid function $\sigma$. What are the differences between $\phi$ and $\sigma$?
+
+6. There do not appear to be terms related to the perturbation constant or the number of ensemble $S$ in the regret bound. Do these quantities not affect the regret? Additionally, how are these hyperparameters chosen in the experiments, and how do they impact performance?
+
+---
+
+**[Typo]**
+
+- line 295: “Line 7” should be “Line 8.”
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper studies the conservative contextual bandit problem, where the goal is to minimize the regret as in classical contextual bandit problems and, with high probability, be competitive against a baseline policy. Importantly, the cost functions are not assumed to be linear in the contexts. This extends previous study of conservative bandit problems beyond the multi-armed setting and the contextual linear setting. 
+
+The two proposed algorithms, C-SquareCB and C-FastCB, combine the inverse gap weighting based exploration and online regression oracles (e.g. function approximation using neural network+ gradient descent), and achieve regret which is sublinear in the horizon T or sublinear in the optimal loss L*, while being (1+alpha)-competitive against the baseline with high probability. Experiments using real world data show that C-SquareCB and C-FastCB have smaller regret compared to algorithms for contextual linear bandit problems, and are more competitive against the baselines.
+
+### Strengths
+The paper is overall well written, with clearly presented problem formulations, algorithms and results. The setup considered fills in the gap of current conservative bandit literature, and the proposed algorithms have provably sublinear (in T or L*) regret while being (1+alpha) competitive against the baseline. Experiments further demonstrate their superior performance as compared to algorithms designed for conservative linear contextual bandits and for classical settings without baselines.
+
+### Weaknesses
+The proofs/assumptions may lack rigor. In particular, the proofs cite results in other works without carefully checking the assumptions under which those results hold. For instance, in line 914-915, lemma 2 in Foster & Rakhlin (2020) is invoked. If my understanding is correct, that lemma requires Assumption 3 to hold for all possible sequences. Nevertheless, in line 1736-1737, Assumption 3 is only proved to hold with high probability. 
+
+One contribution of the work is to use neural network for function approximation to deal with the non-linearity in the cost functions. However, to achieve the stated performance as in Theorem 5.2, the width (and thus the number of parameters) of the neural network is Omega(poly(T)), which might be too large for long-horizon problems, making the algorithms less practical. 
+
+The paper could benefit from more detailed discussion on the significance of the regret bounds in Theorem 3.1 and Theorem 4.1. In particular, it appears that for any algorithm which has regret upper bounded by alpha * y_l * t, under Assumption 2, equation (2) is automatically satisfied. Some discussions on the range of parameters might help the readers better understand the significance of the results.
+
+### Questions
+In Definition 2.2, line 124, alpha is chosen to be <1. I’m wondering if this is just a simplifying assumption, or there is difficulty in extending the results of this paper to settings where alpha>=1? 
+
+In line 160, what is H in the ``inf’’?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This authors consider Conservative Contextual Bandits with general value function class and develop two algorithms C-SquareCB and C-FastCB, using Inverse Gap Weighting and online regression oracles. They show that the safety constraint is satisfied with high probability (the performance is not worse than a baseline policy by more than $(1+\alpha)$ factor). They also show the regret for C-SquareCB is $\tilde O(\sqrt{KT}+K/\alpha)$ and the regret for C-FastCB is $\tilde O(\sqrt{KL^*}+K/\alpha)$, where $L^*$ is the cumulative loss of the optimal
+policy. The efficacy of the proposed algorithms is validated on real-world data.
+
+### Strengths
+- The paper is clearly written and mostly easy to follow, with proof roadmap and intuition moderately provided.
+- The provided solution nicely connects safe conservative bandits and contextual bandits with general functions class.
+- Analysis is sound and rigorous. Numerical experiment is convincing.
+
+### Weaknesses
+As a paper that combines two established sub-fields in bandits, it is a bit unclear the novelty in algorithmic design and theoretical analysis. I would like to see authors provide and emphasize more detailed discussions if possible. In particular, what is your technical/methodological contribution compared to Kazerouni et al. (2017), Foster & Rakhlin (2020), Foster & Krishnamurthy (2021)? What are the challenges of adapting/extending their tools? From what I understand, the novelty appears in: a delicate safety criterion and a time-varying exploration schedule. The authors can add more if necessary.
+
+### Questions
+Some clarification questions apart from the concern in Weaknesses:
+
+- Data-dependent bound
+  - When $L^*=O(1)$, it seems the regret bound becomes $\tilde O(\ln T)$. Is there any instance-dependent metric (similar to the mean gap in the stochastic MAB setting) hiding in the constant term?
+  - $\gamma_t$ seems to be requiring the knowledge of not only $L^*$ but also each component that sums up to $L^*$. Would that be too strong?
+
+- Experiments
+  - Can you provide more details on how you set $\delta$ and tune hyperparameters? How do you run C-FastCB if the optimal cost is not known?
+
+- Typos:
+  - Algorithm 2 Line 8: Is there a term related with $\sqrt{\log\delta^{-1}}$ missing?
+  - Line 836 missing a sqrt term consisting of $m_{\tau-1}$ and $\texttt{Reg}$.
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3

human_reviews/TH4gKbZS1E.md

@@ -0,0 +1,149 @@
+## Human Reviewer 1
+
+### Summary
+This work empirically compares Kolmogorov-Arnold Networks with Multi-Layer Perceptron on
+learning irregular or noisy functions. The experiment results show that KAN do not always perform
+the best.
+
+### Strengths
+Experiment codes are provided for reproducibility.
+
+Do provide some insight on what KAN may be good at modeling.
+
+### Weaknesses
+The finding is purely empirical.
+
+The paper does not clearly state the experiment setting in the main text.
+
+The experiment does not provide conclusive results.
+
+The experiment only tries to fit relatively simple functions. The result may not be relevant to real-world problems.
+
+### Questions
+It is possible to include more challenging problems for comparison? It is well established that MLP can model fairly complicated functions.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+1
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+Authors compare the performance of Kolmogorov-Arnold Networks (KAN) and Multi-Layer Perceptron (MLP) networks on irregular or noisy functions. The author experimentally demonstrated that KAN does not always outperform MLP.
+
+### Strengths
+- The author compared KAN and MLP on various irregular and noisy functions and experimentally demonstrated in which cases KAN is worse than MLP.
+
+### Weaknesses
+- The author merely compared KAN and MLP experimentally but did not analyze why KAN or MLP performs poorly in certain situations.
+
+- The author experimentally demonstrated that KAN is sometimes inferior to MLP. It would be better to propose a new, improved KAN model to address this.
+
+- There are no experiments on high-dimensional functions. In one dimension, both KAN and MLP are likely to approximate well to some extent, but more experiments are needed to explore how they perform in high-dimensional spaces with irregular points.
+
+- If the experiments are conducted only on univariate functions, many models besides MLP can be compared with KAN. It would be beneficial to include other models commonly used in machine learning in the experiments.
+
+### Questions
+I do not have a complete understanding of KAN, but I think KAN appears to be a generalization of projection pursuit regression. Is this correct?
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper conducts a comparative analysis of experiments between MLP and KANs, discussing the outcomes. It challenges the assumption that KANs consistently outperform MLP in modeling mathematical equations, particularly with irregular functions. The experiments involve applying MLP and KAN to various functions—regular, non-differentiable, discontinuous, singular, and coherent oscillation, with and without noise. These functions are single input and single output. Variations include different training sample sizes, iteration counts, and optimizers. The findings demonstrate that KANs do not always surpass MLP.
+
+While this paper serves as a great exploration to KANs and does establish that KANs are not invariably superior to MLP, it falls short by only providing experimental evidence without introducing new theoretical insights or network structures, thus lacking substantial academic contribution.
+
+### Strengths
+The structure of the paper is clear and well-organized.
+The experimental results are clearly presented.
+The experiments validate that KANs are not consistently superior to MLP.
+
+### Weaknesses
+The experiments could be designed more targeted. For instance, in the experiments for non-differentiability, both functions feature only a single non-differentiable point. A comparison between functions with single versus multiple non-differentiable points would be more insightful, given the focus on the impact of these points.
+
+The discussion lacks depth. Given the simplicity of both the functions and network structures used, there is potential for a more detailed examination of how parameters are trained and the reasons behind specific outcomes.
+
+The discussion section does not yield any intriguing or unexpected conclusions, nor does it propose any novel theories or structures.
+
+### Questions
+Given that the structures of both MLP and KANs are well-known, a deeper analysis of their capabilities and limitations in the related work section would be beneficial. More thorough research could uncover more significant findings. For instance, some limitations of KANs identified in the paper are not due to the Kolmogorov-Arnold Theorem but rather due to B-spline, a critical component in KANs that is not discussed in the paper at all.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+In this empirical study, the authors compare the performance of two ad-hoc versions of KAN and MLPs, both with identical parameter counts, in learning ten single-dimensional real functions, with and without added noise.
+
+### Strengths
+The work has considered different classes of functions with some common irregularities. The empirical comparison is sound, and extensive plots provided let's the reader to compare the performance of tested KAN and MLPs in each case.
+
+### Weaknesses
+The comparison primarily centers on the resulting test accuracy curves; however, it lacks the necessary theoretical justification and fundamental analysis to substantiate the findings.
+
+While the authors have structured the text well, the plots are somewhat cluttered and could be presented more effectively for better clarity. 
+
+Overall, the work appears basic and does not demonstrate the level of novelty typically expected from submissions to ICLR.
+
+### Questions
+If authors can provide some theoretical insights backing the observed empirical results in all or some of the function classes tested, it would make the work more promising and considerable for this conference.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+4

human_reviews/Vlo3Gad3YP.md

@@ -0,0 +1,169 @@
+## Human Reviewer 1
+
+### Summary
+In this submission, the authors tackle the challenge of black-box optimization. At present, the bulk of the field has concentrated on methods like Bayesian Optimization, which employ a _forward_ model $p(y|\mathbf{x})$ and an acquisition function $\alpha(\mathbf{x})$. Maximizing the latter at each iteration $t$ yields a candidate design $\mathbf{x}_t$ that will in turn give a noisy evaluation $y_t$.
+
+Interestingly, here, the authors use an _inverse_ model $p(\mathbf{x}|y)$ together with an acquisition function $\tilde{\alpha}(y)$. As such, one can learn to generate designs $\mathbf{x}$ conditionally to an output value $y$, such that the evaluated design results in a noisy evaluation $y_t \approx y$. The precise value of $y$ that should be used for conditional generation is obtained by maximizing the acquisition function $\tilde{\alpha}$.
+
+In this work, conditional generation is achieved through conditional diffusion models. While diffusion models have been used in black box optimization previously, building a suitable acquisition function to handle these models has not been done yet. The latter must adequately trade-off exploitation (high values $y$) and exploitation (values $y$ for which uncertainty model uncertainty $p(\mathbf{x}|y)$ is high). The paper's main contribution is a thorough study of uncertainty quantification for conditional diffusion models, leading to a decomposition between _aleatoric_ and _epistemic_ uncertainty. This ultimately leads to an acquisition function that trades off high function values and epistemic uncertainty.
+
+Finally, the performance of the proposed acquisition function is theoretically grounded depending on some assumptions on the target function $f$, and the method itself is shown to outperform concurrent baselines on a number of continuous and discrete datasets.
+
+### Strengths
+- I found the paper to be well-organized and motivated. The technical novelty is light but seems to be theoretically grounded. 
+
+- The proposed method consistently ranks among the best competitors on a benchmark involving multiple baselines and datasets, while maintaining computation times in the same order of magnitude as Gaussian Process-based alternatives.
+
+### Weaknesses
+- From a practical point of view, I am not sure that Theorems 2 and 3 are useful: they assume the $L$-smoothness of the function $f$. This function takes as input $\x$, which might be discrete, or an embedding of a discrete input like a molecule, and there is little to no chance to have $L$-smoothness in this case unless the embedding explicitly enforces that assumption. I believe this should at least be mentioned.
+
+### Questions
+- An interesting ablation study would be to add a factor weight $\beta$ in front of the epistemic uncertainty in Equation 8, and to vary this term. This would give an insight into how important uncertainty is in the acquisition process.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes a novel approach for solving black-box optimization problems. Unlike traditional methods that focus on learning a surrogate to evaluate design decision quality, this approach employs a diffusion model to approximate the distribution within the design space conditioned on a target value. Unlike existing inverse methods that assume access to an offline dataset, this paper studies a dynamic setting where the exploration-exploitation tradeoff must be considered. To address this challenge, an uncertainty-aware exploration method is introduced. The effectiveness of the proposed approach is demonstrated through extensive numerical studies.
+
+### Strengths
+- Overall the paper was very well-written. The key concepts and challenges are clearly introduced, which makes it easy for me to appreciate the contribution.
+    
+- Black-box optimization is an important methodology that has a wide range of applications in engineering and science. 
+
+- The numerical studies are comprehensive and convincing.
+
+### Weaknesses
+- **Soundness of the theoretical analysis**. While I appreciate the authors efforts to justify the proposed approach through a theoretical lens, I found some of the results unsatisfying. 
+    
+  - For example, in Theorem 2, the bound does not depend on the number of samples collected in each iteration $N$. Intuitively, a large $N$ might lead to over-conservative estimates and a small $N$ might renders the estimate too optimistic. Relating the this bound to $N$ may lead to insights into the choice of this important hyper-parameter. The current bound is independent of $N$, suggesting that it might be loose. Furthermore, this theorem assumes the existence of a perfect diffusion model. I suggest the authors add further discussion on the implication/validity of this assumption. This comment applies to Theorem 3 as well.
+  - Under a similar Lipschitz assumption, is it possible to derive surrogate approximation guarantees for the forward-based approach? If so, how does the inverse bound compare to the forward bounds?
+
+
+- **Acquisition function.** This is a minor point, but I was wondering if a weight should be assigned to  $\Delta$ in Equation (8) because (1) the two terms might be in different scales, and (2) the users may dynamically adjust the weight across different iterations to balance exploration and exploitation.
+
+### Questions
+See weaknesses.
+
+### Soundness
+2
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+In traditional BBO a surrogate model $\hat{f}$ is learned to approximate the objective function and then helps optimize an acquisition function which yields the next $x_{k+1}$ where to evaluate $f$. To select the next query point $x_{k+1}$ Diff-BBO instead performs posterior inference in the parameter space $X$ conditioned on a specified target objective value $y$ obtained by maximising an introduced Uncertainty-aware Exploration (UaE) acquisition function.
+
+Conditional diffusion model are trained to learn the conditional distribution $p(x | y)$, where $x$ represents feasible inputs in the parameter space.
+
+$y$ is chosen with a proposed acquisition function, Uncertainty-aware Exploration (UaE), that prioritizes target values $y$ with high expected objective values while minimizing epistemic uncertainty. This acquisition function balances exploration and exploitation. The paper provides theoretical proofs demonstrating that UaE achieves a near-optimal solution for the BBO problem.
+
+Numerical experiments to support the work are presented
+
+### Strengths
+The paper introduces a new approach supported by solid theoretical results and empirical validation across diverse tasks. Difference with existing methods is clearly presented.
+
+### Weaknesses
+An important part of the procedure is how is assembled the candidate set $\mathcal{Y}$ and its corresponding weights $w$. The paper does not specify a principled method for choosing or tuning these weights, which makes this important aspect somewhat empirical given that for too high weights, the model may focus excessively on unfeasibly high values of $y$ while too low weights might limit the search space.
+
+
+A diffusion model requires a large dataset to effectively learn the data manifold in the design space. If the function $f$ is expensive to evaluate, building a large dataset may be computationally expensive. Most of the experiments are run with a relatively high number of evaluations. How would the method perform on a smaller dataset?
+
+Diffusion models are usually susceptible to mode collapse, where generated samples fail to cover the full distribution of the data. Was this observed? This could cause Diff-BBO to overlook potentially optimal regions in the design space.
+
+### Questions
+Is there a systematic approach for choosing the weights $w$ for the candidate set $\mathcal{Y}$, or is this step largely empirical?
+
+Could the method perform effectively with a smaller accumulated dataset, as opposed to the relatively high number of evaluations used in the experiments?
+
+Was this issue of mode collapse observed in Diff-BBO? If so, how does it impact the model’s ability to explore potentially optimal regions in the design space?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This work proposes an inverse approach leveraging diffusion models for online BBO problem. Specifically, this paper introduces a new acquisition function to propose objective function values, and employ a conditional diffusion model to generate samples. The authors conduct experiments in design-bench to verify the effectiveness of their method.
+
+### Strengths
+1. This paper is easy to follow. The structure of this paper is clear.
+2. This paper provides a solution for online black-box optimization (BBO) called Diffusion-based inverse modeling for black-box optimization (DIFF-BBO), which uses objective function values to generate solutions. This approach is interesting and has the potential to improve the performance of online BBO.
+
+### Weaknesses
+1. The core difference between the main idea of the proposed method (the inverse method) and LLAMBO [1] should be explicitly discussed.  While there are differences in the specific implementation details compared to LLAMBO, this approach appears to be more of an aggregation of the previous methods. Specifically, the inverse approach was originally utilized in MINS [2], while the conditional diffusion model was incorporated in DDOM [3].
+
+2. Why do the experimental tasks of online BBO methods use offline BBO benchmarks? Is it because there is no suitable benchmark for online BBO?
+
+3. While the background on online BBO is quite solid, the paper does not sufficiently explore prior work in offline BBO. Expanding the discussion to include them could provide a more comprehensive literature review and motivation support.
+
+4. Superconductor, Ant and D’Kitty and so on are high dimensional problem. So, it would be better if this paper could compare with more high dimensional Bayesian optimization in recent years rather than simple black-box optimization.
+
+[1] Large Language Models to Enhance Bayesian Optimization. In Proceeding of the 12th International Conference on Learning Representations, Vienna, Austria, 2024.
+
+[2] Model inversion networks for model-based optimization. In Advances in Neural Information Processing Systems 33, pp. 5126–5137, Virtual, 2020.
+
+[3] Diffusion models for black box optimization. In Proceedings of the 40th International Conference on Machine Learning, pp. 17842–17857, Honolulu, HI, 2023.
+
+### Questions
+Please see Weaknesses part.
+
+Besides, “They struggle with steering clear of out-of-distribution and invalid inputs” is often discussed in offline BBO instead of in online BBO. The reason why this paper presents it here for online setting should be discussed.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+5

human_reviews/VxIetsMu3G.md

@@ -0,0 +1,178 @@
+## Human Reviewer 1
+
+### Summary
+This theoretical paper presents new contributions aimed at enhancing our understanding of how pre-training schemes improve fine-tuning performance. It introduces some theoretical results about SimCLR, one of the most popular contrastive learning methods for vision tasks. The paper demonstrates that, under specific conditions regarding the amount of labeled data, SimCLR pre-training coupled with supervised fine-tuning can achieve nearly optimal test loss. The main theoretical results are presented in Theorem 4.2 and 4.3. The paper and its appendix contain the proofs of these theorems. The paper also contains many novel analysis tools that enable the study of the SimCLR algorithm. The theoretical results are valid in a specific setting (simple binary classification task and a two-layers ConvNet), which is not a standard setting and may limit the potential impact of the contributions.
+
+### Strengths
+- This paper introduces new theoretical contributions to understand the advantage of SimCLR pre-training in fine-tuning stage. The paper demonstrates that, under specific conditions regarding the amount of labeled data, SimCLR pre-training coupled with supervised fine-tuning can achieve nearly optimal test loss for two-layers ConvNets. Under certain conditions related to the quantity of labeled and unlabeled data, as well as the signal-to-noise ratio (SNR), the convergence of training loss and a low test loss are assured. These theoretical results indicate that SimCLR pre-training during the fine-tuning stage can reduce label complexity, leading to a lower test loss.
+- The paper and its appendix contain the proofs of the theorems 4.2 and 4.3, which are the main results of this paper.
+- The paper contains many novel analysis tools that enable the study of the SimCLR algorithm. 
+- The appendix contains some experiments on synthetic and MNIST datasets to confirm the theoretical results.
+
+### Weaknesses
+**Impact of the contributions.** The paper studied a very specific setting: simple binary classification task and a two-layers ConvNet with $RELU^q$ output activation. In practice, this setting is not used often. For instance, SimCLR paper uses a ResNet-50 and was used on multiple multiclass classification benchmarks. There is a big gap between the theoretical setting used in this paper and the practical setting, so it is difficult to understand the potential impact of this paper contributions. It may be difficult to do, but it would be interesting to relax some of the assumptions to be able to get theoretical for a family of models which include ResNet-50. 
+
+
+**Lack of justification/motivation.** The paper has many assumptions that are not well motivated, or claims that are not well justified. Here are a few examples:
+- L162: "this data model is particularly suitable to study SimCLR, which is originally proposed for vision tasks." This claim appears to lack sufficient justification. It may be beneficial to include an analysis to support this claim.
+- L153: The patches are represented as vector rather than matrix in Definition 3.1. It seems to ignore the 2d nature of the patches. It could be valuable to justify this choice. 
+- L215: The paper focuses on two-layers ConvNet with $ReLU^q$ activation function with q > 2. This type of activation is not the most popular choice so it could be interesting to explain this choice. 
+- The assumptions regarding data augmentation for pre-training are not clearly stated.
+- Definition 3.1 does not explain the assumptions on $\mu$ 
+
+
+**Clarity.** The presentation and clarity of the paper could be enhanced. The paper includes a considerable number of notations, which can make it challenging to keep track of them all. Some notations appear to be undefined, while others are introduced later in the text, making it difficult to fully understand them until the subsequent paragraphs are read. Here are a few examples:
+- $\Omega$ symbol is used L84, but it is defined at L106
+- $W$ is used at L171 but it does not seem to be defined. 
+- $w_r$ is used at L172 but it is defined at L196
+- $[F(W, x)]_r$ is used at L172 but it does not seem to be defined. 
+- $[n_0]$ is used L184 but it does not seem to be defined. 
+- L200: some variables like the unlabeled dataset are used to compute the loss, but they are not shown as input: $L(W, S_{unlabeled})$
+- L158: The third bullet point in Definition 3.1 is not clear. It looks like a something is missing.
+
+
+**Things to improve the paper that did not impact the score:**
+- A lot of equations do not have number so it is difficult to reference them in the review.
+
+### Questions
+The questions are arranged in order of importance, with the first question being the most important.
+
+- The paper should discuss the scope of the contributions, as the theoretical setting seems quite different from the practical setting. In particular, is it possible to remove or relax some assumptions to include a larger family of models? 
+- L279: "we remark that most of these assumptions are non-essential." It seems to be counter-intuitive to make some assumptions that are non-essential. How do the theoretical results change if these non-essential assumptions are removed? 
+- What are the assumptions about the data augmentation? 
+- Why does the paper focus $ReLU^q$ activation function for the theoretical analysis? Is it possible to include other activation functions? 
+- What are the assumptions about $\mu$? Can $\mu$ be the zero/null vector?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 2
+
+### Summary
+In this paper, the authors present a theoretical analysis to establish the benefits of SimCLR pretraining for downstream finetuning of classification tasks. Their analysis is based on several assumptions, including isotropic Gaussian distributions of the representations, the signal-to-noise ratio (SNR) of the training and testing data as well as number of labeled data. Focusing on a two-layer convolutional neural network (CNN) setting, the authors provide detailed proofs and explanations to support their main argument: SimCLR pretraining provably guarantees convergence of the training loss and enhances test-time robustness.
+
+### Strengths
+The paper is clearly and coherently written, allowing readers to easily follow the presented ideas. Despite being based on several simplified assumptions, the analysis and proofs provided are detailed and reasonable, making the work theoretically sound in the setting of a two-layer CNN.
+
+### Weaknesses
+There are several weaknesses in the paper that require clarification from the authors:
+
+1) The paper aims to understand how SSL methods like SimCLR benefit downstream fine-tuning. However, the literature review provided is insufficient to comprehensively cover this topic, which has been widely studied both theoretically and empirically. A more thorough literature review on SSL's impact on fine-tuning performance is necessary.
+2) The problem setting of 2-layer CNN presented appears overly simplified, potentially deviating from the practical deep learning problems that commonly employ SSL methods such as SimCLR. While deriving theoretical guarantees for modern architectures like transformers is challenging, a more realistic framing of the problem would better justify its relevance and importance.
+3) The experiments in the appendix seem questionable in terms of their validity and relevance to the stated problem. Typically, SimCLR pretraining leverages large-scale unlabeled datasets. However, the experiments employ only 250 and 200 unlabeled data points for the synthetic data experiment and real data experiment, respectively. These small dataset sizes are atypical for SimCLR pretraining and may not represent the true behaviors of SimCLR pretraining combined with downstream fine-tuning tasks.
+
+If the authors provide large-scale experiment results for the real data experiment and explain how the theorems can be applied to more modern backbone architectures, I would be willing to raise the score.
+
+### Questions
+1. How would the theorem transfer to more modern backbone architectures such as ViT? Can you provide more experiment ablation on more architectures?
+2. Can you provide more ablation on the SNR rate and the number of the unlabeld data / labeled data?
+3. What happens if the representation does not follow isotropic Gaussian distribution? Typically SimCLR type of loss results in non-isotropic representations, in that case, does the theorem still hold just by simply replacing the variance term in the SNR equation with the maximum variance across different embedding dimensions?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper extends a prior analysis of Cao et al. (2022) (regarding supervised learning) with a particular focus of SimCLR self-supervised learning. Specifically, it revisits a toy binary classification problem designed by Cao et al. (2022), and consider a SimCLR training using a bespoke 2-layer convolutional neural network. As the result, the paper theoretically show that fine-tuning a SimCLR pre-trained two-layer convolutional neural network requires significantly less label complexity compared to the supervised counterpart (of Cao et al. (2022)) to achieve low training/test losses, in a sense that the complexity is independent to the signal-to-noise ratio of the data at fine-tuning stage provided that it is sufficiently accounted at the pre-training stage. Although the main body of the paper is focused on theoretical results, some supporting experimental results are also provided, e.g., on synthetic data and MNIST, in Appendix.  
+
+Cao et al., Benign overfitting in two-layer convolutional neural networks, NeurIPS 2022.
+
+### Strengths
+- The paper is well-motivated and clearly written.
+- The paper presents solid theoretical results, which indeed provide insight about SimCLR.
+- The area of study, i.e., theoretical understanding of recent contrastive learning methods, is still a timely topic in my opinion.
+- Although not present in the main text, the paper also provides empirical supports as well.
+
+### Weaknesses
+- To my knowledge, there have been several works that study contrastive learning in theoretical aspects, although I could not find a discussion about them from the paper: e.g., [Bansal et al., 2021; HaoChen et al., 2021; Tan et al., 2024]. I think the paper should incorporate any discussions about such a line of research. For example, I am curious whether the paper’s observation about SimCLR as a matrix power method can be related to the spectral view of contrastive learning [HaoChen et al., 2021].
+- I generally feel a lack of discussion about how the theory presented in this paper can be related to the real-world use of SimCLR. I think this is because the paper omits enough context about how can we really ensure that the simplifying assumptions made here are mild enough, so that it can be eventually connected to the real-world.
+- The main body of the paper is quite theory-biased, and allocates its significant portion with the proof sketch. This may narrow down the potential audience, and I personally think the authors might consider to give more highlight on their empirical results as well as the theoretical ones.
+
+Bansal et al., For self-supervised learning, Rationality implies generalization, provably, ICLR 2021.
+
+HaoChen et al., Provable Guarantees for Self-Supervised Deep Learning with Spectral Contrastive Loss, NeurIPS 2021.
+
+Tan et al., Contrastive Learning Is Spectral Clustering On Similarity Graph, ICLR 2024.
+
+### Questions
+- Is there any potential that the analysis presented in this paper can be generalized to more complex architecture?
+- The overall setup assumes that the labeled data at fine-tuning stage is sampled from the same distribution where the pre-training data is sampled, considering a pure unsupervised learning scenario. But one important aspect of SimCLR like training is in its transferrability; as such, could the analysis can be generalized in any way if we relax the assumption of the labeled data to another distribution? For example, I am curious if the analysis could be extended when the labeled data is more noisy that the pre-training data.
+- In my understanding, Theorem 4.3 implies that one still require the unlabeled sample size for SimCLR, $n_0$, at the similar rate of those for supervised learning to optimize the loss. But could this result be due to that the data considered is too easy to discriminate, even without labels, and perhaps the complexity of SimCLR can be much worse under a more challenging data assumption?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+5
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper presents a theoretical analysis of SimCLR pretraining, focusing on training a two-layer convolutional neural network to learn a signal-noise model, as explored in recent works like Cao et al., 2022. The authors derive conditions under which SimCLR pretraining and finetuning can achieve near-optimal test loss. Their findings indicate that SimCLR pretraining, followed by finetuning, can reach nearly optimal error rates with fewer labeled data than required for supervised learning.
+
+### Strengths
+- The theoretical results and proof techniques are novel. In particular, the connection between SimCLR and the matrix power method is especially interesting.
+
+### Weaknesses
+- Unlike other works focused on similar data distributions and network architectures, such as Cao et al., 2022, and Kou et al., 2023, the authors provide only a sufficient condition on the amount of unlabeled data, whereas previous studies offer tight necessary and sufficient conditions. I agree that providing only a sufficient amount of pretraining data effectively demonstrates SimCLR’s advantage over supervised learning. However, I believe that characterizing tight necessary and sufficient conditions would be valuable, as it may facilitate comparisons with other pretraining methods in future work.
+- The data augmentation techniques introduced in the problem setting is toounrealistic. While I understand that defining data augmentation in an abstract manner within a signal-noise model is necessary, changing only the noise patch to new i.i.d. noise seems overly simplistic. This approach leads to augmented data from different pretraining samples with the same (unseen) label following the same distribution, which differs from practical scenarios. I believe a more practical definition for data augmentation could be used—such as applying a linear transformation to the noise patch.
+- There are limited practical insights provided regarding the theoretical findings (see Question 1).
+
+### Questions
+- Could you provide some practical insights into the theoretical findings? Specifically, can the main proof technique—interpreting SimCLR as a matrix power method and characterizing top eigenvectors—be translated to real scenarios to offer intuition on why SimCLR is effective?
+- Can the same proof technique be extended to ReLU networks and the 0-1 test loss considered in Kou et al.2023?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+4
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/XCugWIuHR8.md

@@ -0,0 +1,145 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces Convex Distillation, a model compression method that replaces the non-convex layers of deep neural networks with convex approximations. By leveraging convex optimization, the method achieves efficient compression without the need for labeled data or post-compression fine-tuning.
+
+### Strengths
+- Paper is written well and easy to follow.
+
+- The idea of bridging convex optimization, distillation and compression is interesting.
+
+### Weaknesses
+- Practical contributions of convex optimization in model compression are limited. The convexity conversation is only valid and tested up to 3-layer DNNs. It significantly restricts the objective landscape. For simple tasks, it might be fine, while for more complex tasks, it often leads sub-optimal performance. 
+
+- Experimental results are not satisfactory to justify the efficacy of the proposed methods. Only small datasets are included. Meanwhile, the ResNet18 baseline seems not well tuned (with low acc less than 90%).
+
+### Questions
+See the weakness.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This article proposes a novel distillation technique that efficiently compresses deep neural network models through convex optimization. This method eliminates intermediate non-convex activation functions and uses only the intermediate activations of the original model, enabling distillation without the need for labeled data, and achieving comparable performance to the original model without fine-tuning. Experimental results show that this approach not only maintains model performance when compressing image classification models on multiple standard datasets but also performs better and optimizes faster compared to traditional non-convex distillation methods. This work opens up new avenues for future research at the intersection of convex optimization and deep learning.
+
+### Strengths
+-  The paper introduces a novel approach to knowledge distillation that leverages convex optimization for efficient compression of deep neural networks. 
+
+- The authors provide extensive empirical evidence to support their claims, demonstrating the effectiveness of their method across multiple standard datasets and in various scenarios.
+
+- The author provided Google Colab code with very detailed experimental instructions.
+
+### Weaknesses
+- Using existing convex neural network packages, there is a lack of originality and workload.
+
+- There is an issue with the network configuration. For datasets with small image sizes like CIFAR-10, the configuration used for ResNet on ImageNet should not be applied. It should not downsample by 4x from the start, which results in feature maps that are too small.
+
+- The experiments were only conducted on small datasets and very small networks. Can they be scaled up to larger datasets such as ImageNet?
+
+### Questions
+See Weakness, is there a more reasonable network configuration, more comprehensive experiments?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes a convex distillation method by combining the representational power of large non-convex DNNs with the favorable optimization landscape of convex NNs. The proposed method can distill the models in a label-free manner without requiring post-compression fine-tuning on the training data. Experiments on several image classification datasets show that convex student models can achieve high compression rates without sacrificing accuracy and outperform non-convex compression methods in low-sample and high-compression regimes.
+
+### Strengths
+1.A simple yet effective to distill classification models via convex networks
+
+2.An effective distillation acceleration tool and polishing are used to improve convex solver.
+
+### Weaknesses
+1.Activation matching is not novel for knowledge distillation.
+
+2.Experimental comparison is not sufficient to support the effectiveness of the proposed method. It lacks SOTA KD methods for fair comparison.
+
+3. It is not clear that how to distill all blocks. If the proposed convex distillation performs block-wise distillation, it requires a complex and time-consuming knowledge distillation for handling the entire networks.
+
+### Questions
+See the weaknesses.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper presents a new distillation method that efficiently compresses the model by convex optimization-eliminating intermediate non-convex activation function.
+
+### Strengths
+This manuscript is very clear about the background knowledge and the motivation for undertaking this work is clear.
+
+### Weaknesses
+1.	In the section on related work, there is a lack of information on the most recent work, and the related work is introduced too little.
+2.	The notation in Eq. 1 and Eq. 3 is used incorrectly ($D_i \in {D}' \in \mathcal{D}_x$).
+3.	Some of the textual content in the figures is too small.
+4.	The innovative content of the article is not sufficient.
+5.	In the experimental part, there is a lack of validation results on large datasets such as ImageNet. Also, using only ResNet18 and MobileNet V3 for experiments is not convincing enough.
+6.	The results in Fig. 4 do not intuitively show the superiority of the proposed approach.
+7.	There is a lack of experiments to compare with other methods, only ablation experiments are performed.
+
+### Questions
+1. Why “while DNNs have the capacity to memorize the training dataset, they often end up learning basic solutions that generalize well to test datasets” can “motivate using smaller, compressed models”?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4

human_reviews/dD6b5RREws.md

@@ -0,0 +1,170 @@
+## Human Reviewer 1
+
+### Summary
+The paper investigates the impact of the bootstrap rate (BR) greater than 1.0 on Random Forest (RF) performance, concluding that higher BR values may lead to statistically significant improvements. The paper is purely empirical. I have a few concerns primarily regarding the lack of insight and the generalizability of the findings.
+
+### Strengths
+Extensive experimental studies
+
+### Weaknesses
+1. Lack of Insight: The study would greatly benefit from a discussion on the underlying rationale behind the observed improvements in classification accuracy with BR > 1. Although the (very limited) empirical results suggest that higher BR values could potentially improve performance, the paper does not offer insights into why this might be the case. Introducing a simplified, theoretical case, such as a linear Gaussian model, could help illustrate the mechanisms at play and provide a foundation for these empirical observations. This addition would make the findings more interpretable and provide value beyond experimental results.
+
+2. Experimental Setup and Generalizability: While the authors employ a diverse dataset collection, the setup appears somewhat arbitrary, raising questions about the generalizability of the results. For example, RF performance is known to vary significantly with the number and complexity of categorical variables, which can greatly influence tree-based methods. Controlling for these factors or providing more context on dataset selection criteria would help clarify the scope of the findings and allow for a better assessment of their relevance to broader cases.
+
+3. Classifier for BR Selection: The approach to developing a classifier to predict optimal BR usage lacks a solid rationale. The selection of 18 RF configurations and 10 BR values per dataset does not align with the principles of cross-validation, where data should be permutable to make the validation results meaningful/generalizable. This choice weakens the validity of the classification model and its potential for reliable application across diverse datasets.
+
+### Questions
+None
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+1
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper explores the use of bootstrap sampling in Random Forests, specifically focusing on the bootstrap rate (BR), which is the ratio of the number of observations in each bootstrap sample to the total number of training instances. While traditional methods use a BR of 1 (sampling equal to the original dataset size), the authors investigate the effects of higher BR values, ranging from 1.2 to 5.0, on classification accuracy. Additionally, the authors create a binary classifier to predict whether the optimal BR for a given dataset is less than or equal to 1 or greater than 1.
+
+### Strengths
+1)	Overall, this work re-examines the bootstrap rate in random forests, considering a wide range of BR values and multiple datasets, which throw light on focusing some basic machine learning configs.
+
+2)	Sufficient experiments on multiple datasets are carried out, which involves various aspect on BR values, such as the relationship between BR and classification accuracy, the influence of different RF configurations, and the dependence of the optimal BR on the dataset.
+
+### Weaknesses
+1)	Although different BR values have been extensively studied, the explanations for why certain datasets exhibit specific behaviors at specific BR values may not be deep and comprehensive enough, especially for some complex patterns and anomalies. For example, for a model like RF (nf all) that exhibits special behaviors, although some hypotheses have been proposed, more research may be needed to understand the exact mechanism behind it.
+
+2)	When analyzing the relationship between the optimal BR value and the dataset, although some attributes and characteristics of the dataset are taken into account, there may be other undiscovered factors that affect the selection of the optimal BR value, which may play an important role in different datasets and application scenarios.
+
+3)	The article does not analyze time performance related issues in detail. In practical applications, different BR values and model configurations may have a significant impact on training and prediction time, which may be a direction for future research.
+
+### Questions
+1)	Relationship between BR value and classification accuracy should be supported by more theoretical analysis and illustrated in a more rigorous form.
+
+2)	It is presumed that “RF(nf all) will perform well on datasets with a high proportion of insignificant or less significant features, as it may avoid building trees primarily based on these features”. Could you provide some stronger evidence or more in-depth analysis?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This article investigates the impact of the bootstrap rate (BR) on the predictive performances of random forests. BR is a hyperparameter of random forests: each tree of a random forests are built by drawing with replacement BR*n observations among the n original ones. 
+Previous works have tested the impact of BR values ranging from 0.2 to 1.2. In this paper, the authors compare the performance of various RF configurations (varying the number of trees, maximal depth, number of features randomly selected in each cell...) together with different BR values ranging from 0.2 to 5. 
+
+Experimental results over 36 open source data sets show that a BR larger than one is associated to good predictive performances for most data sets. The authors explain these results by introducing a statistics measure over a data set, corresponding to the inhomogeneity of the data set. More precisely, the parameter $k\_l$ stands for the number of observations in the data set that have $l$ observations with the given label among their $k$ nearest neighbors. The authors demonstrate that the optimal BR is correlated to this measure. 
+
+Finally, the authors also introduce two binary classifiers that predict if a given data set has an optimal BR greater or lower than 1. These classifiers are respectively trained on 36 and 24 data sets.
+
+### Strengths
+The paper studies the impact of BR on random forests performances and provide insights about the characteristics of data sets for which BR >1 should be preferred. Extensive simulations are done to corroborate the proposed statements.
+
+### Weaknesses
+- Several related works are missing. Besides, it is not clear from the related work section which contributions are theoretical or practical. See for example: 
+    - For empirical performances, see Section 4 in Random Forests for Big Data https://arxiv.org/abs/1511.08327
+    - For theoretical results quantifying the uncertainty of random forests, depending on BR, see https://jmlr.org/papers/volume17/14-168/14-168.pdf or https://arxiv.org/pdf/1405.0352
+    - In https://www.esaim-ps.org/articles/ps/pdf/2018/01/ps170099.pdf, the authors provide the expression of the optimal subsampling rate for median forests
+    - Regarding Breiman's forests, the analysis of https://arxiv.org/pdf/2006.06998 takes into account bootstrap
+
+- Experiments are done for classification data sets only. The paper would be strengthened if experiments were done also for regression problems. 
+
+I have several concerns related to the design of experiments, and more specifically with Table 1: 
+- No standard deviations are displayed. Thus, we are not sure if the best method is really the best or if many other RF configurations have performances close to the best one. 
+- Looking at Table 2 in Martinez-Munoz and Suarez (2010), which displays the average test error for different BR ratios, the standard deviations are quite important. For example, the standard deviation of the error (and thus the accuracy) is between 7 and 8 for the audio data set, whereas you display the accuracy with 3 decimals. 
+- Besides, the test set is used for two different objectives: selecting the best RF configuration and computing its error. Thus, the error of the best RF configuration is rather optimistic. It would be more sound to evaluate the performance on a different data set (or to use the out-of-bag error of the forest). Thus, t-test results do not seem conclusive. Besides, as you consider several t-test, you should probably correct the significance level $\alpha$ accordingly, taking into account multiple testing issues.
+- Given the performances displayed in Table 1, I would like to see the same experiment run with a Baseline RF whose number of trees is set to 500. Indeed, this parameter seems to be the most important one, whereas in practice it is fairly easy to tune (the higher the better in terms of accuracy). 
+- A Table with all values of RF configuration / BR rate with standard deviations would give more precise information. 
+- l.335 The fact that optimal performances are reached for two very different values of BR (0.2 and 5) by slightly changing the data set is worrying. One reason can be that the two values are good for the two data sets, which is in contradiction with your study, showing that BR >1 is better than the opposite. This is why metrics for all configurations with standard deviations should be displayed.
+
+### Questions
+- l.47 "we expect 36.8% of observations to be absent in each bootstrap sample." We know that the probability for one observation to be absent of a bootstrap sample is 0.368. However, I am not sure that this corresponds to the percentage of observations absent from the bootstrap sample, as the events $B_i$ ("presence of the $i$th observation in the bootstrap sample") are not independent. Can you prove the statement or give a reference? 
+- Looking at the paper of Martinez-Munoz and Suarez (2010), it appears that they do not use split randomization in their random forests implementation. This could explain the difference between their results and the results obtained in the present paper. This should be clearly mentioned somewhere. 
+- l.98 what is the rationale behind standardizing one-hot encoded features? Tree based methods do not usually need input standardization to work. 
+- The ranges of values for the different hyperparameters are not consistent: the max depth varies between 10 and 25 (which results in leaf containing between n/2^10 and n/2^25 observations if the tree is balanced) while the minimum number of observations per leaf varies between 2 and 5. These ranges of values should be harmonized.  
+- A twofold cross-validation is performed 200 times. Usually, a 5-fold cross validation is performed. Is there any reason for choosing this value? Can you display standard deviation averaged across runs? 
+- A BR value much larger than one would lead to use almost all observations for each tree construction. This would probably damage the computation of the out-of-bag error, which is an important feature of random forests, as it allows us not to divide the data set into a train and a test set to compute the RF error. Could you study how the out-of-bag error is influenced by the BR value?  
+- l.206 "Finally, BR= 1, defined in the original formulation of the bootstrapping procedure and the most frequently used
+value, performed relatively poorly." This is not shown by the experiments: the only things shown here with respect to BR=1 is that it is often not the best configuration. But it can be the second-best most of the time. 
+- l.340 notation $k\_l$ is uncommon and misleading as it counts a number of observations. I would rename it $n_{k,l}$ for clarity.
+- Table 2. What is the rationale for studying very low values of $k$ and $l$ with respect to the data set size? Choosing a fix number of nearest neighbors $k$ for all data sets corresponds to considering neighborhood of different sizes across data sets. Thus, I wonder if such a measure $k\_l$ corresponds to a characteristic of the data set.
+- l.431 Regarding the bootstrap rate prediction, I am not sure to understand the final methodology. More precisely, I do not understand how features are selected: "based on the absolute value of the Spearman rank-order coefficient, calculated separately for each run on the training instances". Does it mean that, at first, 12,000 features are computed, then the correlation with the best BR is computed, and only the ten best are kept? Since computing the best BR requires a validation set, performances should be computed on a test set. Besides, does the procedure require running random forest with all BR values (in order to compute the best ten values among the 12,000 original features) in order to build a classifier to choose the best BR value? 
+- The resulting classifier predict whether the best BR is larger than one. More practically, I would be more interested in estimating the best BR. How does your analysis help the practitioners?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+1
+
+### Rating
+1
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper studies the effects of the Bootstrap Rate (BR) parameter on the test set accuracy of random forest classifiers. The authors claim that it is possible to obtain improvements (not always) by setting BR > 1, which is rarely done in practice. The authors develop a classifier to predict if the optimal BR is larger than 1 based on the dataset.
+
+### Strengths
+* Good scope of study in terms of datasets and covering a wide range of hyperparameters
+* Exploration of the effects of a bootstrap rate that is generally understudied
+* The idea of determining the BR based on data statistics is interesting
+
+### Weaknesses
+* The analysis carried out in table 1 is not explained in a clear enough way. For instance, the paired t-test is not fully described and is hard to understand. For a result listed as a key contribution in the paper, the procedure should be explicitly described.
+* Figure two does not include error bars, and in almost all cases the effect of the BR parameter seems very small in terms of accuracy.  
+* The classifier only predicts if the optimal BR (a term that need to be formally defined) is larger than 1. It would be more useful for practitioners to predict the actual rate.
+* Generally I feel as if the findings in this study align well with the hypothesis that BR > 1 is as good, but not better, than BR=1. When making a claim that the optimal BR is equal to a certain value, the authors rely on finite sample analysis, which might be susceptible to random fluctuations.
+
+### Questions
+* By "2-fold stratified cross-validation, repeated 200 times, was applied, yielding 400 results", do the authors mean that they did a 50/50 (train/test) split and ran two experiments for each configuration?
+* Can the authors please fully describe their hypothesis testing procedure carried out in Table 1? 
+* Can the authors please formally define the optimal BR?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+4

human_reviews/ed75tWzgt0.md

@@ -0,0 +1,176 @@
+## Human Reviewer 1
+
+### Summary
+This work introduces a theoretical two-player zero-sum game framework for reinforcement learning from human feedback (RLHF), where the max-player approximates the Nash equilibrium and the min-player approximates the best response by each choosing their own optimistic rewards. The theoretical algorithm achieves $\tilde{O}(\sqrt{T})$ regret in linear environments. It also proposes two practical algorithms: a two-agent algorithm TANPO and a single-agent algorithm SADPO to approximate TANPO. These algorithms outperform several baseline methods in experiments.
+
+### Strengths
+1. This paper gives a theoretical sound result for self-play style RLHF algorithm, achieving a sub-linear regret in linear two-player zero-sum games.
+2. The algorithms have better performance than baselines as shown in experiments.
+
+### Weaknesses
+1. The discussion on data diversity is not persuasive enough.
+2. The motivation of using a single-agent approximation is not discussed.
+3. As usually Nash learning results do not assume BT model or other transitive preferences, this work only focuses on BT model, lacking discussions on non-transitive preferences.
+4. It would be better if the authors can plot a curve of regret to show that it is $\tilde{O} (\sqrt{T})$ (even if in toy environments).
+
+### Questions
+1. On top of page 4, why is $\mathcal{R}$ defined on two actions instead of one action as in Equation (3)?
+2. Is the final step of Equation (7) $V_r (\pi^t, \mu)$ instead of $V_r (\pi^t, \dagger)$?
+3. What is the sampling strategy of online DPO?
+4. Regarding data diversity: Why does larger $|\log \pi_\text{ref} (a^1) - \log \pi_\text{ref} (a^2)|$ reflect higher data diversity? In my opinion, diversity is about the **overall distribution** of the sampled responses (more like $D_{KL} (\pi_\text{ref} (\cdot | \text{TANPO}) || \pi_\text{ref} (\cdot | \text{online DPO}))$), instead of the difference in a **local pair**.
+5. Regarding single-agent approximation: What is the motivation of using a single-agent approximation? Is it more time and space efficient than two-player version?
+6. Previous Nash learning results do not assume transitive preferences. Can your results applied to non-transitive preferences?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper studies provably efficient self-play training algorithms of RLHF, which aims to 1) derive theoretical guarantees for the self-play framework and 2) improve with active exploration. The authors propose both an easy-to-implement two-agent algorithm and its single-agent version. Numerical results are provided to demonstrate the effectiveness of the proposed algorithms.
+
+### Strengths
+The proposed algorithms show improvements to the baselines in the numerical experiments.
+
+### Weaknesses
+1. The self-play framework is motivated by general preference models, while this paper limits to the standard BT reward model;
+2. Confusion and (elementary) mathematical typos appear in critical definitions and analysis, indicating the paper's current version may not be ready for publication (see Questions below).
+
+### Questions
+1. Policies $\pi$ and $\mu$ are fully decoupled in the value function (3), and the minimax value is always zero, which makes the use of Nash equilibrium as a solution concept quite confusing;
+2. There are typos in deriving the second equations in both (5) and (7). For instance, $\arg\max_\pi\min_\mu V_{\hat{r}_t^1}(\pi,\mu)$ in (5) is always zero while the RHS is non-zero. Additionally, there is a typo in Eq.(12).
+3. A minor suggestion on writing: In Section 2, I think it would be better to introduce "Two-Agent Zero-Sum Games" in the context of the self-play training framework.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes a new self-play RLHF algorithm. It effectively balances the trade-off between exploration and exploitation.  In TANPO, two players are trained using different loss functions to ensure more diverse and informative data collection. And, in SADPO,  a single-agent approximation of  TANPO, which is supported by both theoretical analysis and empirical evidence. Experiments are conducted on multiple evaluation benchmarks, including AlpacaEval 2.0, MT Bench, and PairRM. It is a good paper.
+
+### Strengths
+This paper proposes a new self-play RLHF algorithm. It effectively balances the trade-off between exploration and exploitation.  In TANPO, two players are trained using different loss functions to ensure more diverse and informative data collection. And, in SADPO,  a single-agent approximation of  TANPO, which is supported by both theoretical analysis and empirical evidence. Experiments are conducted on multiple evaluation benchmarks, including AlpacaEval 2.0, MT Bench, and PairRM.
+
+### Weaknesses
+No
+
+### Questions
+No
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The submission investigates a game theoretic approach to RLHF to increase data diversity.
+
+### Strengths
+RLHF is an important problem.
+
+---
+
+Review summary: I'm not an expert in RLHF, so I'll defer to other reviewers for the strength of the submission along that axis. But as far as its game-theoretic approach, the submission seems a bit confused, as articulated in the weaknesses section. Still, RLHF is an empirical field, so the its empirical results (about which I am not qualified to speak) shouldn't be underweighted.
+
+### Weaknesses
+> the max-player aims to maximize the summation of (i) the expected Nash equilibrium value function and (ii) the negative estimation loss of that reward function. Similarly, the min-player seeks to maximize the summation of (i) the expected best response value function based on the max-player’s strategy and (ii) the negative estimation loss of that reward function.
+
+These sentences doesn't make sense:
+1. Value functions are already expected values.
+2. Both the Nash equilibrium value function and the best response value function are a unique functions mapping from decision points to values. They are not an objective that a player can maximize.
+3. The submission's use the language "that reward function", indicating a reference back to a previously mentioned reward function. But there is no such previously mentioned reward function.
+
+---
+
+As a aesthetic matter, "Two-Agent Nash Policy Optimization" is a pretty tasteless name. There is a large literature of policy-based algorithms for computing Nash equilibria in zero-sum games.
+
+---
+
+> In the Nash equilibrium, the max-player’s strategy and the min-player’s strategy are mutual best responses, meaning each is optimal given the strategy of the other (Nash et al., 1950).
+
+-> given the other
+
+---
+
+> where V (π, µ) is a general function that captures the payoffs based on the strategies π and µ.
+
+"general" is superfluous here.
+
+---
+
+The submission's usage of regret, while not wrong, is a bit atypical for games. Rather than quantifying a player's external regret over the iterations actually played, it quantifies that player's exploitability. Under this usage of regret, the relationship between sublinearity and convergence to Nash is less fundamental than it is games literature. Specifically, while in games literature, sublinearity is necessary for average iterate convergence, here it is just quantifying convergence speed.
+
+---
+
+I don't understand the motivation of setting up the equation (3) as a zero-sum game. The objective the submission describes is additively factorable across players, so you can just drop all the min terms and maximize to find the Nash equilibrium.
+
+---
+
+> To simplify this setup, we propose the Single-Agent Diversity driven Policy Optimization (SADPO) algorithm as a single-agent approximation of TANPO. The SADPO optimization objective is similar to min-player objective (13).
+
+If we just care about data diversity, why bother with a game theoretic setup in the first place?
+
+---
+
+I don't understand Table 3. Why are there min players and max players on both axes?
+
+### Questions
+I included some in the weaknesses section above.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+3

human_reviews/esf4Lduba2.md

@@ -0,0 +1,170 @@
+## Human Reviewer 1
+
+### Summary
+The paper introduces FIMP, a framework that adapts pretrained non-textual foundation models for graph-based tasks. FIMP repurposes existing pretrained non-textual foundation models for message-passing on graphs. FIMP is evaluated across diverse domains, such as image classification, spatial transcriptomics, and fMRI brain activity recordings, utilizing state-of-the-art foundation models like ViTs, scGPT, and BrainLM. It demonstrates competitive performance in a zero-shot setting for embedding image networks, achieving notable results without task-specific retraining.
+
+### Strengths
+1. The concept of leveraging pretrained foundation models for message passing in GNNs is innovative. This approach shows how pretrained knowledge can inform finer-grained interactions in graph structures.
+2. The suggested method achieves improvements over the discussed baselines.
+
+### Weaknesses
+1. The paper is not easy to follow, with several concepts and sections lacking clarity (see questions below).
+2. While the idea of integrating pretrained foundation models with MPNNs is interesting, the proposed approach is overly simplistic (a straightforward approach -- if I understood correctly –- see questions below). Simplicity in methodology is not inherently problematic in my opinion; however, the combination of simplicity with a lack of a clear motivation for addressing the problem raises concerns — see Weakness 4 for further details.
+3. The experimental results are not entirely convincing, as the baselines used for comparison (e.g., GCN, GAT) are somewhat outdated. It would strengthen the paper to include comparisons with more recent state-of-the-art graph learning (e.g.,  GRIT [1]) methods that could offer a more rigorous benchmark for evaluating FIMP's performance.
+4. In my opinion, the paper lacks sufficient motivation. The tasks presented appear to be crafted primarily to emphasize FIMP’s strengths, rather than addressing meaningful, real-world applications. If I understand correctly, the authors construct a graph from publicly available datasets (e.g., the Mapillary image dataset). For example, line 173 states: “images form a geographical proximity graph.” However, this implies that the method is evaluated exclusively on datasets where the graphs were constructed by the authors, and these graphs are not publicly accessible.
+To strengthen the paper, the authors can either (1) provide a compelling rationale for how such a setting could arise naturally in real-world scenarios, or (2) identify and utilize publicly available datasets that align with their setup.
+
+### Questions
+1. Lines 191-193: "By aligning the tokenization in FIMP with the tokenization scheme of pretrained foundation models, we reduce distribution shift in token representation when applying these models to graph-structured data." Does this mean that you are directly adopting the tokenization strategy of the pretrained foundation model? Clarification on how closely you follow the original tokenization would be helpful.
+2. In Algorithm 1, which of the weights are learned from scratch, and which are (possibly) borrowed from the pretrained foundation model? A clearer distinction between learned and reused weights would improve understanding.
+3. How does FIMP-base specifically leverage the knowledge embedded in a foundation model? From what I understand, given lines 251-253: “In its base formulation, cross-attention message passing can be done with a simple cross-attention mechanism which is learned from scratch during training. We denote this base version of our architecture as FIMP-base in our experiments.”, the knowledge transfer in this case occurs primarily through the tokenization process. 
+4. On the FIMP versions that rely on a foundation model’s weights in the cross attention, It seems unclear how effective using the pretrained weights of a foundation model would be, given that cross-node attention differs from what the foundation model encountered during training (e.g., ViTs typically see only single images per training step). Could you elaborate on how the cross-node attention mechanism adapts to this discrepancy? Or did it just appear to work well in practice?
+
+**References**:
+
+[1] Graph Inductive Biases in Transformers without Message Passing. Ma et. al. 2023
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces a message-passing framework that leverages pretrained non-textual foundation models for graph-based tasks. The proposed method, FIMP, aligns node tokenization in Graph Neural Networks with tokenization schemes used by foundation models, allowing for more granular feature-level interactions during message passing. The authors demonstrate FIMP's effectiveness across diverse domains, including image networks, single-cell RNA sequencing, and fMRI brain activity recordings, showing improvements over strong baselines and highlighting the potential of repurposing non-textual foundation models for graph tasks.
+
+### Strengths
+1. The expression is very fluent. Readers can easily understand the content of the work.
+
+1. The study of graph foundation models and large graph models is very meaningful.
+
+1. The paper demostrates FIMP outperforms traditonal GNN models on image networks, single-cell RNA sequencing, and fMRI brain activity recordings. FIMP can effectively leverage SOTA foundation models in graph tasks.
+
+### Weaknesses
+1. The algorithm proposed in this paper appears to be incremental. Although the paper discusses the differences from works like GAT, in essence, it merely transforms node features from vectors($x_v \in \mathbb{R}^d$) to matrices($x_v \in \mathbb{R}^{f \times d}$).
+
+2. The current experiments are limited to image networks, single-cell RNA sequencing, and fMRI brain activity. It is unclear whether this method is effective in other domains, especially traditional graph datasets. I hope to see experiments conducted on larger practical datasets like OGB.
+
+3. As shown in Table 1, the gene expression prediction results on the mouse hippocampus and human heart spatial transcriptomics datasets, FIMP + scGPT performs much better than FIMP-base and FIMP + ViT. This indicates that to unleash the potential of the FIMP algorithm, a suitable and powerful foundation model is needed. However, in many scenarios, there may not be powerful foundation models available. Pretraining a suitable foundation model from scratch is also expensive and challenging. This limits the practical application of FIMP.
+
+### Questions
+The practical application of traditional GNNs is hampered by a significant computational burden. As shown in Figure 6, the time cost of FIMP is twice that of traditional GNNs. This concerns me about the practical value of FIMP. I would like to see a comparison of F1-Score and consumed time between FIMP + ViT and the ViT model alone on image classification dataset. I also hope to see the training time expenditure of FIMP on larger graph datasets.
+
+In summary, if the authors can address my concerns, I would be open to increase my score.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper incorporates pre-trained LM for cross-node message creation for a non-textual graph neural network. To this end, it tokenizes each node's content into a sequence and employs self-attention between sequences from pairs of nodes. Evaluations on genomics, brain, and street-view images demonstrate its effectiveness.
+
+### Strengths
+1. The writing and organization are good. The presentation is easy to follow.2. 
+2. The proposed methods can be applied to many fields. 
+3. Experimental evaluations from different fields demonstrate its effectiveness.
+
+### Weaknesses
+1. The novelty is very limited. The proposed method is the extension of attention in GAT to non-textual node content. This aims at jointly considering source and target nodes for message passing. However, this is not novel.
+2. The proposed method lacks clear motivation. It is direct for non-textual node content. Thus, the significance is weak. It is not clear why the cross-node attention can improve the performance. 
+3. The evaluations are not convincing. Only the GNN baselines are compared. Besides, it is not clear how the node attribute is constructed for GNNs. How about using the embedding from LM as the node’s attribute?
+4. Figure 3 lacks a description in the main body.
+
+### Questions
+See weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The authors introduce Foundation-Informed Message Passing (FIMP), a novel message-passing framework designed to leverage pretrained non-textual foundation models for graph neural networks (GNNs). 
+Unlike existing GNNs, which utilize a single embedding for nodes, FIMP represents nodes as sequences of feature tokens. 
+This enables cross-node attention-based message-passing using the self-attention layers from pretrained models. 
+The proposed method is evaluated on tasks across image networks, spatial transcriptomics, and fMRI brain activity recordings, showing significant improvements over baseline GNNs. 
+Additionally, FIMP demonstrates zero-shot capabilities on image classification tasks using pretrained ViT.
+
+### Strengths
+I think it seems like the authors are trying hard to emphasize how FIMP is fundamentally different from existing attention-based GNNs like GATs and graph transformers. The statement with boldface might indicate that they have faced challenges in communicating the novelty of their work or distinguishing it from prior approaches in earlier submissions.
+**However, in my opinion, this research is both timely and important because it aligns well with the current era of foundation models dominating various domains like NLP, computer vision, and even biomedicine.**
+Existing foundation models excel with unstructured data (like text and images), but applying them directly to graph-structured data is challenging. 
+FIMP addresses this gap by allowing foundation models pretrained on unstructured data to be repurposed for graph-based tasks through its innovative tokenization and message-passing approach.
+The authors provide comprehensive evaluations across diverse datasets, demonstrating that FIMP outperforms existing GNN models like GCN, GraphSAGE, and GATs. Notably, FIMP with pretrained models (e.g., scGPT and BrainLM) achieves the best results in gene expression prediction and brain activity reconstruction.
+The zero-shot embedding experiments on the Mapillary image dataset highlight FIMP's ability to leverage pretrained foundation models without additional training, which is a significant advantage in scenarios where labeled graph data is scarce.
+
+### Weaknesses
+- While I acknowledge the novelty and potential impact of the research direction taken in this work, I must also express some concerns. The proposed module, despite its creative approach (direction), appears somewhat simplistic in its current form. It would benefit greatly from a deeper theoretical analysis to clarify why the architecture is effective and how it fundamentally differs from existing methods. This would not only strengthen the authors’ claims but also provide a clearer understanding of the model’s underlying principles.
+
+
+
+- I noticed that while the paper benchmarks against various GNNs, it does not include comparisons with several main baselines (non-foundation models) that are well-established within each specific domain. Is there a particular reason why these comparisons were omitted? Including these baselines would provide a more comprehensive evaluation and help validate the effectiveness of FIMP beyond the GNN-specific context.
+
+
+- I would like to address the reporting of experimental results. The performance metrics are averaged over five runs, but it is unclear if this number of repetitions is statistically sufficient to ensure robust conclusions, especially given the variability in results. For example, in the Mouse Embryo cell type classification task, the reported standard deviations appear quite significant. Increasing the number of runs or providing a justification for why five is adequate would enhance the reliability of the reported outcomes.
+
+
+- (minor) Additionally, regarding the reference to Graph Transformer Networks (GTNs) on line 803, I believe there might be a mischaracterization of the original work by Yun et al. (2019). The GTNs discussed in that paper are not graph transformers in the sense typically understood in this context. Instead, they are better described as a graph-adapted version of Spatial Transformer Networks [1]. As stated by Yun et al.:
+> Yun et al. (2019): "GTNs can be viewed as a graph analogue of Spatial Transformer Networks which explicitly learn spatial transformations of input images or features."
+
+[1] M. Jaderberg, K. Simonyan, A. Zisserman, et al. Spatial transformer networks. In Advances in neural information processing systems, pages 2017–2025, 2015.
+
+### Questions
+please see the above section.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+4
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/ga4LyaucKr.md

@@ -0,0 +1,191 @@
+## Human Reviewer 1
+
+### Summary
+This paper uses the MenuNet (Shen et al., 2018) idea to optimize a general quasi-linear objective over multiple players and multiple items. In particular, the feasible allocation set X is the product of the feasible allocation set for each individual player i, which implies that there is no constraint on the total supply of each item (each item could potentially allocate to multiple players).
+
+The results are compared with other mechanisms empirically.
+
+### Strengths
+Unless I misunderstood the feasible allocation set, I didn’t see much strength.
+
+### Weaknesses
+The problem being solved in this paper is a trivial extension of the MenuNet (Shen et al., 2018) to multiplayer setting with individual-wise allocation constraints. The key challenge of extending MenuNet to general multi-bidder setting is to handle the feasibility constraint properly (i.e., each item can only be allocated to at most one bidder). 
+
+This challenge is referred to as “menu compatibility” and first solved by GemNet (Wang et al, 2024b), who solve the compatibility issue through a combination of a price adjustment and MIP (mixed integer program).
+
+This paper, however, drops the only challenge of generalizing MenuNet to the multi-bidder setting. So I cannot see any real contribution to the literature (unless I misunderstood this part).
+
+### Questions
+Please properly mention the key result of MenuNet in your second paragraph, and explicitly compare your approach with theirs.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+1
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+There has been a lot of recent research on designing revenue-optimal, strategy-proof auctions through the use of neural networks and machine learning tools. However, existing approaches often fall short of meeting all desired properties: exact truthfulness, expressiveness, and efficiency. For instance, RegretNet does not ensure exact truthfulness, AMA mechanisms lack expressive power, and MenuNet can be computationally inefficient. This paper presents PFM-Net, a framework designed to address all three objectives through a learning-based approach.
+
+The authors propose a full-menu mechanism that uses neural networks to parameterize pricing functions—determining how much agents are charged for specific bundles based on their valuations. This framework incorporates insights from economic theory, such as agent independence, convexity and monotonicity, to achieve incentive compatibility and a no-buy-no-pay rule to satisfy individual rationality. The optimization is an alternating process: first, allocations are optimized for the players given fixed pricing functions; then, the neural network parameters are adjusted to optimize revenue for the auctioneer.
+
+The framework is initially evaluated in a single-bidder, multiple-goods setting. It is adaptable to other objectives, such as social welfare maximization. To demonstrate this flexibility, the authors include an experiment with a social planner setting involving multiple agents and multiple goods.
+
+### Strengths
+S1. PFM-Net leverages insights from economic theory, including agent independence, convexity, and monotonicity, to ensure incentive compatibility and no-buy no-pay rule to satisfy individual rationality. These designs seem to be an improvement over architectures presented in RegretNet
+
+S2. Avoidance of Explicit Menu Enumeration  
+Traditional menu-based mechanisms often require enumerating all possible menu options, which can be computationally prohibitive, especially as the number of items grows. For instance, even a deterministic mechanism for a single buyer with $m$ items would need $2^m$ menu options, creating scalability challenges. PFM-Net, however, avoids this by not requiring explicit enumeration of the menu. For each auction instance, it optimizes the agent's objective directly based on the pricing functions. This means allocations are determined dynamically by maximizing the agent’s utility under the current pricing function, allowing the model to handle large settings without incurring the overhead of menu enumeration.
+
+### Weaknesses
+W1. Missing Baselines  
+RochetNet, the current state-of-the-art for single-buyer settings, should be included as a benchmark, as other methods generally reduce to RochetNet in this context, making it a sufficient point of comparison. Additionally, the optimal mechanism for up to six items is given by the SJA mechanism (referenced in the paper as Giannakopoulos and Koutsoupias). Please include this under OPT for $S_5$. This mechanism can be extended for larger m using a recursive formula, with results available up to m = 10 in the RegretNet paper, where it is also conjectured to be optimal. This makes SJA an essential baseline for evaluating the proposed approach.
+
+Moreover, it would be interesting to test the model’s performance in a setting with a single additive bidder and two items, where the bidder's values are independently drawn from a Beta distribution (α=1, β=2). Prior work [2] has shown that the optimal mechanism in this setup involves an infinitely sized menu, providing a valuable test case. Additionally, including settings where randomization is essential would show how well this approach performs for non-deterministic settings.
+
+W2. Lack of Moderate/Large-Scale Experiments  
+The paper currently lacks experiments involving moderate to large-scale settings. RegretNet already performs well with very low regret for the small scale settings shown in the paper. For smaller settings at least, one could consider Regretnet to be potentially exactly truthful. To fully demonstrate PFM-Net's advantages over regretnet, it would be beneficial to include tests with multiple agents and items (e.g., n,m≥2) 
+
+W3. Writing and Clarity  
+The writing is generally clear and accessible until Section 4. I found myself frequently switching between the appendix and the main paper to fully understand the methodology. Including the learning algorithm or a pseudo-code in the main paper would improve readability. Additionally, clearly noting in the main text when specific technical details, such as the handling of over-allocations, are explained in the appendix would be helpful as well.
+
+---
+
+### References
+[1] Yiannis Giannakopoulos and Elias Koutsoupias. Duality and optimality of auctions for uniform
+distributions. In Proceedings of the fifteenth ACM conference on Economics and computation,
+pp. 259–276, 2014.
+
+[2] Daskalakis, C., Deckelbaum, A., and Tzamos, C. (2017). Strong duality for a multiple-good monopolist. Econometrica, 85:735–767
+
+### Questions
+This approach shares notable similarities with RegretNet. Rather than having separate networks for allocation and payment, PFM-Net combines these into a single payment network with hardcoded constraints like convexity and monotonicity. The training process is also comparable: the proposed approach alternates between computing the allocation (the analogous step is finding the misreport in regretnet) and optimizing the payment function (updating the weights of RegretNet). Equation 11 is the also the same as RegretNet's objective (with the missing L2 penalty term). 
+
+For larger settings, it’s likely that PFM-Net would encounter similar issues to RegretNet with gradient-based allocation computation. Testing the proposed approach's exact violation (i.e., the term involving ReLU in Equation 11) at test time, preferably with multiple initializations of x, would be informative. For smaller settings, RegretNet incurs very low regret (and recovers optimal solutions wherever known). For larger configurations, further evaluation is needed to verify that PFM-Net’s allocation computation can overcome these scaling challenges that RegretNet faces in computing the misreports.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper explores deep learning methods applied to mechanism design, focusing on multi-buyer, multi-item scenarios where $ n $ represents the number of buyers and $ m $ the number of items. The primary focus is on menu mechanisms. For example, in a single-buyer context, a menu mechanism is defined by a set $ X \subseteq [0,1]^m $, where each allocation vector $ \vec{x} \in X $ indicates the probability that the buyer will receive each item. This mechanism is coupled with a pricing function $ p(\vec{x}) $, and the buyer selects the allocation vector that maximizes their utility.
+
+The authors attempt to establish that the class of truthful mechanisms is equivalent to the class of menu mechanisms with convex pricing functions (I found this proof somewhat difficult to follow, as I elaborate in the Weaknesses section.)  Building on this theoretical foundation, the authors design a neural architecture. While the main text does not provide details on the architecture or algorithm, the core idea is to use a convex function to represent the payment function, which is optimized during training.
+
+### Strengths
+The paper's experiments suggest that the authors may be onto something promising with their architectural design. The revenue results in Table 1 indicate strong performance relative to baselines like UM-GemNet, showcasing the potential effectiveness of their approach.
+
+### Weaknesses
+- I found the proof of the main theoretical result, Theorem 3.4, challenging to follow. This theorem claims that the class of truthful mechanisms is equivalent to the class of menu mechanisms with convex pricing functions, but several parts of the proof were confusing:
+  - On line 1018, it’s unclear what is meant by treating $ x_i^d $ and $ p_i^d $ as free variables $ x_i $ and $ p_i $. First, since these are functions, it’s confusing to call them variables, and second, because they are defined by the input mechanism $ M^d $, it’s even more confusing to refer to them as “free” variables.
+  - On line 1022, I’m unsure what is meant by saying $ \tilde{u}_i(t) $ is constant with respect to $ x_i $ and $ p_i $. By definition, $ \tilde{u}_i(t) $ depends on $ x_i^d $ and $ p_i^d $, so it doesn’t appear to be constant with respect to these terms.
+  - In Equation (5), the supremum is taken over $ t_i $, but the line before mentions $ p_i $ should be minimized. The connection between this minimization and the supremum in Equation (5) is unclear.
+  - The paragraph titled “Prove the first statement” on line 1049 is also difficult to interpret. Since $ \tilde{u}_i(t) $ is a function of $ x_i^d $ and $ p_i^d $, whether $ \tilde{u}_i(t) $ is convex should depend on the properties of $ x_i^d $ and $ p_i^d $ (e.g., whether or not they themselves convexity).
+- Section 4 would benefit from more information on the algorithm.
+- In terms of experiments, prior work (e.g., UM-GemNet) evaluates performance on a wider range of distributions beyond $ U([0,1]) $. This paper should expand its set of benchmarks to allow for a more comprehensive comparison.
+- There are also numerous grammatical issues throughout the paper. I recommend using a tool like Grammarly to identify and correct these. Lastly, it’s advisable to avoid terms like “ingenious” in the abstract when describing one’s own method.
+
+### Questions
+Could you please address my confusions regarding the proof of Theorem 3.4?
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+There has been much recent progress in using neural networks for mechanism design. Current successful approaches are either not fully expressive, are restricted to a single agent, can't guarantee strategyproofness, or require costly postprocessing. The authors of this paper present a new method that is fully expressive and works in general settings, and is strategyproof, yet is claimed not to require costly postprocessing. The key idea is rather than searching for allocation/payment rules as functions of types, they work in the dual space and compute a pricing function over each possible allocation of the mechanism. They show that truthfulness is equivalent to a convexity property of this pricing function. There are many good methods for enforcing convexity of neural networks, so the authors use these to train neural networks representing the pricing function on a couple mechanism design problems, and achieve good performance.
+
+### Strengths
+The paper tackles a key problem in mechanism design and tries to push things forward in a very creative way. The core idea is clever and original. The authors do have some successful experiments and successfully prove many important mechanism design properties for their method.
+
+### Weaknesses
+# Presentation
+
+The biggest weakness of this paper is in its presentation.
+
+There are frequent grammar and usage errors, and it might be good to fix them, but these errors don’t harm comprehension, so this is not so important.
+
+However, the organization and structure of the paper is extremely difficult to follow, far beyond the usual problems of space-limited conference papers. The definitions are unusual and non-standard, many points jump around frequently, the proofs are not well-organized, and I find myself trying to guess at what the authors are doing based on my knowledge of mechanism design, rather than learning it from the paper itself.
+
+Overall the presentation is confusing enough that I have a hard time following the paper, even though I completely understand all background work. There are many good aspects to this work but the poor presentation makes it hard to really tell what’s going on.
+
+# Experiments
+
+One of the main exciting things about automated mechanism design is its use in solving the wide-open problem of revenue-maximizing DSIC auction design. There are many auction design benchmark problems in the papers the authors cite (GemNet, RegretNet, AMA) but the authors choose to compare to none of these benchmarks, instead picking only two problems, one of which is less interesting (single buyer) and one of which is not standard.
+
+If the authors could run their method on some of the same benchmarks as in the RegretNet/GemNet papers, and hopefully produce similar plots visualizing the learned mechanisms, it would significantly increase confidence that their method works well. I think any claimed competitor to GemNet must tackle some of those problems shown in the GemNet paper (e.g. recover the Yao auction, or 2x2 uniform additive buyers, at the very least).
+
+# Full Expressiveness and Supply Constraints
+
+The authors claim their method is fully expressive, which seems to be true as far as I can tell. It is also true of GemNet (their main point of comparison).
+
+The main weakness they point out with GemNet is that it requires a costly post-processing step on a discrete grid. The purpose of this post-processing step is to achieve “menu compatibility” — during learning GemNet may choose menus such that when all bidders choose their favorite menu item, some items are oversold, and the post-processing step adjusts prices to prevent this.
+
+So the key point is the post-processing step is only required for problems where menu compatibility is an issue. But in both of the problems studied in the experiments in this paper, menu compatibility would not be an issue — GemNet also is fully expressive and requires no postprocessing!
+
+Although they don’t deal with it in experiments, in principle their method can deal with supply constraints of the sort that show up in auctions. This is discussed quite briefly in section F.1 (I think it belongs in the main body or at least should be mentioned more prominently as it is very important). However there are unresolved issues to make this work in practice. Perhaps these issues can be overcome more efficiently than the GemNet postprocessing, but the paper gives no evidence one way or the other. (Also, side note — equation 13 seems not to “type check” — a vector proj(x, X) is subtracted from a scalar)
+
+Overall as it stands, the authors seem to have pursued a clever idea with some partial success, but this is not a good paper as a paper. I think the authors should significantly rewrite it, as well as pushing experiments much further, and deal more straightforwardly with the issue of handling supply constraints.
+
+### Questions
+Although this review is somewhat harsh, I do want to give the authors encouragement for pursuing a clever and original idea, and I think in the future this could become a great paper.
+
+The authors are welcome to respond to any points in my review they want to, and to correct any errors or misunderstandings of mine. I would be happy to engage in discussion.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4

human_reviews/h24XT5DOb2.md

@@ -0,0 +1,187 @@
+## Human Reviewer 1
+
+### Summary
+The paper presents a foundational model for imputing PDP data with missing MNAR values using VQ-VAE. This system aims to encode the time series into discretized embeddings that represent latent profiles, which help capture patient behavior and improve performance in downstream tasks.
+
+The results appear comparable to existing alternatives in the literature. However, the strength of this model lies in its flexibility, as it can be applied to tasks for which it was not explicitly trained. This demonstrates that the generated embeddings capture meaningful representations of the input data, allowing their use across a broader range of objectives without requiring fine-tuning.
+
+### Strengths
+- **Model Flexibility**: One of the standout features of the proposed model is its flexibility. The embeddings learned are generalizable across different tasks, even those for which the model was not explicitly trained. This demonstrates its robustness and potential for transfer learning.
+
+- **Potential for Interpretability**: The use of latent profile embeddings offers promise for improving interpretability in clinical and behavioral data, an essential feature for real-world healthcare applications. This adds an additional layer of value to the model beyond just predictive accuracy.
+
+- **Description of the method**: The article presents a detailed explanation of the method.
+
+### Weaknesses
+- **Limited Evaluation on Downstream Tasks**: Although the paper emphasizes the model’s flexibility and its applicability to a variety of downstream tasks, the evaluation is limited to only two tasks. Expanding the analysis to include a broader range of scenarios would have provided stronger evidence of the model's generalizability.
+
+- **Lack of Exploration on Interpretability**: The paper asserts that the learned latent profiles enhance interpretability, yet no concrete evidence or case studies are provided to substantiate this claim. A more thorough investigation, possibly through qualitative analysis or correlations with clinically relevant features, would significantly strengthen the paper's contributions.
+
+- **Omission of Spatio-Temporal Architectures**: While the authors mention that exploring spatio-temporal architectures is left for future work, the choice not to use models such as GNNs, RNNs, or Transformers, which are well-suited for these types of tasks, is not sufficiently justified. This omission limits the model's applicability to more complex time series problems.
+
+- **Clarity in Results Presentation:** The presentation of results could be made clearer. The use of multi-line graphs can be confusing, and this would be better supplemented with numerical information in the main text. Additionally, incorporating a wider range of classification metrics beyond AUC, such as AUPRC, would provide a more comprehensive evaluation of the model's performance.
+
+### Questions
+- When training the model, synthetic missing values are introduced. Why are these not also used during the calculation of the loss function to optimize their imputation, potentially improving the model’s capacity to handle missing values in the time series?
+
+- What is the class imbalance in the Suicide Detection task? If there is significant imbalance, the AUC score might be misleading. In clinical settings, the AUPRC is often used as an alternative metric for imbalanced data [1]. Perhaps it could be considered for this problem as well.
+
+- The imputation errors in Table 6 seem quite high, although this may be due to the nature of the data. Could the authors provide results from comparison techniques? It would be useful to include at least one method of a common benchmark [2], and possibly show results from methods like linear interpolation and mean interpolation.
+
+- The paper mentions that these profile embeddings could enhance the interpretability of individual behaviors. Have the authors explored this? Were they able to find any correlation between the presence of specific embeddings and labels? If not, could they outline how they would approach this in future work, if it falls outside the scope of this paper?
+
+- Why were more suitable architectures for capturing spatio-temporal representations not used? While it is mentioned that this is left for future work, it’s unclear why models such as GNNs, RNNs, Transformers, or even newer blocks like Mamba were not considered [2] [3].
+
+- Figure 4 is difficult to interpret in terms of the quality of imputations and reconstructions. It would be helpful to complement this with a table, similar to Table 6.
+
+
+[1] Moor, M., Rieck, B., Horn, M., Jutzeler, C. R., & Borgwardt, K. (2021). Early prediction of sepsis in the ICU using machine learning: a systematic review. Frontiers in medicine, 8, 607952.
+
+[2] Cini, A., Marisca, I., & Alippi, C. (2021). Filling the g_ap_s: Multivariate time series imputation by graph neural networks. arXiv preprint arXiv:2108.00298.
+
+[3] Liu, M., Huang, H., Feng, H., Sun, L., Du, B., & Fu, Y. (2023, April). Pristi: A conditional diffusion framework for spatiotemporal imputation. In 2023 IEEE 39th International Conference on Data Engineering (ICDE) (pp. 1927-1939). IEEE.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper explores patient behavior monitoring and suicide detection by adapting the vector quantized variational autoencoder (VQ-VAE) to analyze data from real-world wearable devices. By reconstructing heterogeneous multi-source time-series data and accounting for patterns of missing data, the proposed model is validated to be effective in detecting change-time events within a cohort of suicidal patients.
+
+### Strengths
+S1. This research provides significant medical potential by modeling heterogeneous, multi-source time-series data from wearable devices, aiming to develop advanced approaches for identifying potential behavioral shifts that signal serious mental health risks.
+
+S2. The proposed model, which leverages VQ-VAE as the foundation model, captures the missing data patterns, and incorporates a change-point detection algorithm, is overall reasonable from a technical standpoint for this specific medical application.
+
+S3. The experimental evaluation includes a demonstration of imputation performance, an assessment of downstream tasks, results from statistical tests, and several representative case studies.
+
+### Weaknesses
+W1. My primary concern regarding this paper is its technical novelty. Although the proposal adeptly uses the VQ-VAE model as the foundational framework, modifies its design to capture missing data patterns in time-series wearable device data, and connects to change-point detection as a downstream task, it relies heavily on existing techniques. This reliance diminishes the perceived novelty of the proposed model. Moreover, the paper does not sufficiently articulate the challenges and complexities associated with integrating these techniques, which is crucial for demonstrating its innovation.
+
+W2. The evaluation of imputation performance lacks comparative baselines, which undermines the robustness of the findings. Additionally, the downstream task evaluation includes only one baseline, HetMM, which is insufficient. Incorporating a broader range of advanced and recent baselines would help demonstrate the superiority of the proposal in this research area.
+
+W3. The evaluation relies solely on a single proprietary dataset, limiting the generalizability of the findings. Expanding the evaluation to include multiple datasets would enhance the credibility and applicability of the proposed model across different contexts.
+
+W4. The performance improvements offered by the proposed model are marginal when compared with HetMM. Given that the proposal claims advantages in scalability and efficiency, it is essential to substantiate these claims with experiments on larger datasets or real-time analysis scenarios.
+
+W5. As the proposal targets a specific healthcare application—suicide detection—it would be advantageous to include medical validation of the findings or to offer insights into how the proposed model could inform real-world clinical decision-making. Such contributions would significantly elevate the practical relevance of the proposal.
+
+W6. The writing of this paper could be improved by addressing the following aspects:
+
+(i) The paper should clearly identify which model among Model A0, Model A1, and Model A2 performs best and should be recognized as the final proposed model.
+
+(ii) The inclusion of a detailed related work section is necessary. This section would help situate the paper within the existing body of knowledge, providing a comprehensive backdrop against which the current research can be assessed and appreciated.
+
+### Questions
+Please refer to points W1 through W6 for detailed concerns and recommendations.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+3
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces a foundation model for patient behavior monitoring and suicide detection based on a modified Vector Quantized Variational Autoencoder (VQ-VAE). This model effectively reconstructs heterogeneous time-series data from wearable devices and captures missing data patterns. The paper highlights its application in predicting suicide attempts using a probabilistic change-point detection (CPD) algorithm on data from a cohort of psychiatric patients. The model shows promise for scalable, efficient patient monitoring by outperforming patient-specific methods without requiring fine-tuning.
+
+### Strengths
+The approach is novel in integrating VQ-VAE with CPD for behavioral analysis and suicide prediction, particularly in using discrete latent representations for temporal health data. The paper is methodologically strong, with well-designed experiments and performance evaluations. The work has substantial potential in clinical monitoring and predictive modeling, addressing a critical healthcare application with an approach that could be extended to other domains. The paper is well-written, with clear articulation of concepts and objectives, making complex methodologies accessible. Figures and tables are effectively used to illustrate the architecture and experimental results.
+
+### Weaknesses
+The evaluation is conducted on a single private dataset. 
+The use of a proprietary dataset limits reproducibility and prevents validation or benchmarking with public datasets, reducing transparency. Expanding to public datasets could provide a more comprehensive assessment of the model's generalization and facilitate future comparisons.
+
+The study lacks a comparison with other generative models for reconstruction tasks, which could demonstrate the superiority or limitations of the proposed VQ-VAE model, comparisons with more generative models in the reconstruction phase could bolster the paper’s robustness.
+
+Although three VQ-VAE variants (A0, A1, and A2) are tested, there is limited discussion on the specific performance differences across these variants and their impact on handling missing data.
+
+Certain technical details, such as the choice of hyperparameters, could be clarified to enhance reproducibility.
+
+### Questions
+1. Could you elaborate on the performance differences among the three VQ-VAE variants (A0, A1, and A2) in terms of handling specific missing data patterns?
+
+2. Have you considered evaluating the model on additional, publicly available behavioral datasets to better assess generalizability and facilitate reproducibility? Do the authors plan to publish or share the dataset to reproduce and enable further research validation?
+
+3. Have you explored using alternative generative models, such as GANs or VAEs, for comparison on data reconstruction? How do you envision these models performing relative to VQ-VAE in terms of capturing behavioral patterns?
+
+4. Can the authors clarify the rationale for choosing specific hyperparameters for CPD (e.g., the hazard function parameter λ)?
+
+A minor typo: Line 188 - The term "D" is used without prior definition or context.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The authors introduce a model based on VQ-VAE, trained in a self-supervised way with different (informative) ways of modeling missingness. The authors look at reconstruction and imputation accuracy of the model on a behavioral data set of wearable readings. Using the trained model the authors implement a Bayesian online learning procedure to infer the MAP of the presence of a change point in the data with proposed application to suicide prevention, which is benchmarked against HetMM on their behavioral data set in terms of ROC curves.
+
+### Strengths
+- Time series modeling and change point detection are important problems in clinical applications
+
+### Weaknesses
+- A related work section for other time series models is missing, but important to place the work within the context of the broader literature
+- For imputation, the model should be compared to baselines.
+- Other baselines for change-point detection than HetMM would be beneficial, e.g. point process frameworks, time series foundation models
+- The novelty of the work (as opposed to applying a known framework to a novel interesting problem) is not entirely clear to me.
+
+### Questions
+- What is the calibration of the pseudo-probabilities as compared to HetMM?
+- How does the model deal with censoring due to loss of follow-up?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+1
+
+### Rating
+3
+
+### Confidence
+3

human_reviews/jTEKTdI3K9.md

@@ -0,0 +1,210 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposed AVHBench, a benchmark to assess cross-modal hallucinations in audio-visual LLMs. The benchmark is built on existing datasets VALOR and AudioCaps, and contains about 6k QnA pairs and 1k audio-visual captions across four cross-modal tasks, including audio-driven video hallucination, video-driven audio hallucination, audio-visual matching and audio-visual captioning. The authors design a semi-automatic pipeline for data anntation. Several open-source audio-visual LLMs are evaluated on AVHBench, and the results show that most existing audio-visual LLMs suffer from cross-modal hallucinations. To alleviate this problem, the authors further enhances Video-LLaMA through audio feature alignment and LoRA fine-tuning, proving that audio-visual hallucinations might come from insufficient training on paired audio-visual data.
+
+### Strengths
+- This paper proposed a audio-visual hallucination benchmark mainly focusing on cross-modal hallucination evaluation, which is an aspect that has received little attention.
+
+- Some tasks in the benchmark provide new perspectives on the study of multi-modal hallucinations, like *Audio-driven Video Hallucination* and *Video-driven Audio Hallucination*. The inability of the model to distinguish between information from audio or video may be a vital reason that causes multi-modal hallucination.
+
+- The paper is well-written, clear and easy to understand.
+
+### Weaknesses
+- The benchmark uses existing datasets VALOR and AudioCaps, which may introduce biases inherent to those datasets, potentially affecting the generalizability and validity of the evaluation.
+
+- In this benchmark, for human speech, the authors seem to have only considered the event of "one person is speaking," instead of taking into account the content of what is being said. The content of speech contains a wealth of information and is very likely to contribute to the hallucinations of audio-visual LLMs. However, the paper seems to overlook this scenario.
+    
+    For example, consider a scenario where a person in a video is saying to himself "Yesterday I heard a dog barking", but neither the dog nor the barking sound appears in the video or audio. Models are prone to hallucinations in this scenario. 
+
+- It is recommended to evaluate more recent audio-visual LLMs, like Video-LLaMA 2, or video-SALMONN.
+
+- Since the video is so rich in content, long text is required to descirbe the video completely. However, for the "audio-visual captioning" task, only a short caption is provided as the groundtruth. This suggests that the groundtruth caption is likely to contain only the important information in the video and omit the secondary information. However, it is still possible for the model to describe something that is present in the video but is not hallucination. That's why I'm concerned about the correctness of the "audio-visual captioning" task of AVHBench.
+
+### Questions
+- After LoRA fine-tuning, the results of the model on AVHBench become significantly better. Does this suggest that the model may just not be able to do the judgement questions, or that it hasn't seen the mismatched audio/video and therefore performs poorly on that test set, rather than the model having a large number of hallucinations?
+
+- Do the authors consider providing Gemini's results on the benchmark?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+**Summary of this paper:**
+
+This work proposes a cross-modal hallucination evaluation benchmark called AVHBench, which comprises four different tasks: audio-driven video hallucination, video-driven audio hallucination, audio-visual matching, and audio-visual captioning. Besides, the paper analyzes the presence of cross-modal hallucinations and investigating their potential causes using the proposed benchmark on six recent audio-visual LLMs.
+
+**Strengths:**
+
+1.	This paper introduces the first comprehensive benchmark specifically designed to evaluate the perception and comprehension capabilities of audio-visual LLMs.
+
+2.	Authors include a clear organization of the related literature on multimodal large language models (MLLMs) and hallucinations in MLLMs.
+
+3.	The figures in this paper are well-designed.
+
+**Weakness:**
+
+The datasets used in this paper are relatively limited in terms of scene diversity, which may hinder the benchmark's ability to evaluate the model's performance in a broader range of real-world scenarios.
+
+**Comments, Suggestions And Typos:**
+
+To provide a more comprehensive evaluation of this benchmark, it is recommended to increase the number of evaluation models and diversify the scenarios.
+
+### Strengths
+Refer to the summary
+
+### Weaknesses
+Refer to the summary
+
+### Questions
+Refer to the summary
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+5
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces AVHBench, a novel benchmark designed to evaluate cross-modal hallucinations in audio-visual large language models (LLMs). Addressing a critical gap, AVHBench tests models on their ability to handle complex interactions between audio and visual cues without generating erroneous outputs, known as cross-modal hallucinations. It includes four tasks: audio-driven video hallucination, video-driven audio hallucination, audio-visual matching, and captioning. Using a semi-automated pipeline for dataset curation, AVHBench facilitates robust assessment and enhancement of model accuracy in handling multimodal inputs.
+
+### Strengths
+1. This paper is well-written.
+2. The motivation for proposing benchmark for audio-visual hallucination is clear.
+3. The tasks proposed by the benchmark are valuable, and the semi-automatic solutions designed are also reasonable.
+
+### Weaknesses
+As a benchmark paper, more evaluation models are needed. For example, the recent video-salmonn[1], advanced Gemini[2], and unimodal models like Qwen-audio[3,4] for audio, llava-onevision[5] for vision.
+
+1. Sun, Guangzhi, et al. "video-SALMONN: Speech-enhanced audio-visual large language models." arXiv preprint arXiv:2406.15704 (2024).
+2. Team, Gemini, et al. "Gemini: a family of highly capable multimodal models." arXiv preprint arXiv:2312.11805 (2023).
+3. Chu, Yunfei, et al. "Qwen-audio: Advancing universal audio understanding via unified large-scale audio-language models." arXiv preprint arXiv:2311.07919 (2023).
+4. Chu, Yunfei, et al. "Qwen2-audio technical report." arXiv preprint arXiv:2407.10759 (2024).
+5. Li B, Zhang Y, Guo D, et al. Llava-onevision: Easy visual task transfer[J]. arXiv preprint arXiv:2408.03326, 2024.
+
+### Questions
+see weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This article proposes a comprehensive benchmark for evaluating the perceptual and understanding capabilities of audiovisual LLMs, which includes four tasks: Audio-driven Video Hallucination, Video-driven Audio Hallucination, and Audio-visual Matching. It assesses the hallucination phenomena of existing AV-LLMs, as well as the cross-modal matching and reasoning abilities of these models, and provides relevant analyses and conclusions.
+
+### Strengths
+1. Well-Written and Accessible: The article is well-written, well-motivated, and easy to follow.
+2. Comprehensive Benchmark: The proposed benchmark is comprehensive, containing several complementary dimensions and devised tasks.
+3. Valuable Takeaways: The takeaways provide valuable analyses, conclusions, and insights.
+4. The paper presents three methods to improve the trustworthiness of multimodal large language models (MLLMs).
+
+### Weaknesses
+1. How do the authors ensure the quality of the dataset? Are there any evaluation measures in place?
+
+2. For the tasks of Audio-driven Video Hallucination and Video-driven Audio Hallucination, how do the authors ensure that the visuals or audio contain objects that are either silent or not present? Most events, objects, and sounds in videos are quite singular, with the sound-producing objects being consistent and uniform in both video and audio.
+
+3. How should the presence of ambient sounds, such as wind or rain, which do not correspond to specific "objects" in the visuals, be handled?
+
+4. If there are multiple objects of the same type in the video that do not make sounds, for example, several dogs that are silent while a dog off-screen is barking, or if the audio consists largely of narration while the video shows people, would such cases still be classified as "present in both video and audio"?
+
+5. How is the situation handled when the audio is background music? Additionally, in cases where the samples themselves are inconsistent between video and audio but originate from the same video source, the model may determine them as not matching, while the ground truth indicates they do match. How is this discrepancy addressed?
+
+### Questions
+See Weaknesses.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 5
+
+### Summary
+The paper proposes AVHBench, a benchmark for evaluating cross-modal hallucination for audio-visual language models. The paper finds that the current models fall short when it comes to evaluations designed to test hallucination, with a performance close to random guesses. The eval bench comes out of a GPT-4 aided data generation pipeline with human verification. The author fine-tuned their model using the data coming out of the same pipeline without any evaluation and found that they were able to improve the hallucination issue significantly.
+
+### Strengths
+Looking into visual-audio-language model cross-modal hallucination seems to be novel. The paper is well-written and clearly motivated.
+
+### Weaknesses
+* The synthetic dataset seems to be the important component for both the evaluation set and the training set.  There seem to be several sources of error that the author did not either discuss or give some analysis on:
+ 1. For Audio-Visual disentanglement: (a) error might come from the visual tagging process (b) The prompt in Table s1 can not distinguish the sound or appearance of multiple instances of the same type of object. (e.g. given two people and the sound of a human talking, it will recognize someone is talking but have no idea whether the one who is talking is in view)
+2.  For audio-visual caption generation. Given two unaligned audio and visual captions, likewise, the language model is not guaranteed to be able to capture the correspondence between visual and audio information.
+Given the above, I think it's crucial that the author give some error analysis of the proposed pipeline (e.g. from the verification data of manual labour.) 
+
+* Related to the previous point, the proposed pipeline is not able to generate audio-visual captions/questions that require temporal reasoning. Would be good to have some discussion on this.
+
+* Also related to the above point, to show the proposed pipeline is truly useful for generating data to fine-tune the audio-visual model, it would be nice to see some more results on how the fine-tuned model performs on other audio-visual benchmarks.(In Table 4 beyond VAST Captioning dataset)
+
+### Questions
+My concern majorly lies in the error analysis and limitation of the data generation pipeline as well as the results on the fine-tuned models.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3