Add files using upload-large-folder tool

human_reviews/0kHbD6ad07.md

@@ -0,0 +1,180 @@
+## Human Reviewer 1
+
+### Summary
+This is a primarily theory paper that discusses decoder only transformers being injective. Specifically that each unique prompt maps to a unique last token embedding (from final transformer layer). Paper provides theoretical setup and justification for this argument under acceptable assumptions at initialization (Theorem 2.2) and post SGD-like gradient based training (Theorem 2.3). It is to be noted, that the theorems state how injectivity is very high probability (1 or near 1) event under stated assumption and not proven. For eg., paper discusses how collisions (two different input mapping to same last token representation) can still occur if an adversarial crafts data specifically so (end of Section 2). 
+
+Paper also introduces a prompt recovery method from last token representation. The method "Sequential Inverse Prompt via ITerative updates" or SIP-IT. This method says that at each index position given a current + previous token hidden representation, check each token in vocab to see which will produce current index hidden representation (formal algorithm describe in Algorithm 1). The paper further discusses computational cost of SIP IT.
+
+Review Note: Since the primary results are theoretical, I reduce my confidence to 3 in the reading and review.
+
+### Strengths
+- This is a unique and very science paper on LLMs, something we rarely see. Thanks for working on this. (I have what takeaways from this concern but more on that below).
+- It cites most related work I know and use proper baselines in experiments.
+
+### Weaknesses
+The paper seems poorly organized, I felt lot of important results were pushed into appendix and main paper felt very repetitive at times. I have listed my main qualms in the questions section.
+
+### Questions
+1. What is "measure-zero parameter choices"? 
+1. If you consider the prompt length progressively growing (i.e. --> \inf) with no upper bound, but the model width (i.e. embedding or hidden representation size) is fixed, then the model's representation has to be overloaded to some extent. This would challenge the injectivity directly since it won't be lossless anymore. So I would argue, in the limit, since the vector dim of hidden dim is capped, transformers are not injective and infact more and more lossy. What do the authors think?
+1. Since the language follows power law of Zipf distribution [1], the words occurs as per an exponentially decaying frequency. One example, let us have LLM predict the next token on "Complete this: The quick brown fox jumps over the lazy [dog]" where "dog" is not part of prompt. If this exact phrase is used use with two distinct prompts (i.e. prefixes of this phrase), the next token would always be "dog" for most useful LLMs. Can you please run this as experiment with varied internet data used as prefix? If the next token prediction is same, I expect the last token to be super close if not identical and see if it meets the threshold of 1e-6 used in paper. There can be other examples as well, like cases where "should" is followed by "have" (assuming "should have" token is not part of vocab) with vastly different prompts. This is more inline with Zipfian distribution argument.
+1. Why threshold of 1e-6?
+1. Can you explain what exactly is the BRUTEFORCE method in Table 2?
+1. "distinct prompts produce distinct hidden states under standard initialization and training" is good summary of the paper.
+1. I think that under high precision, lookup of hidden activation against data is already expected. This steps from how deterministic models are. Can you clarify, under your threat model, the extractor of data have ONLY the last token embedding/representation with none of the prompt (and no knowledge of it's length either) or its access to each of previous token's embedding/representation as well? Because latter is kind of trivial (to me at least) and former is very difficult and the experiments might be convincing in that case. So yeah, the exact threat model is unclear to me, and kind of not stated clearly enough in the paper.
+
+[1] https://en.wikipedia.org/wiki/Zipf%27s_law
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The authors claim that a common thought is that language models are have a many to one relationship with regard to input and latent representation (respectively). The authors argue that language models are injective and that this suggests that there is an invertible relationship such that given a latent state, the input provided to arrive at this state is recoverable. 
+
+The authors base the analysis on the belief that initialized transformers don't typically create collisions between prompts and that after a fixed number of training updates, the parameters have not been changed sufficiently to induce collisions. 
+
+The authors propose a method to obtain the input to a language model. However, the method requires knowledge of the total last layer hidden state of the model for every substring of the prompt whose first token is the first token in the prompt.
+
+### Strengths
+This paper presents an interesting analysis I have not seen of the transformer architecture and convinced me that the transformer is often real analytic (though that depends on the choice of activation functions).
+
+### Weaknesses
+Terms like measure-zero and injective need to be defined. 
+
+A collision must be defined. It's not sufficient to say that a collision didn't occur because a value differed in any way. Some bound must be placed on the distance between the output representations that constitutes a collision. 
+
+Many activation functions are not real analytic functions. 
+
+My understanding is that the zero set necessarily is measure-zero (given that the model is real analytic and we know the model is not the zero function). This doesn't seem to necessarily imply that no two inputs to the function will provide a similar output. I believe the authors intend for the reader to understand that the difference of two activations given dissimilar prompts is itself a real analytic function which then suggests that the difference between any two prompts whose value is zero is a measure-zero set. The problem with this logic is that it is possible that the difference between the last stage activations for two prompts is either the zero function or epsilon close to the zero function for some small epsilon without violating anything in the proof. 
+
+My primary contention is that showing that the zero set is measure-zero is not sufficient to make the strong claim that there are no collisions. A collision should reasonably be considered anything within some small distance of zero rather than those items which have identical values given the reality of executing floating point operations a real machine.   
+
+The SipIT method is trivial - given one knows the latent representation of every substring (which is less probable than simply knowing the input text to begin with), the algorithm can produce the substring by evaluating every possible token in this location. Further, the analysis of the algorithm seems to be either obscured or incorrect. For a prompt of length N and vocabulary of k, the algorithm will in the worst case require k^N time. This is exponential where the paper claims a linear time guarantee. Perhaps the paper meant linear with respect to vocabulary? But this is meaningless since the vocabulary is stationary and the input length is the changing factor.
+
+### Questions
+Decoder-only models are claimed to be a lossless representations of their input. This seems to be in contradiction to existing results that transformers are universal function approximators - a universal function approximation ought to be able to learn a many to one mapping if appropriate. However, function approximation is based on an epsilon bound. This seems to suggest that if instead of examining the size of the zero set, one were to examine the set which is epsilon close to the zero set, this would have no theoretic guarantee of being measure-zero. How does this not eliminate the practicality of the paper's analysis?
+
+What degree of fidelity is necessary for this result to hold? If lower precision approximations are used, online batching for practical serving, or significant context information is present, these will create random noise which ought to obscure any such precise mapping as is necessary for the analysis to hold.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+1
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper presents a strong theoretical study showing that decoder-only Transformer architectures are (almost surely) injective: different prompts map to distinct last-token hidden states. Building on this, the authors introduce SIPIT, an algorithm that recovers the exact input prompt from the final hidden state. The method is theory-driven and, across extensive experiments on a wide variety of models, the authors find no collisions, providing compelling empirical support for injectivity.
+
+### Strengths
+The main strength of the paper are as follows:
+
+1. This paper makes a significant theoretical contribution by proving that standard decoder-only Transformer language models are almost surely injective – different input prompts will (with probability one) produce distinct hidden representations. The authors use real analysis to show that collisions (two different prompts yielding the same last-token state).
+2. Building on the injectivity result, the novel algorithm (SipIt) to recover the exact input text from a model’s hidden activations with provable efficiency is introduced. As prior approaches could only approximate prompts via heavy training of inversion models, whereas SipIt achieves exact recovery by directly using the model's own representations.
+3. The paper backs up its theory with comprehensive experiments on two models (a kind of old GPT-2 and more novel Gemma). The authors performed an exhaustive search for representation collisions using 100k prompts drawn from diverse sources, amounting to billions of pairwise comparisons. They report no collisions in any model or layer tested, distinct prompts always yielded distinct last-token embeddings, with clear separation margins.
+4. The SipIt algorithm is shown to be not just theoretically sound but practically effective. On GPT-2, SipIt was able to reconstruct 20-token prompts perfectly (100% token-wise accuracy) in reasonable time, without any additional training or approximation.
+5. By establishing invertibility as a fundamental property, the work has broad implications. It provides a sound basis for interpretability: knowing that the full input is encoded in the last-layer state means any failure to probe knowledge is due to method limits, not information loss.
+
+### Weaknesses
+I would highlight the following weaknesses of this paper:
+
+1. Large vocabulary scaling. How does SipIt handle very large vocabularies (e.g., 100k+ tokens)? Does runtime grow linearly in practice, or do the gradient-based heuristics keep it manageable?
+2. Uncertain theoretical result of not-analytic estimation.
Most modern models use SwiGLU or SiLU activations. Since your proofs assume analytic activations, can you confirm that these fit the theory?
+3. In continuation to the previous point, it is not clear, what happens with the quantized models. It seems that it can be the main source of the collision.
+4. The experiments are provided for models of relatively small size, thus, we cannot asses what empirically happens with the larger number of parameters.
+
+### Questions
+My questions to the authors are as following:
+
+1. Have you estimated, how SipIt works on datasets that were seen by the models during training, and on OOD samples, that were not observed by the model. Or even some random sequences of tokens? Does inversion speed or accuracy change compared to natural text?
+2. Instruction-tuned models with identical answers:
For instruction models (or even pre-trained ones) where many prompts lead to the same answer (e.g., "yes" or "no"), do the hidden states remain well separated? Have you measured how close they get?
+3. Did you ever find prompts whose hidden states were almost identical? If so, what kind of prompts were they?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The authors make a significant theoretical and practical contribution by rigorously establishing that standard decoder-only transformer language models are almost surely injective - meaning different input prompts map to distinct last-token hidden representations under common initialization and training regimes. The authors leverage real-analyticity to prove that collisions (non-injective behavior) occur only on a measure-zero set of parameters, and they further demonstrate that gradient descent preserves this injectivity. Building on this foundation, the paper introduces SipIt, a novel algorithm that efficiently reconstructs the exact input prompt from hidden activations with provable linear-time guarantees. This work bridges theory and practice, offering new insights into model transparency, interpretability, and safety.
+
+### Strengths
+The following strong points can be highlighted:
+1. The theoretical analysis is rigorous, leveraging real-analyticity and measure theory to derive almost-sure guarantees.
+2. The injectivity results are novel and counter widespread assumptions about information loss in Transformers.
+3. SipIt is both theoretically grounded and empirically validated, demonstrating exact recovery with linear-time complexity.
+4. Extensive experiments across multiple models (e.g., GPT-2, Gemma, Llama) confirm the absence of collisions in practice.
+5. The paper clearly discusses implications for privacy, interpretability, and model auditing.
+
+The paper’s core contribution is proving injectivity and enabling exact inversion, which addresses foundational questions in deep learning and has immediate relevance to interpretability and safety. The proofs are meticulous, and the experiments are comprehensive, spanning multiple model families and scales. The introduction of SipIt provides a tangible tool for future research, while the theoretical guarantees are robust and well-supported. These qualities align with ICLR’s emphasis on impactful, rigorously evaluated work.
+
+### Weaknesses
+Whereas the paper is technically solid, there are several weak points I would like to mention:
+1. The scope is limited to decoder-only Transformers with analytic components, excluding architectures with non-analytic activations (e.g., ReLU) or encoder-decoder models.
+2. The practical utility of SipIt, while theoretically appealing, is primarily evaluated in a noiseless setting; its robustness to quantization or approximate hidden states is less explored.
+3. The discussion of related work, while adequate, could better contextualize how these results complement or challenge prior beliefs about Transformer expressivity.
+4. Some proofs in the appendix are highly technical and may be inaccessible to readers without deep mathematical backgrounds.
+5. It will be very interesting to analyze injectivity correlation with other internal transformer characteristics like anisotropy and intrinsic feature dimension using such frameworks as LLM-Microscope (https://arxiv.org/abs/2502.15007, https://github.com/AIRI-Institute/LLM-Microscope)
+
+### Questions
+1. How does SipIt perform under noisy or quantized hidden states, and can the theory be extended to account for such perturbations?
+2. Could the injectivity results generalize to encoder-decoder Transformers or models with non-analytic components (e.g., ReLU)?
+3. The paper claims gradient descent preserves injectivity - does this hold for adaptive optimizers like Adam, or only for GD?
+4. Are there practical scenarios where the linear-time complexity of SipIt becomes prohibitive, e.g., for very large vocabularies?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+4
+
+### Rating
+6
+
+### Confidence
+5

human_reviews/0obDxMfeYu.md

@@ -0,0 +1,186 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces Medix, a method for Out-of-Distribution (OOD) detection that utilizes unlabeled "in-the-wild" data. Its core contribution is a two-stage process: (1) filtering potential OOD samples from the unlabeled data by identifying points whose gradient's element-wise median deviates significantly from the In-Distribution (InD) mean gradient, and (2) training a binary OOD detector using the identified outliers and labeled InD data. The authors support their method with theoretical error bounds and extensive empirical evaluations against numerous baselines.
+
+### Strengths
+1. A solid theoretical foundation is provided, with proven bounds on both inlier and outlier misclassification rates for the median-based filtering stage. This analysis offers valuable insight into the method's robustness, particularly under the Huber contamination model.
+2. The paper includes valuable investigations into hyperparameter sensitivity, pseudo-label quality, and computational efficiency, which bolster the credibility and practical understanding of the proposed approach.
+
+### Weaknesses
+1. The most significant weakness is the insufficient discussion and comparison with the most relevant prior work. The core mechanism of using gradients from an InD model to identify OOD samples in unlabeled data is the central contribution of Du et al. ("How Does Unlabeled Data Provably Help Out-of-Distribution Detection?"). While Medix employs a median-based objective instead of a mean-based one, the paper fails to compellingly argue why this constitutes a fundamental advance rather than an incremental variation. The absence of a direct, quantitative comparison with this key work (and its subsequent developments) makes it impossible to assess the true marginal contribution of the median-centric approach.
+2. The empirical comparisons, while broad, overlook several recent and highly relevant state-of-the-art methods (e.g., [a,b]). This omission weakens the claim of superior performance, as it remains unclear how Medix fares against its most direct competitors.
+3. The iterative, greedy filtering algorithm (Algorithm 1) is computationally intensive. Although profiling results are provided, the cost may be prohibitive for very large-scale datasets or applications requiring frequent model updates, which is a common scenario in real-world deployments.
+4. The theoretical guarantees rely on assumptions such as i.i.d. samples and sub-Gaussian gradient coordinates. The practical validity of these assumptions, especially for modern deep networks and complex data distributions, is not thoroughly justified and could limit the real-world applicability of the bounds.
+
+References:  
+[a] Du et al. How does unlabeled data provably help out-of-distribution detection?. ICLR'2024.  
+[b] Geng et al. LoD: Loss-difference OOD Detection by Intentionally Label-Noisifying Unlabeled Wild Data. IJCAI'2025.
+
+### Questions
+1. The mechanism of using InD model gradients to extract OOD candidates from unlabeled data is highly similar to the approach in Du et al. Could you explicitly clarify the fundamental conceptual difference between Medix and this line of work? Please provide a quantitative comparison on common benchmarks to demonstrate the advantage of using the median over other filtering criteria proposed in prior art.
+2. How does the theoretical analysis, particularly the error bounds, differ from or improve upon the theoretical framework already established in Du et al.? Is the primary theoretical contribution the adaptation of the analysis to the median operator, or are there new foundational insights?
+3. Why were the most directly comparable methods, specifically Du et al. and its subsequent developments, not included in the empirical comparison? Can you run these comparisons and report the results?
+4. Under what conditions would the median-based filtering be expected to perform poorly? For instance, how sensitive is the method to the quality of the initial InD model or scenarios where the OOD samples do not induce a significant shift in the gradient median?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper investigates how to enhance the OOD detection performance of a type of method which utilizes unlabeled wild data to train the model to distinguish between ID and OOD data. It proposes Medix, a median-based approach that filters outliers from the unlabeled wild data and then trains an OOD detector on the identified outliers and the ID samples. The authors provide theoretical guarantees for the robustness of median-based filtering and conduct experiments on CIFAR OOD benchmarks to verify its effectiveness.
+
+### Strengths
+1. The topic of utilizing unlabeled wild data for OOD detection is interesting.
+2. This paper provides a theoretical analysis of the proposed method.
+
+### Weaknesses
+1. **This paper should provide a more detailed discussion on a very important related work, SAL [1], which has only been briefly mentioned in the Related Work Section.** Both SAL and Medix share the same overall “Separate and Learn" framework, which first identifies OOD data from unlabeled wild data and then regularizes the model with them, and also leverages a similar reference gradient technique. If the reviewer understands correctly, the difference between SAL and Medix only lies in the procedures to use gradient information to identify candidate OOD data from the wild data. SAL can also be applied in the setting of dataset-level mixing. Therefore, the authors should **provide an in-depth discussion on the advantages of Medix over SAL, reinforced with additional experiment results or theoretical analysis.** 
+
+2. **This paper omits a direct experimental comparison between SAL and Medix.** In fact, the reported results in the SAL paper indicate that SAL achieves nearly zero FPR95 on both CIFAR-10 and CIFAR-100 OOD benchmarks. But this paper claims that there are substantial opportunities for further advancement on utilizing unlabeled wild data. Therefore, the reviewer suggests that the authors **include a direct comparison with SAL and extend experiments on more challenging OOD benchmarks**, such as ImageNet-1k OOD benchmarks.
+
+3. The reviewer suggests that the authors present a detailed mathematical formulation of the element-wise median (EWM) used in the proposed algorithm, which will make it easier for readers to understand the implementation details.
+
+4. The process of filtering potential outliers from unlabeled wild data incurs non-negligible computational overhead, since it involves calculating the gradient of the loss function with respect to model parameters for every sample in the wild data. The reviewer suggests that the authors provide an experiment to quantitatively evaluate the actual time cost of filtering OOD data from unlabeled wild data with Medix and compare it with the time cost to train the OOD detector.
+
+[1]. Xuefeng Du, et al. How Does Unlabeled Data Provably Help Out-of-Distribution Detection?
+
+### Questions
+Please see the weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The work proposed to use median to address wild OOD detection. Median can used against noise and outliers for achieving robustness. The paper also provides error bounds to prove the effectiveness of proposed method.
+
+### Strengths
+A trivial method to address wild OOD with a strong theory.
+
+### Weaknesses
+I have not too many issues. I just want to ask why your bound has the Contamination term. It is possible to withdraw this term? In Du's work, there is not such constant. So it seems your bound is loss. Please address this issue, I will raise my score.
+
+### Questions
+See weakness.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper proposes ​​Medix​​, a median-based framework to identify OOD samples from unlabeled wild data and train robust OOD detectors. Theoretical analysis is provided along with the method to demonstrate the benifits of Medix.
+
+### Strengths
+Theoretical analysis are provided along with Medix. The authors derived an error bounds for understanding the proposed method under sub-Gaussian assumptions.
+
+### Weaknesses
+1. There is no comparison on challenging near-OOD scenarios and large-scale datasets (e.g., ImageNet), raising the concerns of the scalability of the proposed method.
+2. ​​Limited novelty differentiation​​. Medix’s core design choice, i.e., (median filtering for OOD detector training) is closely related to​​OpenMatch​​[1] and ​​SAL​​[, with insufficient distinction in motivation and mechanics. The reviewer generally feel the novelty of the proposed method does not meet the bar of NeurIPS. More comprehensive comparison may be necessary.
+3. ​​Inadequate evaluation​​. Strong baselines​​ are missing. Initial comparisons used outdated methods (e.g., KNN+ 2022). Some recent SOTA (e.g., CIDER[3], POEM[4]) are missing.
+
+[1] OpenMatch: Open-set Consistency Regularization for Semi-supervised Learning with Outliers
+
+[2] How Does Unlabeled Data Provably Help Out-of-Distribution Detection?
+
+[3] How to Exploit Hyperspherical Embeddings for Out-of-Distribution Detection?
+
+[4] POEM: Out-of-Distribution Detection with Posterior Sampling
+
+### Questions
+see weakness
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 5
+
+### Summary
+The paper proposed Medix, which leverages the unlabled wild data to build OOD detector. Medix first utilizes the graidents from the ERM loss and an element-wise median (EWM) to extract candidate outliers from the wild unlabeled data, adopting the greedy leave-one-out strategy. The OOD detector is then trained on the ID data and the extacted data. Theoretical bounds are provided to gaurantee low error rate for outlier selection.
+
+### Strengths
+1. Principled method for filtering outliers: The authors propose Medix which can cleanly extract outliers from wild dataset, being well theoretically grounded on misclassification rate, separating contamination.
+2. This paper provides a new perspective for OOD detection in a more realistic scenario, additional but unlabeled data is provided during training and deployment.
+3. The proposed Medix achieves strong performance across various OOD datasets.
+
+### Weaknesses
+1. Computational concern. While the gradients can be efficient for extracting outliers, it may introduce unaffordable computational cost. Time and Space complexity analysis will further justify the effect of Medix. The gradient quality may also influence the extraction, which is closely related to the amount of ID data used.
+2. Modest improvement. The baselines can already achieve great performance (e.g., WOODS on CIFAR10 vs. SVHN and LSUN-C), indicating that the improvement of Medix may be dataset-based and may shrink in some other domains(harder cases).
+3. Concern of greedy selection algorithm: while the leave-one-out strategy is intuitive and straightforward and the theoretical bounds are provided, there still be failure modes where ID data is wrongly removed.
+
+### Questions
+See weaknesses above.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/1CR1MTIgmq.md

@@ -0,0 +1,97 @@
+## Human Reviewer 1
+
+### Summary
+I think this paper should have been desk rejected.
+It is essentially an attack on another article published in TPAMI, without any context or scientific contribution. The topic (brain–computer interface) is also outside my area of expertise.
+
+### Strengths
+N/A
+
+### Weaknesses
+N/A
+
+### Questions
+N/A
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+In 2021, the following paper was published in TPAMI: Palazzo et al., "Decoding Brain Representations by Multimodal Learning of Neural Activity and Visual Features," TPAMI, 43(11): 3833–3849 (2021). Since its publication, several papers have identified substantial flaws in this work. In particular, two comments papers appeared in TPAMI (Ahmed et al., 2022; Bharadwaj et al., 2023). In 2024, Palazzo et al. published a rebuttal in TPAMI presenting their response to these criticisms. The present submission is a continuation of this discourse, aiming to highlight the issues in Palazzo et al. (2024). Each of Sections 2–8 examines individual claims from Palazzo et al. (2024) and argues why they are invalid, partially providing support through experimental results.
+
+### Strengths
+The paper provides a critical discussion of several technical details of a previous TPAMI publication, following up on the subsequent comments and rebuttal papers that also appeared in TPAMI. Thus, it may help clarify the technical issues that have been raised.
+
+It also contributes to increasing general awareness of these and similar methodological challenges in neuroimaging and other fields.
+
+### Weaknesses
+This is a very unusual submission. A substantial portion of the text largely reiterates claims from the Palazzo et al. (2024) paper, as well as from the subsequent comment and rebuttal papers. The arguments presented for each cited claim are very brief and lack sufficient detail and discussion. In its current form, the paper is not self-contained, which makes it difficult to properly assess its content.
+
+It is highly questionable whether ICLR or similar conferences would be an appropriate venue for this submission, as it does not present any novel technical contributions. Considering the follow-up history of the original Palazzo et al. (2021) paper, TPAMI appears to be the more suitable venue for continuing this discussion.
+
+### Questions
+Why not consider contributing this text to TPAMI? 
+
+Alternatively, why not consider a suitable online forum established for this ongoing discussion, potentially moderated by an authoritative figure in the field? This would allow many researchers to participate, fostering a collaborative effort aimed at arriving at the truth. In this context, such a platform might be even much more effective than publishing the work as a standalone paper.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+0
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper addresses statements in a recent publication.
+
+### Strengths
+The paper seems to address a major shortcoming in the published literature.
+
+### Weaknesses
+As far as I can see, no research contribution is made and the paper is off topic for ICLR. It is also over the page limit. The paper should be sent to TPAMI.
+
+### Questions
+n/a
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+0
+
+### Confidence
+3

human_reviews/23Ea4fVGiI.md

@@ -0,0 +1,110 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces a prompting approach based on the Task–Method–Knowledge (TMK) framework, originally from educational research. The authors argue that LLMs often fail on multi-step planning tasks because typical prompts provide only shallow textual goals. TMK decomposes each problem into structured layers: task (goal and conditions), method (procedure), and knowledge (objects and relations). The authors uses PlanBench’s Blocksworld domain in a TMK-formatted json and compare it with the plain prompts on GPT family (GPT-4, GPT-4o, o1-mini, o1, and GPT-5). Results show modest but consistent improvements.
+
+### Strengths
+• The paper introduces a simple and interpretable idea: representing planning tasks in a structured TMK format may align with how procedural knowledge is expressed in model pre-training data
+
+• The approach is prompt-based and requires no fine-tuning or external resources, making it easy to reproduce and extend.
+
+• The empirical trend (larger gains for weaker models) is intuitive and suggests the TMK structure provides helpful inductive bias.
+
+### Weaknesses
+• The experiments are narrow: only one domain and one benchmark family. Claims about general planning improvement are therefore not well supported.
+
+• There are no comparisons with so many other structured prompting methods (eg CoS, ReAct, least-to-most, chain-of-thought scaffolding and so on). It is unclear whether TMK offers advantages beyond simply using more structured json templates.
+
+• The explanation of why TMK helps remains speculative. No ablation isolates whether improvements come from the educational knowledge model framing or just from better formatting and key-word cues.
+
+### Questions
+The writing quality and technical presentation are weak, with many typos, missing citations, and incorrect or inconsistent latex usage (eg mismatched quotes and unescaped underscores). The paper would need careful proofreading and formatting cleanup.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper studies integrating the Task-Knowledge-Method (TMK) framework into prompting for LLM reasoning tasks. It focused on the blocksworld task in Planbench and experimented with OpenAI models.
+
+### Strengths
+The paper shows the potential of applying a prompting technique to improve the performance of models on the blocksworld task.
+
+### Weaknesses
+1. The reviewer finds it difficult to understand what is the core idea/contribution of TMK prompting method proposed in the paper. The paper did poorly in explaining the proposed prompting method.
+2. The paper proposes a prompting method, yet there are no examples of the prompt in the paper.
+3. Experimental evidence of the advantage of the proposed method is very limited: only OpenAI models, only one task (blocksworld), and many numbers are missing (Table 2)
+
+### Questions
+The presentation of the paper needs a lot more work, to list a few:
+1. Missing citation in line 39.
+2. Missing space in line 50.
+3. Missing citation in line 394.
+4. A formatting suggestion: most citations in the paper should use `\citep{}` instead of `\cite{}`.
+5. Figure 1 is too big, and the resolution of the image is too low.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper proposes TMK, a framework to capture specific reasoning structures in LLM reasoning tasks. It features explicit task decomposition, and benefits the LLM reasoning tasks. It proposes a PlanBench benchmark to conduct experiments.
+
+### Strengths
+The method focuses on the key problem of long reasoning LLMs, which does not have clear task decomposition during thinking.
+
+### Weaknesses
++ Many fields are missing in Table 2, especially Plain Text + One Shot. Note that it's unfair to compare the other two columns (TMK + One Shot vs Plain Text + Zero Shot). The paper does not have other main results.
++ It does not make sense to replace standard description of blockworlds into irrelevant mystery or random words. No LLM learns it during pre-training, nor people will use those words to describe tasks, nor they will use LLMs like this.
++ It's not easy to understand what TMK is doing (lacking a concrete example of the prompt/formatting).
++ The method is only experimented on blockworlds, which is a toy dataset easily solvable by non-machine learning algorithms like DFS. It's unknown whether the method can be generalized to other meaningful domains, such as mathematical, logical, legal, scientific reasonings.
+
+### Questions
++ There is only 1 item in the list (Line 054)
++ Gpt5 -> GPT-5 (Line 308)
++ Unknown citation (?) (Line 394)
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+3

human_reviews/2PJkG6aV4A.md

@@ -0,0 +1,156 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes a method to evaluate societal bias in LVLMs that have strong safety guardrails. The authors note that modern LVLMs often refuse to answer traditional bias prompts that ask them to infer a person's attributes from an image. To bypass these refusals, their proposed method uses person-irrelevant tasks while providing a user's photo as implicit demographic context. Applying this framework to 20 LVLMs, the authors find it successfully avoids refusals and reveals that all tested models exhibit bias by altering their responses based on the user's perceived demographics.
+
+### Strengths
+**1. An effective framework for evaluating guarded models**
+
+The paper correctly identifies a critical problem: existing bias benchmarks are failing because modern models with safety guardrails simply refuse to answer. The proposed method is a novel and effective solution to this issue. By changing the evaluation paradigm, it successfully bypasses these guardrails to achieve a zero refusal rate, offering a practical way forward for evaluating heavily-guarded models.
+
+**2. A comprehensive, large-scale empirical study**
+
+The paper presents a comprehensive and large-scale empirical study of societal bias across 20 recent LVLMs, including both open-source and proprietary models. The evaluation is well-structured, using three distinct tasks to probe different facets of bias. This broad study provides a valuable snapshot of the current bias landscape and yields interesting comparative results between different types of models.
+
+### Weaknesses
+**1. Limited Methodological Novelty**
+
+The paper's core methodological contribution is limited. The proposed framework is largely an adaptation of the established "persona-based evaluation" paradigm from the text-only LLM domain. The primary modification is substituting an explicit textual persona with an implicit visual one via a user's photo. This is a straightforward and incremental step, rather than a fundamental innovation in evaluation methodology.
+
+**2. Potential for Confounding Variables in User Images**
+
+The evaluation framework does not adequately account for confounding variables within the user images. While the paper assumes the model's biased outputs are a reaction to core demographic traits like gender and race, the images contain many other correlated cues. The paper does not present any analysis to disentangle these factors, making it unclear if the model is reacting to the intended demographic variable or to these other confounding visual features.
+
+**3. Over-constrained "Story Generation" Task May Induce and Exaggerate Bias**
+
+One of the paper's key findings relies on the "story generation" task, which is presented as the most open-ended evaluation. However, the prompt for this task is highly constrained, explicitly forcing the model to fill in seven sensitive attributes, including job, economic status, and education. This setup is less a measure of bias in creative, open-ended generation and more a "stereotype fill-in-the-blanks" exercise. This task design may itself induce bias and exaggerate the extent to which the model would apply stereotypes in more natural, unconstrained scenarios.
+
+### Questions
+The paper's central claim is that the model's outputs are influenced by the user's demographic traits. Could the authors provide any analysis or stronger arguments to rule out the influence of potential confounding visual variables that might be spuriously correlated with demographics in the input images?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+1
+
+### Rating
+4
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper identifies a critical flaw in existing methods for evaluating societal bias in Large Vision-Language Models: the increasing prevalence of safety guardrails. Models like GPT and Claude often refuse to answer attribute-inferring prompts, making traditional bias benchmarks unreliable. To solve this refusal problem, the authors propose a novel guardrail-agnostic evaluation method. The key idea is to decouple the task from the person in the image. Instead of using the image as a target for an attribute-inferring prompt, the method uses the image as provisional user information paired with a person-irrelevant prompt. The authors instantiate this framework across three tasks (Story Generation, Term Explanation, Exam-Style QA) and apply it to 20 recent LVLMs (both open-source and proprietary). Their method successfully achieves a 0% refusal rate. The results show that all models exhibit gender and racial bias and that proprietary models, while still biased, show lower levels of bias than open-source ones, which the authors hypothesize is due to continuous monitoring and iterative refinement rather than just one-time safety training.
+
+### Strengths
+1. Clarity and Writing: The paper is exceptionally well-written, clear, and logically structured. The problem is motivated perfectly with concrete examples (Figure 1) and data (Table 1), making the paper's contribution easy to understand and appreciate.
+2. Novel and Necessary Methodology: The "refusal problem" is a very real and growing challenge for AI safety and fairness research. As models become more locked down, prior evaluation methods are becoming obsolete. The proposed guardrail-agnostic method is an elegant, simple, and highly effective solution. It cleverly reframes the evaluation to probe implicit associations without triggering the models' explicit safety guardrails. The fact that it achieves zero refusals is a strong validation of the approach.
+3. Significant and Nuanced Findings: The paper's comprehensive evaluation of 20 LVLMs produces several nuanced and important findings for the community. The observations are not monolithic ("all models are biased") but are carefully broken down.
+4. Thorough and Insightful Discussion: The discussion in Section 5 is a major strength. The authors move beyond just reporting results to hypothesize why these results occur. The argument that continuous monitoring and iterative refinement are key factors in bias reduction —more so than just static safety alignment —is a mature and important takeaway for the field, especially for the open-source community.
+
+### Weaknesses
+The paper is of very high quality overall. Here are some questions:
+1. The validation of the LLM-as-judge showed a 97% human agreement for the term explanation task. Was a similar human-agreement study performed for the story generation attribute extraction? This seems like a more complex and potentially ambiguous extraction task (e.g., parsing "personality" or "economic situation"), and it would be valuable to know the human-LLM alignment there as well.
+2. In the "Term Explanation" and "Exam-Style QA" tasks, did the models ever comment on the irrelevance of the attached photo? For example, a response like, "I'm happy to explain linear algebra, but your photo is not relevant to the query." While not an explicit refusal, this would be an interesting behavior to note, as it sits between a full refusal and the (biased)-but-compliant behavior observed.
+
+### Questions
+See weakness
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper aims to address the problem of diagnosing social biases in LVLMs which have strong safety guardrails, The authors find that leading commercial LVLMs as well as some recent open-source LVLMs refuse to answer queries that are commonly used for diagnosing bias in these models (e.g., refusing to identify the occupation of a person). To avoid this issue, the authors propose a simple approach where they prompt LVLMs with an image and a text prompt which does not ask direct questions about the person, thereby eliminating the problem of model refusal. Despite the prompt not referencing the person in the image, significant differences in story generation, term explanation, and exam-style QA responses are observed depending upon the race and gender of the person depicted in the image. Experiments are conducted across a wide range of open-source and commercial LVLMs to quantify these differences.
+
+### Strengths
+1. The problem of studying intrinsic model biases in the face of safety guardrails is important and timely
+2. The authors propose a simple but clever solution which to the best of my knowledge is novel and original
+3. Several interesting findings are provided throughout the paper, such as discrepancies in the presence of technical jargon depending upon the race/gender depicted in the image (observation 2.2)
+4. Additional analyses are provided investigating the relationship between bias and model size, gender/race bias interdependence, and bias generalization across tasks.
+
+### Weaknesses
+1. The study only investigates race and gender bias, which seems limiting considering that image datasets exist with annotations for other types of social attributes.
+2. The proposed method itself is quite simple, which is not necessarily a weakness. However, the simplicity of the approach makes the first few pages of the manuscript feel repetitive as the same basic idea is described multiple times.
+3. The term explanation and exam-style QA tasks could be viewed as somewhat contrived in that they are not realistic prompting scenarios for LVLMs (i.e., it's unlikely for an image of a person to be provided as context along with a math of physics question). However, I understand the need to use creative prompting techniques to circumvent refusals. 
+4. This work is limited only to bias evaluation and did not explore any strategies for mitigating bias.
+
+### Questions
+1. Why did you limit your study to only race & gender attributes and a single dataset (FairFace)? Other datasets such as FACET and SocialCounterfactuals contain images with annotations for other types of social attributes which would be interesting to investigate.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes a new evaluation protocol of societal bias in *guardrailed* large vision-language models. The authors show that LVLMs with safety guardrails refuse to answer direct attribute inference prompts used in conventional bias evaluation, and propose to use indirect prompts (e.g., story generation) merely using the image as a context. The new benchmark is shown to eliminate model refusal, yet still exposes hidden demographic biases in the LVLM responses.
+
+### Strengths
+- Unlike existing work that probe direct bias of models, the paper exposes indirect biases in LVLMs even through irrelevant questions to the input image. Such bias can potentially result in disparities in the response style and quality of the model, and likely cannot be eliminated by mitigation methods that target direct bias.
+- The evaluation is reasonably large-scale, spanning 3 probe tasks and 20 open-source and proprietary models. Results are extensive with detailed breakdown of bias types and qualitative examples. Some of the findings, such as the breakdown of bias per subject and demographic group, expose the stereotypes of existing VLMs that merit further research.
+- Overall writing of the paper is clear and easy to follow.
+
+### Weaknesses
+- Comparison to existing open-ended bias evaluations: I wonder if the proposed approach differs substantially from VisBias and other open-ended benchmarks, beyond using a different set of prompts. While the new tasks are irrelevant to the image by design, the prompts do refer to the user photo explicitly ("I've attached my photo...") and I wonder if the models may be misled to believe its response (e.g., generated story) should relate to the image, which leads to the same spurious correlations that direct bias probing suffers from. I would have liked to see some quantitative evaluations or comparisons between open-ended probing methods, similar to how the authors compared refusal rate to closed-form prompts in table 1.
+- It would be more convincing if the LLM judge used an ensemble of multiple models to reduce potential bias and variance from the Qwen3 32B model. While the model is shown to agree with human 97% of the time, it is unclear if this is high enough as the bias measurements for the task (term explanation) are mostly under 5%.
+- Most of the qualitative analysis was presented for story generation and term explanation and make intuitive sense. However, it remains unclear to me how the exam QA task could also be affected by demographics depicted in the image. The paper does not provide a detailed hypothesis or analysis on this task. Or could it be that the measured TVD of 1-2% is already within the margin of error that one cannot conclude the bias is present with certainty at all?
+
+### Questions
+(in addition to questions raised in weaknesses)
+While the experiments are already extensive, there are a few open questions worth further exploring:
+- Does the bias originate from LLM pretraining, and how multimodal training amplifies it? Authors may apply the same benchmark on text-only models, with image replaced with detailed caption or user profile in the context.
+- How effective are existing mitigation methods on the indirect bias? Adding system prompts that instruct the LVLM to avoid stereotypes from the input image could be a starting point.
+- Likewise, the authors can compare the bias of the same model architecture in different training phases (e.g., PT->SFT->RL) to study whether each phase reduces or amplifies bias.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4

human_reviews/2uTxLC4LmC.md

@@ -0,0 +1,218 @@
+## Human Reviewer 1
+
+### Summary
+Reinforcement learning on chain-of-thoughts to improve reasoning capabilities have led to very strong thinking models like OpenAI-o1 and Deepseek-R1. These models generally produce a long sequence of “thinking” tokens (enclosed by <think> </think> or similar tags) followed by the final answer, and being trained to spend additional compute for thinking has dramatically improved these models capabilities in complex tasks like math and coding.
+
+However, this increase in capabilities also raises concerns about their misuse and safety issues. This work particularly studies the safety of both the chain-of-thought and the final answer. It shows that even when the final answer is safe, the chain-of-thought generated by the LLM can still be safe. The paper claims potential misuse of the unsafe chain-of-thoughts (COTs), and studies how to make COTs themselves safe. Specifically, they identify two mechanisms, namely safety triggers (where the model generates a sequence of tokens detecting potential harm of its chain-of-thought and stops) and compliance cues (where the model complies with potentially harmful requests) that affect safety of the chain-of-thoughts. The work then proposes to generate safe responses from unsafe ones by replacing the first “compliance cue” with a “safety trigger” and completing the response, use the (unsafe response, edited safe response) as a potential preference pair, and finally run DPO style training to induce safety in the chain-of-thoughts. The paper calls this method Intervened Preference Optimization (IPO), and demonstrates its performance in both safety benchmarks as well as demonstrates that it preserves performance across regular reasoning tasks.
+
+### Strengths
+1. The paper is well-written and executed. The idea of how to generate the preference pairs is well motivated by describing why simple reinforcement learning does not suffice (section 2.3), then showing a comparison of compliance cues and safety triggers after formally defining them (Section 3.1, 3.2), all of which leads to the final method of Intervened preference optimization (Section 3.3 and 3.4). I think the way the paper is structured is quite intuitive and illuminating.
+
+2. The results shown in Table 2 are comprehensive and thorough. Across a wide range of datasets and baselines, the proposed method shows strong performance. I also like the evidence shown in favor of the proposed method not harming regular reasoning performance on math datasets.
+
+3. The explanations about safety dynamics, and analyzing the safety of chain-of-thought besides the final answer is interesting and can be useful for the community.
+
+### Weaknesses
+I will list my concerns about this paper in the order of most to least severity.
+
+## Potential Flaw in the Problem Setting
+
+The biggest weakness of this paper is actually in the problem setting of attempting to make the chain-of-thoughts themselves safe, instead of just keeping the final response safe. I do not fully buy the authors’ argument, presented in 2.2, for why we should train models to keep their chain-of-thoughts safe. Particularly, prior works [1, 2] have described how making chain-of-thoughts safer can harm monitorability of modern reasoning LLMs. [1] shows that for a strong reasoning model like o3-mini, prolonged training to apply optimization pressure on the COT directly can lead to obfuscation, where the model can take harmful/unsafe actions while obfuscating their thoughts in the COT, which may make it **much harder** to actually prevent unsafe behaviors. In other words, making COTs safe does not mean that very strong reasoning models like o3-mini will become safe — these models might just learn to obfuscate their chain-of-thoughts. **A very strong counter-argument will be needed to supporting why we need to make COTs safe instead of just the final answers, given the results from frontier reasoning models like o3-mini presented in [1], to justify why COTs should be safe as well, given the risk to monitorability that this process may produce.**
+
+A simple way for current reasoning models, adopted by most frontier LLMs, is to just not reveal their chain-of-thoughts to the user. In this way, chain-of-thoughts act like scratch pads, and just training the model to backtrack/reverse safety issues in the final answer by spending additional compute during inference [4, 5] seems like a better option compared to specifically training for COT safety.
+
+## Lacking contextualization and novelty
+
+Firstly, just teaching the model to backtrack in its chain-of-thought to correct potential unsafe generations, such as Backtrack [4] might result in better performance. This paper does not cite [4] nor compare with it. Moreover, the idea of adding interventions to unsafe responses to potentially edit them to generate a preferred response, then creating a (preferred, dispreferred) pair for DPO-style training has already been explored in [4]. Without properly citing/contextualizing the current work with respect to [4] makes it significantly weaker.
+
+## Inheriting regular problems of DPO
+
+The paper describes why regular RL/GRPO does not work (base model lacks diversity) and then proposes the alternative method of IPO. However, this seems to me to be a similar reasoning of why people were using DPO style learning for math/reasoning earlier. Specifically, if base models lack sufficient coverage and capacity, then online RL might not induce the correct learning behavior [7]. However, the proper solution of that is to train a better base model, as shown in [7], and then to use online RL on top of it, instead of using offline methods like DPO.
+
+More specifically, an offline DPO loop will inherit all the problems of such an algorithm, like unintentional unalignment [8, 9] and problems of off-policy data [10, 11, 12]. These works and the concerns they raise about offline DPO-class of algorithms are not properly addressed in the paper. For example, what happens to the probability of positive sequences as IPO training goes on?
+
+## Lack of baselines
+
+This is a very minor concern, but instead of running DPO, if one just runs SFT on the constructed preferred responses, how does that perform? How about SFT + DPO (i.e., RPO [9])?
+
+### Questions
+1. Is there any reason to use GPT-4o, instead of the specific OpenAI safety moderation API (https://platform.openai.com/docs/guides/moderation) to judge safety? I believe the safety moderation API would be a better judge since it is specifically trained for this purpose, and having a strong enough judge is important for the numbers presented to be believable.
+
+2. How does this work compare to more recent methods specifically designed to instill safety into reasoning LLMs, such as TARS [5]? More specifically, I am interested in something like Figure 2 in [5], where one can clearly see the safety vs compliance tradeoff of thinking models. Does this paper’s proposed method, IPO, reduce performance on borderline/more tricky questions, which may seem unsafe at first but the model should think/spend more compute to actually find a satisfying answer? How does IPO perform on “ambiguous” prompts defined in [5]? Possibly using prompts from OR-Bench [6], that evaluates LLMs on over-refusal, can be useful here.
+
+3. A continuation to the previous question, but how does the model perform under tricky questions such as the one shown in Figure 1 of Deliberative Alignment [3], where only after thinking through/processing the question can the model understand that the prompt is asking the model to do something unsafe?
+
+# References
+
+[1] Monitoring Reasoning Models for Misbehavior and the Risks of Promoting Obfuscation, https://cdn.openai.com/pdf/34f2ada6-870f-4c26-9790-fd8def56387f/CoT_Monitoring.pdf
+
+[2] Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety, https://arxiv.org/abs/2507.11473
+
+[3] Deliberative Alignment: Reasoning Enables Safer Language Models, https://arxiv.org/abs/2412.16339
+
+[4] Backtracking Improves Generation Safety, https://arxiv.org/abs/2409.14586
+
+[5] Reasoning as an Adaptive Defense for Safety, https://arxiv.org/abs/2507.00971
+
+[6] OR-Bench: An over-refusal benchmark for large language models, https://arxiv.org/abs/2405.20947
+
+[7] Cognitive Behaviors that Enable Self-Improving Reasoners, or, Four Habits of Highly Effective STaRs, https://arxiv.org/abs/2503.01307v1
+
+[8] Unintentional Unalignment: Likelihood Displacement in Direct Preference Optimization, https://arxiv.org/abs/2410.08847
+
+[9] Iterative Reasoning Preference Optimization, https://arxiv.org/abs/2404.19733
+
+[10] Is DPO Superior to PPO for LLM Alignment? A Comprehensive Study, https://arxiv.org/abs/2404.10719
+
+[11] Preference Fine-Tuning of LLMs Should Leverage Suboptimal, On-Policy Data, https://arxiv.org/abs/2404.14367
+
+[12] Bridging Offline and Online Reinforcement Learning for LLMs, https://arxiv.org/abs/2506.21495
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+IPO proposes a DPO style training approach for safety in reasoning models by first identifying compliance cues and replacing them with safety triggers. Interestingly, IPO only trains on preference pairs of reasoning traces but show that this also leads to safe answers. The paper first conducts studies into SFT and RL based approaches for safety training, showing the limitations of both approaches on reasoning safety. Then they analyze the CSR by looking into turning points to see that replacing compliance cues with safety triggers could lead to reasoning traces that generate safe answers. Finally, they construct a DPO (RPO) preference dataset with prefixes including the reasoning trace up to the first compliance cue, chosen responses being a reasoning completion when replaced with a safety trigger, and rejected response being the original reasoning completion.
+
+### Strengths
+The paper is strongly motivated through analysis on SFT based and RL based approaches with RL experiments rewarding the safety of the reasoning directly. Analysis on CSR also provides valuable insight into the mechanism of how reasoning leads to a safe answer. Although it has been known that reasoning helps with safety, it has not been clear how or why it helps. This paper gives insight into this problem. Most interestingly, training only on the reasoning leads to safer answers, which is a sufficiently different paradigm from RLVR training in math domains. It shows that reasoning in more open-ended domains could work differently.
+
+### Weaknesses
+1. Although the paper includes sufficient baselines for SFT-based approaches, it does not include prior open-source work that uses RL-based approaches for safety training such as TARS [1]. Since part of the motivation is that RL training also does not provide sufficient signal, comparing against prior RL-based approaches for both reasoning trace analysis (Figure 2) and performance (Table 2) would help strengthen the paper.
+2. In Table 2, is the Avg. safety score across both Rsng. and Resp. or just Resp.? It would be helpful if the table includes averages for them separately to see how the effect of IPO is different for the reasoning and answer.
+3. I am not convinced about the claim in lines 189-191 “However, the absolute safety scores remain limited on adversarial datasets, suggesting the challenge of imposing process supervision”. How much effort was put into the training for tuning hyperparameters and exploring context guidelines in the reward model? The authors should explain what methods they tried to get the best performance out of GRPO with process rewards. 
+4. The experiments on safety triggers and compliance cues were conducted on DS-8B. However, DS models are known to be relatively unsafe compared to other models [2,3], exhibiting different behavior. Does the trend also hold on other base reasoning models such as Qwen3-8B which are known to be safer?
+
+
+**References**
+
+[1] Kim, Taeyoun, et al. "Reasoning as an Adaptive Defense for Safety." arXiv preprint arXiv:2507.00971 (2025).
+
+[2] Zhou, Kaiwen, et al. "The hidden risks of large reasoning models: A safety assessment of r1." arXiv preprint arXiv:2502.12659 (2025).
+
+[3] Zhang, Zhexin, et al. "How Should We Enhance the Safety of Large Reasoning Models: An Empirical Study." arXiv preprint arXiv:2505.15404 (2025).
+
+### Questions
+General comments:
+
+1. The qualitative examples in the Appendix should be referenced in the main body. I was generally confused in lines 255-257, 274, 287-290 before I saw those examples.
+2. It would be helpful to also mention in words in section 3.4 that IPO only contains reasoning traces in the preference pairs. 
+3. It would be helpful to include RL-based approaches in the related works section.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+4
+
+### Rating
+4
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper tackles the problem of reasoning-level safety alignment in Large Reasoning Models (LRMs). It introduces Intervened Preference Optimization (IPO), a novel DPO-based method that supervises reasoning by replacing unsafe compliance cues with safe triggers to form preference pairs at safety-critical steps. Extensive experiments across multiple LRMs show that IPO substantially reduces harmful reasoning and response content while preserving or even improving reasoning capabilities.
+
+### Strengths
+- Clear technical novelty: The proposed IPO method is conceptually distinct from existing SFT or RL-based alignment methods, offering an integration of intervention-based preference optimization for reasoning safety.
+
+- Strong empirical grounding: The preliminary analyses (Sections 2 and 3) are concrete and provide interesting insights about safety of LRMs — establishing a solid foundation for the proposed approach.
+
+- Effective results: IPO consistently improves reasoning-level and response-level safety across multiple benchmarks without degrading reasoning capability.
+
+### Weaknesses
+- Limited model scale evaluation: Experiments are only conducted on mid-sized LRMs (DS-7B/8B and Qwen3-8B). Evaluating IPO on smaller (e.g., DS-1.5B) and larger models (e.g., DS-14B, DS-32B) would help assess scalability and generalizability.
+
+- Lack of adaptive robustness testing: It would be more convincing to include adaptive or adversarial tests such as obfuscation, paraphrased jailbreaks, or multi-turn exploit prompts, which better reflect real-world attack surfaces.
+
+- Limited capability assessment: Although math and coding benchmarks are included, there is no evaluation of general language capabilities after alignment — e.g., factual QA, instruction-following coverage, multilingual understanding, or knowledge recall. This raises questions about whether IPO maintains general utility beyond reasoning-heavy tasks.
+
+### Questions
+Why are the training dataset sizes for IPO different across models (DS-8B, DS-7B, Qwen3-8B)?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper explores how to achieve safe reasoning in LLMs. The authors argue that most existing safety-aligned models focus only on surface-level refusals or single-turn judgments, ignoring how unsafe reasoning trajectories evolve internally.
+To address this, the paper introduces a reasoning-aware alignment framework that dynamically monitors intermediate reasoning states, detects unsafe reasoning drifts, and reinforces safety alignment during the reasoning process itself.
+
+### Strengths
+1. The paper addresses a crucial and emerging challenge — safety alignment for reasoning models. This topic is becoming increasingly important as reasoning-capable models are deployed in real-world agentic settings.
+
+2. The paper makes a clear distinction between response-level and reasoning-level safety, offering a conceptual advance that will help structure future research in this area.
+
+3. The experiments show consistent improvements on multiple safety benchmarks, while maintaining competitive reasoning performance. The trade-off between safety and capability is well-balanced, outperforming SFT-based and RL-based baselines.
+
+4. The paper is clearly written, with intuitive diagrams and examples that help readers understand complex reasoning dynamics. Figures illustrating unsafe reasoning chains and safety corrections are particularly effective.
+
+### Weaknesses
+1. The paper would benefit from a deeper theoretical discussion on why certain reasoning paths become unsafe, and why the IPO method work.
+
+2. Empirical evaluation could be strengthened by including more open-ended reasoning scenarios to demonstrate generalization beyond factual or mathematical reasoning.
+
+### Questions
+1. The paper shows that Intervention by Prompt Optimization (IPO) significantly enhances reasoning safety by replacing compliance cues with safety triggers. However, do these interventions perform differently across attack types—for instance, between direct jailbreaks and indirect persuasion attacks? Are certain categories (e.g., violence, misinformation) more effectively mitigated than others?
+
+2. While IPO successfully reduces unsafe reasoning, how does it avoid over-refusal on benign requests? The paper mentions supplementing benign datasets—could you elaborate on whether this mechanism sufficiently addresses the trade-off? Might a more fine-grained calibration strategy further improve the balance?
+
+3. Some experiments (e.g., on GPQA-Diamond) suggest that IPO not only improves safety but also enhances reasoning performance. Does this imply that safety triggers might indirectly reinforce logical consistency or reduce reasoning shortcuts? It would be valuable to further explore this potential synergy between safety alignment and reasoning robustness.
+
+4. In adversarial contexts where attackers attempt to bypass safety triggers (e.g., through indirect phrasing or prompt obfuscation), how resilient is IPO? 
+
+5. While the paper responsibly avoids releasing harmful examples, could the intervention mechanism itself be reverse-engineered (e.g., by comparing original and intervened reasoning traces)? How might future deployments ensure the transparency–security balance, especially in open-model ecosystems?
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+4

human_reviews/2za3iNkwXn.md

@@ -0,0 +1,231 @@
+## Human Reviewer 1
+
+### Summary
+This paper investigates how compression methods including quantization, distillation, and pruning, affect the reasoning abilities of large reasoning models (LRMs).
+
+### Strengths
+(1) The paper is clearly written, with a well-structured presentation that is easy to follow.
+
+(2) The motivation is articulated in a clear and convincing manner.
+
+(3) The study provides a comprehensive analysis and offers valuable new insights.
+
+### Weaknesses
+(1) The chosen reasoning models are all based on DeepSeek-R1 or its distilled variants, which may constrain the generality of the claims. It is unclear whether the findings would also hold for other reasoning models such as GPT-OSS-20B or GPT-OSS-120B.
+
+(2) The proposed layer-importance locating method is not entirely convincing. For example, Figure 2 suggests that the first-layer weights are relatively unimportant, yet Table 3 shows that quantizing the 1_up component causes the largest performance drop on AIME 2024. This apparent contradiction raises concerns about the reliability of the identified importance scores.
+
+(3) Minor concerns: the tick labels in Figure 2 are not clearly visible, and there is a typo in line 483 (“imrpving” → “improving”).
+
+### Questions
+(1) Can the findings in this paper also hold for other reasoning models such as GPT-OSS-20B/120B?
+
+(2) Why quantizing the 1_up component causes the largest performance drop on AIME 2024 but still keeps relatively good performence on other tasks.  
+
+(3) Does the identified important component also depend on the task type/task difficulties?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper investigates how compression methods affect large reasoning models (LRMs), specifically focusing on DeepSeek-R1 and its variants. The authors benchmark compressed models on four reasoning datasets and employ mechanistic interpretability techniques (difference of means and attribution patching) to identify which weights are most important for reasoning capabilities. Key findings include: (1) weight count impacts knowledge memorization more than reasoning, (2) the MLP up-projection in the final layer is critically important for distilled LRMs, and (3) current quantization methods overly compress final-layer modules and MLP gate projections.
+
+### Strengths
+**Comprehensive scope.** The paper systematically evaluates three major compression paradigms (quantization, distillation, and pruning) on LRMs, addressing a timely and important research gap.
+
+**Easy to read and well-structured.** The paper maintains a clear narrative flow with consistent notation, provides context for key design choices, and connects results to takeaways, which significantly improves readability.
+
+### Weaknesses
+**Limited model coverage.**
+The analysis centers almost exclusively on DeepSeek-R1 and its distilled variants, which makes several findings read as DeepSeek-specific behaviors rather than properties of compression that generalize across LRMs. Validating the conclusions on additional open-sourced LRM families (e.g., QwQ variants) would strengthen the generalizability claims and help disentangle model-specific effects from compression-induced phenomena.
+
+**Imbalanced treatment across compression methods.**
+The paper allocates uneven coverage across the three compression families. Distillation is represented by off-the-shelf distilled checkpoints for black-box open-source models, quantization is explored with four distinct methods, while pruning is evaluated only with SparseGPT. This asymmetry makes the comparisons look method-specific rather than family-level and may hurt perceived fairness.
+
+**Coarse knowledge vs. reasoning disentanglement.**
+The paper infers that parameter count chiefly affects knowledge based largely on lower MuSiQue EM/F1 compared to other tasks. However, this single contrast is not sufficient to attribute effects to knowledge retention versus reasoning capability. MuSiQue itself blends multi-hop reasoning with retrieval-like knowledge and is sensitive to prompt/context choices. As a result, the “parameter count affects knowledge” conclusion feels somewhat overstated. More fine-grained experiments would strengthen the claim, e.g., RAG vs. closed-book ablations across model sizes or other synthetic tasks that decouple reasoning from memorized facts.
+
+**Minor presentation issues.** There’s a typo (“imrpving”) in Section 6.
+
+### Questions
+See weaknesses above.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper systematically studies the effective of different model compression strategies on the reasoning capability of LRM, by applying reasoning-diagnosis metrics such as the steering vectors. From extensive empirical studies the authors propose certain properties and locate certain modules in LRM that are most critical to reasoning, providing insights into future designs on compression methods.
+
+After reading the paper, I am under the impression that the paper has meaningful motivation, supported by well-designed empirical studies with strong results, and the claims are well-presented. However, the paper 
+
+1. Lacks methodological novelty and diversity in the analytical framework, using only one previously-proposed metrics. Some of the main claims of the paper, such as the localization of critical modules.
+2. Could be further supported by more rigorous controlled study and evaluation metrics. 
+
+Therefore I recommend weak reject, but with room for further improvement after obtaining more insights from the authors during discussions.
+
+### Strengths
+1. The motivation of the paper is meaningful, as the authors focus on the effect of compression methods to (hard) reasoning tasks, which is both critical and not well-studied.
+2. From comprehensive experiments the authors provides valuable insights on the design of compression methods on LRMs, that is to pay attention to certain modules important to reasoning.
+
+### Weaknesses
+1. The paper makes, in my opinion, a quite strong claim on localizing the reasoning ability to certain model layers. The claim that later layer is more critical to reasoning/performances is intuitive, as they have greater impact on the final output, and could be sensitive to perturbations/compressions. However the authors do not go deeper into why certain deep layers are important to reasoning. 
+2. For a systematic study, the authors should consider using a more diverse set of diagnosing tools, rather than only using one framework, as it would put the general applicability of the main claim under question.
+3. In the paper, the locating of tokens related to certain reasoning behaviors seems not rigorous enough, as the authors only use GPT-4o to identify related tokens, which itself may have certain biases.
+4. Some missing discussions on the layer/module-wise compression methods [1, 2], and identifying weights important to reasoning [3]
+
+[1] Using Heavy-Tailed Self Regularization Theory for Improved Layer-wise Pruning of Large Language Models, https://arxiv.org/abs/2410.10912
+
+[2] Outlier Weighed Layerwise Sparsity (OWL): A Missing Secret Sauce for Pruning LLMs to High Sparsity, https://arxiv.org/abs/2310.05175
+
+[3] Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning, https://arxiv.org/abs/2506.00772
+
+### Questions
+1. Could the authors add more discussions on recent works [1] on the principal weights of model weights that are related to reasoning performances. 
+2. Since the choosing of tokens related to reasoning behaviors may be critical in the determination of reasoning-related modules, I wonder whether the authors have tried using other models (other than GPT-4o) to determine the tokens. In those cases, would the overall claims still hold? If so, it will strengthen the robustness of the findings.
+3. The authors proposes that certain layers/modules are more important in reasoning process, which relates to the discussion on the imbalanced quality among layers/modules. Do the authors believe that module-specific compression methods could potentially address the problem of over-compressing the important layers? There are recent works discussing module-wise compression [1, 2]
+4. For empirical evidence of layer importance (Takeaway 4.1 & 4.3, Fig. 2 & 3, etc.) it would be more intuitive if the authors could present the $I_{ml}^c$ metric for the base model (before distillation) to give the reader a better idea on how the importance of each module changes, rather than only presenting the $RI_{ml}^c$.
+
+[1] Using Heavy-Tailed Self Regularization Theory for Improved Layer-wise Pruning of Large Language Models, https://arxiv.org/abs/2410.10912
+
+[2] Outlier Weighed Layerwise Sparsity (OWL): A Missing Secret Sauce for Pruning LLMs to High Sparsity, https://arxiv.org/abs/2310.05175
+
+[3] Principal Weights Emerge after Rank Reduction for Reasoning-Focused Supervised Fine-Tuning, https://arxiv.org/abs/2506.00772
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper presents a two-fold analysis of the effects of compression on Large Reasoning Models (LRMs), using DeepSeek-R1 as its primary case study. First, it provides a comprehensive benchmark of three major compression families—quantization, distillation, and pruning—evaluating their impact on performance across a diverse set of reasoning datasets (AIME 2024, FOLIO, Temporal Sequences, and MuSiQue). Second, and more notably, the paper employs mechanistic interpretability techniques (adapting difference of means and attribution patching) to identify which specific model components are causally important for reasoning. The authors find that current compression methods, particularly quantization, disproportionately harm these critical components. They validate this finding by showing that selectively protecting these components (e.g., the final-layer MLP) during quantization significantly recovers lost performance.
+
+### Strengths
+Novelty of Interpretation: The paper's primary strength is its application of mechanistic interpretability to the problem of model compression. It moves beyond standard "accuracy vs. bits/size" tables to provide a causal analysis of why and where performance degrades, which is a valuable contribution.
+
+Actionable Findings: The analysis yields clear, actionable insights. The identification of the final-layer mlp.up_proj as a critical component for reasoning (in R1-distilled models) and the finding that popular quantization methods (AWQ, GPTQ) overly compress these final layers are important discoveries.
+
+Strong Validation: The experiment in Section 5.2 is compelling. Demonstrating that protecting just 2% of weights (the final-layer MLPs) from quantization can recover 6.57% in average accuracy on a 3-bit model provides strong validation for the paper's entire interpretability pipeline.
+
+### Weaknesses
+Generalizability: The analysis is heavily centered on DeepSeek-R1 and its specific distilled variants (R1-Distill-Llama, R1-Distill-Qwen). It is unclear if the central findings (e.g., the high importance of the final-layer mlp.up_proj) are a general feature of all LRMs or an artifact of the specific distillation-with-SFT process used to create the R1 models.
+
+Subjectivity in Methodology: The interpretability analysis (Section 2.2) relies on prompting GPT-4o to locate token sequences corresponding to four specific reasoning behaviors. This labeling process seems subjective and could introduce noise or bias. The robustness of the resulting steering vectors is highly dependent on the quality of this heuristic.
+
+Underdeveloped Pruning Analysis: While pruning is introduced as one of the three main compression methods, it is quickly dismissed after benchmarking shows it performs poorly (e.t., at 50% sparsity). The subsequent mechanistic analysis focuses almost exclusively on distillation and quantization, making the pruning aspect of the paper feel incomplete.
+
+### Questions
+1. Can the authors comment on whether the "final-layer importance" finding is specific to the R1-distillation process? Have you tried applying your interpretability analysis to a standard, non-distilled model (e.g., base Llama 3) that has been fine-tuned for reasoning? Would you expect to see the same components identified as critical?
+
+2. Could you elaborate on the validation process for the GPT-4o labeling of reasoning behaviors? How sensitive are the final importance scores (and the resulting conclusions) to potential inaccuracies or inconsistencies in this automated labeling process?
+
+3. The paper notes (Takeaway 3.3) that pruning/distillation (reducing parameter count) hurts knowledge memorization (MuSiQue) more severely than reasoning (AIME, FOLIO). Quantization (reducing precision) seems to have a less detrimental effect on knowledge. Could you expand on this distinction? Why do you hypothesize that parametric knowledge is so much more sensitive to parameter count than to parameter precision?
+
+I would like to improve my scores if authors can solve my questions.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 5
+
+### Summary
+This paper studies how compression (quantization, distillation, pruning) impacts reasoning in LRMs (DeepSeek-R1 and distilled Llama/Qwen) via (i) a broad benchmark on AIME-2024, FOLIO, Temporal Sequences, and MuSiQue, and (ii) mechanistic analyses that compute behavior-specific importance for every linear module using difference-of-means and attribution-patching. Key claims include: 2.51-bit dynamic quantization of R1 attains near-R1 performance; the final-layer MLP up-projection is consistently most reasoning-critical; and protecting ~2% of weights (final-layer MLPs) recovers +6.57% average accuracy for 3-bit AWQ.
+
+### Strengths
+1、 Comprehensive scope across three compression families with clear head-to-head tables and multiple distilled sizes (70B/32B/8B/7B).
+
+2、Fine-grained mechanistic lens (module-level DoM + attribution-patching) tied to four reasoning behaviors (backtracking, uncertainty estimation, example testing, adding knowledge).
+
+3、 Actionable insight: the final-layer up-projection is most critical; selectively quantizing it alone (≈0.7% of weights) causes a large drop, while protecting final-layer MLPs at 16-bit raises a 3-bit model by +6.57% on average.
+
+4、 Useful observations on collapse points (e.g., pruning ≥50% collapses; 3-bit baselines stress on harder tasks) and knowledge vs. reasoning separation (MuSiQue).
+
+### Weaknesses
+1、Robustness & Statistics. Behavior labeling robustness is under-specified (prompt/threshold/seed); key results lack rank-stability and uncertainty (variance, 95% CIs, significance).
+
+2、Metric & Visualization Choice. RI plots zero out increases, risking masked compensations; provide justified alternatives (report both ↑/↓ and net change).
+
+3、Coverage & Generalization. Pruning analysis is shallow (no mechanistic view beyond collapse); external validity is limited (mainly R1-distilled 8B/7B, few non-R1 families).
+
+### Questions
+1) Behavior-label robustness.
+Vary the behavior taxonomy/prompt/thresholds and report rank stability (Kendall’s τ / Spearman ρ) with 95% CIs; include a leave-one-behavior-out analysis.
+2) RI visualization choice.
+Justify zeroing positive RI deltas and provide an alternative view that reports both increases and decreases, plus a net-change summary.
+3) Selective protection trade-offs. Provide an accuracy vs. protected-ratio (0.5–5%) curve and the latency/memory overhead; compare protecting final-layer gate vs. up vs. both.
+4) 2.51-bit dynamic quantization details.
+Specify the skip policy (which layers stay high-precision), calibration set, and per-module bit-allocation histograms; compare to AWQ/GPTQ under identical calibration on quantization error, attribution-preservation, and end-task metrics.
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/36CzrZYEch.md

@@ -0,0 +1,162 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces IRIS, which is a RL reward based on minimizing self-confidence / maximizing self uncertainty. The authors propose this objective and use GRPO to train an autoregressive text to image model. The performance of the IRIS model is on par with other methods which use a reward signal for training.
+
+### Strengths
+- This paper is well written and does a good job of communicating the idea
+- The ablations are good.
+- I appreciate the novelty, and I appreciate the insight that encouraging mode covering behavior in autogregressive image generators trends towards human preferences.
+
+### Weaknesses
+- This approach seems limited to optimizing for human preferences as measured by these specific benchmarks. If for example, I wanted to fit to the preference of some specific group, then this method would not work.
+- The motivation for choosing this method over alternatives isn't entirely clear to me. For instance, would the approach remain effective with a different base model than Janus Pro? Would this generalize better to other human preference benchmarks?
+- I'd appreciate more understanding into the underlying mechanisms of this behavior. Is this an emergent property of the pretraining phase, or something else?
+- I am surprised that figure 2 allows one to draw this conclusion. For example, [0] and other papers indeed observe an increase in entropy in addition to an increase in reward. I notice that the GRPO for the LLM is on a MATH task, which entropy does usually decrease.
+
+- Unimportant but appendix b.1 refers to GRPO as Generative Reward Process Optimization, instead of Group Relative Policy Optimization.
+
+[0] Tonyi Deep Research, https://arxiv.org/pdf/2510.24701
+
+### Questions
+- is IRIS more robust to multiple benchmarks? why would i prefer iris over other methods?
+- what is the hypothesis for this working?
+- does this method work for other families of models?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes IRIS, a label-free RL fine-tuning scheme for text-to-image models (Janus-Pro) that replaces external reward models with an intrinsic token-level signal called Negative Self-Certainty (NSC), defined as the forward KL from a uniform distribution to the model’s next-token distribution; the authors consider both text and image tokens and ultimately use NSC for both. The reward is optimized with a GRPO objective. Experiments compare three reward schemes—external, SC, and NSC—on Janus-Pro and evaluate early training (first ~800 steps) on GenEval, T2I-CompBench, and WISE, where IRIS reports gains over the base models and competitive performance with the external-reward baseline.
+
+### Strengths
+- IRIS uses an intrinsic reward (their NSC) instead of any human labels or domain-specific verifiers, which makes the setup lightweight and potentially scalable to new domains. 
+- The GRPO objective and token-level advantage computation are written explicitly, which lowers the barrier for reproduction or adaptation.
+- The plots in the experimental results consistently report improvements early in training.
+
+### Weaknesses
+- This paper’s statements about self-confidence are mistaken. As written, IRIS is actually maximizing self-confidence, not minimizing it. The paper’s prose repeatedly says the opposite, but the mathematics and the reward used in training point the other way. Eq. (2) defines Self-Certainty (SC) at a token as $SC = -{KL}(U\ \Vert\ \pi_\theta)$, contradicting the typical definition [r1]. 
+- There is a severe mischaracterization of the proposed method IRIS. The papers asserts that forward KL to uniform is “mode-covering,” contrasting it with entropy; but with the paper’s own definition (forward KL: ${KL}(U\Vert \pi)$), increasing NSC pushes peaky (confidence-seeking) distributions, not mode-covering.
+- Metric terminology drifts mid-paper. Fig. 2’s paragraph says they “compute the self-confidence measured by ${KL}(U\Vert \pi)$,” which is NSC by their Eq. (2), while elsewhere “self-certainty” is plotted and discussed, resulting in inconsistent naming (SC vs. NSC vs. “confidence”).
+- Novelty relative to prior work (INTUITOR [r1]) is thin at the algorithmic core. IRIS’s intrinsic reward is exactly the token-level forward KL to uniform, used inside GRPO—the same scalar that INTUITOR optimizes (called “self-certainty” there) using the same GRPO algorithm. The contribution is primarily domain/application (T2I) and pipeline choices (semantic CoT), not a novel objective. This should be explicitly acknowledged and compared.
+- Gains in the experimental results may stem from the semantic CoT stage rather than the intrinsic reward per se. While the appendix shows with/without CoT comparisons, the main results and claims still bundle CoT with IRIS; stronger controls are needed (e.g., CoT-matched baselines and cross-combinations with external-reward pipelines).
+- Although IRIS trains without external verifiers, most metrics are pretrained reward/evaluator models (HPSv2, etc). Improvements against these may reflect alignment with those evaluators rather than human preference; explicit human-preference studies can only be implied with blinded A/B tests.
+
+[r1] Zhao, X., Kang, Z., Feng, A., Levine, S. and Song, D., 2025. Learning to reason without external rewards. arXiv preprint arXiv:2505.19590.
+
+### Questions
+- What exactly is “self-confidence” in Fig. 2? The text says you “compute the self-confidence measured by KL between uniform and the model’s distribution.” Is this NSC (i.e., ${KL}(U\|\pi)$)? If so, please clarify terminology and axes so readers don’t confuse SC with NSC. 
+- Why NSC on both text and image tokens? You conclude “using NSC as the intrinsic reward in both text and image tokens can achieve the best results.” Could you share ablations with mixed choices (e.g., NSC on image, SC on text), modality-specific weights, or temperature controls to support this decision?
+- Many plots focus on the first ~800 steps. Do your gains persist or change at longer horizons, and how does variance across random seeds look over full training? Please include long-run curves and seed-wise spreads. 
+- The reward NSC is differentiable against the model parameters as it’s an analytic function of the model’s softmax probabilities, hence smoothly differentiable w.r.t. the logits/parameters. It means that it can be direclty optimized using gradient-based algorithm like SGD. What is the meaning of using RL to optimize it?
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper introduces IRIS (Intrinsic Reward Image Synthesis), a reinforcement learning framework designed to improve text-to-image generation without relying on external human feedback or labeled data. The key idea is to use intrinsic rewards derived from the model’s own self-uncertainty, showing that maximizing uncertainty—rather than minimizing it—leads to more diverse, detailed, and human-preferred images. The authors demonstrate that autoregressive T2I models with low uncertainty tend to produce oversimplified and uniform results, while optimizing intrinsic uncertainty enhances visual richness and alignment with user intent. Empirical experiments confirm that IRIS effectively improves generation quality.
+
+### Strengths
+1. This paper provides a interesting observation of the relationship between model confidence and reward in the training process, and the different patterns between text generation and image generation.
+
+2. This paper proposes a reinforcement learning method that improves text-to-image generation without external reward or human annotation. After training with intrinsic reward, the model outperforms the baseline model by a large margin, and achieves comparable performance with external reward-guided training.
+
+3. This paper is motivated by a observation presented in Figure2, and fluently leads to the method design.
+
+### Weaknesses
+1. The performance using intrinsic reward is slightly lower than using external reward. It is sometimes acceptable to achieve slightly lower performance without the need of external reward, but in some cases people may prefer a higher overall performance.
+
+2. Captions of figure 5 and 6 need to switch. The captions of some other figures also seem unclear. It would be better to summarize the key observations of each figure in the caption.
+
+3. In the experiment, the authors compare the best checkpoints of T2I-R1 and IRIS, which may lead to unfair comparison. It would be better to compare them at the same training steps. Since RL training are often unstable and costly, sometimes the training step is also an important factor that we want to compare. Besides, in figure 3, it seems T2I-R1 consistently outperform IRIS by a large margin. This does not well align with the claimed results in Table1.
+
+### Questions
+NA
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper introduces IRIS (Intrinsic Reward Image Synthesis), a reinforcement learning framework for text-to-image (T2I) generation that uses negative self-certainty (NSC) as an intrinsic reward. Unlike RLHF or external reward-based approaches, IRIS optimizes autoregressive T2I models without human labels or domain-specific verifiers. The key insight is that reducing self-confidence in image token predictions enhances visual richness and diversity. Experiments on Janus-Pro across GenEval, T2I-CompBench, and WISE benchmarks show IRIS performs competitively or better than models trained with external rewards.
+
+### Strengths
+[+] Challenges standard RL intuition demonstrates that minimizing self-confidence improves image diversity.
+
+[+] Defines intrinsic rewards via forward KL divergence; mathematically grounded and well-motivated.
+
+[+] Broad evaluation across multiple benchmarks and ablations (text vs. image tokens, KL direction, semantic CoT).
+
+### Weaknesses
+[-] Reward aggregation details (text/image scaling, normalization) are unclear.
+
+[-] Not includes the comparison with preference- or diffusion-based RL.
+
+[-] There are no human studies to demonstrate that visual preference.
+
+### Questions
+1. How are text and image token rewards normalized or weighted—are they on comparable scales?
+1. Has IRIS been tested on diffusion or masked models beyond autoregressive Janus-Pro?
+1. Any user preference studies confirming that lower self-confidence yields subjectively better images?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/4fm4Rc5fUK.md

@@ -0,0 +1,112 @@
+## Human Reviewer 1
+
+### Summary
+ATF is a system that turns natural-language math problems into formal Lean 4 statements using feedback tools. It combines **syntax checks** from the Lean compiler and **semantic checks** from multiple LLM judges to iteratively refine results. Trained in three stages, ATF greatly improves both accuracy and consistency over previous models and releases a 750K-sample dataset (**Numina-ATF**) to support further research.
+
+### Strengths
+- The proposed framework, ATF, is clearly structured and the experimental results are reported in a generally comprehensible way.  
+- The topic itself is timely, and the authors make an effort to connect their work to recent trends in LLM-based reasoning and formal verification.
+
+### Weaknesses
+- I want to know what other tool calls, besides the **Syntax Check Tool**, can enhance autoformalization.
+I doubt that there are many tools capable of surpassing **Lean** in terms of checking ability, so the paper should explore **Lean’s potential as a tool** more deeply.
+
+- Lean is not good at performing numerical calculations, but I didn’t see you invoke any **calculator-related tools** in your framework.
+
+- Please provide experiments on **benchmarks that require extensive numerical computation**.
+
+### Questions
+Please refer to Weaknesses.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+2
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper proposes a model called ATF (Autoformalizer with Tool Feedback), which is designed to translate mathematical problems from natural language into formal statements. To this end, the authors design two types of tool feedback mechanisms: the first uses the Lean 4 compiler to check and correct the syntax of the generated formal statements, ensuring syntactic validity; the second adopts a multi-model voting approach to evaluate the semantic consistency of the generated results. When errors occur in the formalization results, the model can iteratively revise its outputs based on the tool feedback. To train ATF, the authors propose a three-stage training process: first, a “cold start” phase on synthetic data to teach the model how to use tools for correction; then an “expert iteration” phase to further improve the model’s capability through simulated expert feedback; and finally, a Direct Preference Optimization (DPO) phase to reduce ineffective modifications. In the experiments, ATF is evaluated on three mainstream benchmark datasets (FormalMath-Lite, ProverBench, and CombiBench), and the results show that ATF significantly outperforms the current best baseline, Goedel-V2-Formalizer-32B, in both syntactic validity and semantic consistency. The authors also release a synthetic dataset containing 750,000 formal statements (Numina-ATF) and conduct detailed human evaluations and ablation studies to verify the effectiveness of each component. Overall, this work demonstrates a new approach to significantly improve automatic mathematical formalization through tool feedback and provides new data resources.
+
+### Strengths
+- High innovation: For the first time, the paper introduces the use of Lean compiler outputs as a syntax verification tool and multi-model collective judgment as a semantic verification tool in the automatic formalization task, effectively combining the strengths of formal systems and large model reasoning.
+- Significant experimental results: The proposed method substantially surpasses the current best approach (Goedel-V2-Formalizer-32B) on multiple popular benchmarks, including FormalMath-Lite, ProverBench, and CombiBench, with particularly notable improvements in semantic consistency metrics, demonstrating the effectiveness of the approach.
+- Well-designed training process: The proposed three-stage training strategy—cold start, expert iteration, and DPO—progressively optimizes the model for different needs, enabling it to learn how to invoke tools and make reasonable corrections based on feedback, showing a thoughtful and well-structured design.
+- Resource contribution: The authors release the Numina-ATF synthetic formalization dataset with 750,000 samples, providing the community with valuable resources for training and evaluation, which holds high practical value.
+- Detailed analysis: The paper includes human evaluations and ablation studies, offering in-depth analysis of the roles of each component and the model’s scalability (such as extension effects in the inference stage), enhancing the credibility of the work. The writing is clear, and the figures are easy to read, making the contributions easy to grasp.
+
+### Weaknesses
+- Concerns about the reliability of the multi-model consistency tool: The paper relies on multiple large language models as “judges” to determine whether the generated statements are semantically consistent with the problems. However, the judgments made by LLMs may be unstable or biased, especially when it comes to subtle logical errors. Although the authors conducted human evaluations, the error rate and potential blind spots of the consistency checking tool remain unclear. It is recommended to further quantify or add verification mechanisms to ensure the accuracy of consistency feedback.
+- Limited generalization ability and scope: The current tool feedback is based on the Lean 4 compiler. If mathematical problems need to be formalized in other languages (such as Isabelle or Coq) or in different versions of Lean, the current approach may not be directly applicable. The authors mention the differences between Lean versions, but there is insufficient study on the adaptability of the method. Future work could explore the method’s transferability across different formal systems or introduce language-agnostic tool interfaces.
+- Training and inference overhead: The three-stage training and multi-round feedback mechanisms of ATF increase computational complexity. In particular, during inference, the repeated invocation of the compiler and multi-model judgments may lead to slower inference speed and higher resource consumption. The paper does not discuss efficiency in detail. In practical applications, fast response is also important, and it would be helpful for the authors to specify the model’s inference cost and latency, as well as its performance under limited computational resources.
+- Interpretability and failure analysis: Although the paper provides overall performance improvement data, it lacks an in-depth analysis of failure cases. For example, it remains unclear what types of problems ATF still struggles to formalize, or when tool feedback fails to correct the output. A detailed analysis of failure cases would help reveal the limitations of the approach and potential directions for improvement.
+
+### Questions
+- The paper mentions using multiple large language models (LLMs) as “judges” for semantic consistency verification. However, I only observed the use of **QWQ-32B** and **Qwen3-32B** in the main text. Could the authors clarify whether these are the only LLMs employed in the consistency check, or if other models were also used but not explicitly mentioned in the paper?
+- In Table 1, the results do not appear to show a clear advantage of the *Ensemble Vote* method compared to using a single LLM as the judge. I would recommend adding a new evaluation metric — **Accuracy** — to the table, which would provide a more intuitive comparison and make it easier to quantify the improvement brought by the ensemble voting method.
+- Since ATF requires multiple rounds of tool calls for iterative correction during inference, does this lead to a significant computational overhead? How does the actual inference speed compare to conventional one-shot formalization models? Moreover, have the authors considered strategies such as reducing the number of iterations or parallelizing the process to enhance scalability for large-scale mathematical libraries?
+- In Section 5.2 (“Tool Analysis”), the paper states: “As shown in Figure 5, the number of tool calls varies by dataset; CombiBench requires the highest average number of tool invocations (8.35) due to its combinatorial complexity, while FormalMath-Lite requires fewer attempts (3.19).”However, I was unable to locate the corresponding values (8.35 and 3.19) in Figure 5. Similarly, the sentence *“ProverBench is an exception where consistency checking (66.34%) outperforms syntax checking (61.65%)”* cites values that also do not appear in the figure. Could the authors verify whether these numbers are accurate or possibly correspond to an earlier version of the figure?
+- In Table 4 (ablation study), the improvement brought by adding the DPO training stage over the *Expert Iteration* stage alone appears rather marginal (around 1% increase). Could the authors elaborate on whether the DPO stage provides additional benefits beyond accuracy improvement, such as better stability, generalization, or robustness in handling ambiguous formalization cases?
+- The experiments report strong results on **FormalMath-Lite**, **ProverBench**, and **CombiBench**. However, other widely used benchmarks for formalization tasks include **MiniF2F** and **ProofNet**. Could the authors share any results or observations on these datasets, or discuss potential challenges in applying ATF to them?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper presents Autoformalizer with Tool Feedback (ATF), a framework that integrates Lean 4 compiler feedback and multi-LLM semantic evaluation to improve mathematical autoformalization. Through three-stage training—cold start, expert iteration, and DPO—ATF learns effective tool usage and revision strategies. Experiments on FormalMath-Lite, ProverBench, and CombiBench show significant gains over prior systems like Goedel-V2 and StepFun-Formalizer, particularly in semantic consistency.
+
+### Strengths
+1. The paper clearly identifies two key bottlenecks in current autoformalization models—syntactic errors and semantic drift—and systematically addresses both through tool feedback.
+2. The experiments show substantial improvements in syntactic validity and semantic consistency, further validated by human evaluation, demonstrating a strong correlation with human judgment.
+3. The system is thoughtfully designed, featuring grouped execution and expert iteration mechanisms that enable efficient syntax checking and progressive tool learning.
+4. The paper is well-written and comprehensive, with detailed appendices and clear figures that effectively illustrate the model’s iterative reasoning and tool interaction process.
+
+### Weaknesses
+1. The paper introduces a meaningful but moderately novel approach by systematizing tool feedback specifically for autoformalization
+2. Training and evaluation datasets are all derived from the Numina ecosystem; although similarity-based decontamination (cosine < 0.8) is performed, stronger guarantees against overlap would make the results more convincing. Include one external dataset (e.g., MiniF2F or PutnamBench) or a stricter decontamination threshold.
+
+### Questions
+Please refer to the Weakness section.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+2

human_reviews/4hkMvkzai5.md

@@ -0,0 +1,252 @@
+## Human Reviewer 1
+
+### Summary
+The paper introduces DeCodec, a method to decouple audio representations into orthogonal subspaces for speech and background sound. It further decomposes speech representations into semantic and paralinguistic components, similar to SpeechTokenizer. To achieve this, the authors propose a subspace orthogonal projection module and a representation swap training strategy. The model is trained on 700 hours of speech data mixed with background sounds selected from ESC-50 and DNS-Noise datasets. Experiments demonstrate competitive performance on tasks including speech reconstruction, speech enhancement, and one-shot voice conversion of noisy speech compared to baselines such as EnCodec, HiFi-Codec, DAC, and SpeechTokenizer.
+
+### Strengths
+1. Decoupling noisy speech into speech and background sound is an interesting and novel idea.
+2. The writing is clear and easy to follow.
+3. DeCodec inherently enables speech enhancement with performance comparable to specialized models.
+
+### Weaknesses
+1. The claim of being an "audio codec" is overstated as the experiments only focus on noisy speech data. Typical audio codecs should cover broader audio content, such as pure sounds or music.
+2. The study lacks experimental comparisons with relevant works, such as UniCodec and FACodec, which share similar decoupling-based approaches mentioned in the introduction and related work.
+3. The experiments are limited to 700 hours of speech data. It is unclear if the method can scale effectively to larger datasets.
+4. One-shot voice conversion results are poor (e.g., WER exceeding 50). The qualitative audio demos also exhibit low sound quality. The paper should include comparisons with stronger voice conversion models like CosyVoice2 (e.g., its S3+flow matching method).
+
+Minor Comments:
+1. Line 266: Correct "S andnN" to "S and N."
+2. Define terms like SRVQ and NRVQ upon first mention.
+
+### Questions
+Please address the issues mentioned in the Weaknesses section. Resolving these concerns may lead to a higher evaluation.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The authors propose, in this paper, the model DeCodec, a disentangled audio codec that separates audio representations into orthogonal subspaces for speech and background sound.  Built on a neural codec framework (DAC model), DeCodec introduces: a subspace orthogonal projection module that factorizes input into decoupled subspaces, and a representation-swap training strategy aligning them with speech and background features. Parallel RVQs independently quantize each subspace, while semantic guidance refines speech disentanglement. Experiments illustrate the performances of DeCodec in terms of reconstruction quality and speech enhancement.
+
+### Strengths
+The main strengths of the paper are:
+
+-	The concept of using audio codecs to build disentangled representations is very interesting (although not novel)
+
+-	The use of a subspace orthogonal decomposition module to project the primary audio embeddings in two orthogonal subspaces is interesting.
+
+-	The authors provide a comprehensible online demo, illustrating the performance of the proposed model.
+
+### Weaknesses
+The main weaknesses of the paper are:
+
+-	The authors are largely overstating the results, and in particular regarding the universality of the disentangled representation obtained or in the semantic representation control.
+
+-	The paper is not well positioned with regards to the appropriate State of the art. Some very related works in building disentangled representations using neural audio codecs are not mentioned nor discussed (for example [1] and [2] below). Similarly, the current work is weakly positioned with the SoA in Speech enhancement.
+
+[1] Omran, N. Zeghidour, Z. Borsos, F. de Chaumont Quitry, M. Slaney, and M. Tagliasacchi, “Disentangling speech from surroundings with neural embeddings,” in Proc. IEEE Int. Conf. Acoust., Speech, Signal Process. (ICASSP), 2023.
+[2] X. Bie, X. Liu and G. Richard, "Learning Source Disentanglement in Neural Audio Codec," 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Hyderabad, India, 2025,
+
+-	No indication is given to assess if the differences between approaches are statistically relevant.  
+
+-	The justification of the interest of “speech / background sound separation” from the neuroscience results (audio perception exploiting two different auditory cortex regions) is not convincing and appears to be quite disconnected.
+
+-	The clarity and writing of the paper could be largely improved (e.g. figure 1 with almost unreadable texts, many typos or inconsistencies)
+
+### Questions
+Questions:
+-	What are the references of the different speech enhancement models used and compared to (Table 2) ?. Are these models strong baselines ?
+-	Typos, some examples: separated, andnN, BRVQ (in figure 2) and NRVQ (in text), Aound, representaions, …
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes DeCodec, a neural audio codec designed as a universal disentangled representation learner. It introduces three key modules — Subspace Orthogonal Projection (SOP) for separating speech and background sound, Representation Swap Training (RST) for enforcing disentanglement, and Semantic Guidance (SG) using HuBERT features to structure speech representations. The model aims to unify multiple audio tasks (speech enhancement, voice conversion, ASR, TTS) within a single controllable representation space. Experiments on synthetic noisy datasets show strong reconstruction and robustness performance, though evidence for true semantic disentanglement remains mostly empirical rather than theoretical. Overall, DeCodec offers an interesting conceptual reframing and practical architecture, but its theoretical grounding and evaluation depth are limited.
+
+### Strengths
+* Reinterprets neural audio codecs as universal disentangled representation learners, bridging low-level compression with high-level semantic audio modeling. The bio-inspired analogy to auditory cortex processing is intellectually engaging.
+
+* Simple design: The combination of Subspace Orthogonal Projection (SOP), Representation Swap Training (RST), and Semantic Guidance (SG) forms an elegant, modular framework for factorized representation learning. The approach is lightweight, differentiable, and compatible with existing codecs.
+
+* Unified controllable representations: Demonstrates a single codec that can handle diverse tasks (reconstruction, enhancement, voice conversion, ASR, TTS) through selective subspace manipulation, showing strong versatility and practical controllability.
+
+* Achieves competitive or superior SDR, WER, and DNSMOS scores compared to baselines, particularly in noisy and low-resource conditions, indicating improved robustness and general usability.
+
+* Clear structure: The paper is well-organized, easy to reproduce, and provides intuitive visualizations linking architectural components to perceptual interpretations.
+
+### Weaknesses
+* Conceptual Validity
+
+The assumption that orthogonality between subspaces guarantees semantic disentanglement is not theoretically grounded.
+Orthogonality only enforces statistical decorrelation, not physical or perceptual independence. The learned projections may encode arbitrary frequency or energy partitions rather than truly separating speech and background sound. The analogy with the auditory cortex (A2) is purely metaphorical. Biological “orthogonality” arises from nonlinear, attention-driven and feedback-modulated mechanisms, whereas the proposed SOP relies solely on linear projections and an L2 orthogonality loss.
+
+* Methodological Concerns
+
+Representation Swap Training (RST) lacks theoretical justification. The swap reconstruction objective can be satisfied without genuine factor separation, as the decoder may simply compensate through overparameterization. The observed “disentanglement” might thus emerge from data correlation or reconstruction bias rather than causal structure. The Semantic Guidance (SG) module only supervises the speech branch using HuBERT features but is not jointly optimized with the SOP/RST mechanisms. Therefore, the claimed “hierarchical disentanglement” (semantic vs. paralinguistic) is loosely coupled and not formally guaranteed.
+
+* Evaluation and Empirical Limitations
+
+Experiments are largely performed on synthetic mixtures (LibriTTS + DNS-Noise), without tests on real-world recordings, unseen noise conditions, or cross-domain generalization. Baseline selection is incomplete — comparisons omit recent disentangling codecs such as RepCodec, FunCodec, or SemanticCodec, which pursue similar objectives. The ablation analysis only reports SDR and WER metrics; no interpretability or information-theoretic metrics (e.g., mutual information, alignment scores) are provided to confirm actual semantic separation. Several claims (e.g., controllable feature recombination, semantic robustness) are demonstrated qualitatively through spectrograms, without perceptual or statistical validation. Subjective listening tests (MOS/CMOS/SMOS) are missing; evaluation relies solely on objective or proxy metrics such as DNSMOS and WER, which may not reflect perceptual quality.
+
+* Conceptual Scope and Framing
+
+The “universal front-end” claim is overstated. The multi-task functionality is achieved through post-hoc recombination of representations rather than a truly unified optimization framework. The connection between DeCodec’s representation learning and human auditory cognition is speculative, offering conceptual inspiration but lacking neuroscientific rigor or measurable alignment.
+
+### Questions
+1. The paper draws inspiration from the auditory cortex (A2) but provides little biological or neuroscientific evidence. Could you clarify whether this analogy is purely conceptual, or if there are empirical findings (e.g., cortical response studies, neural decoding evidence) supporting the proposed “orthogonal subspace” interpretation?
+
+2. Is there any neurophysiological or psychophysical experiment planned (or referenced) to validate the hypothesis that speech and background sound representations in DeCodec correspond to separable cortical processing pathways?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper introduces DeCodec, which frames the audio codec as a disentangled representation learner. It introduces two mechanisms: Subspace Orthogonal Projection and Representation Swap procedure, to ensure that codecs subspaces correlate to speech and background sound. Experiments show that the codec achieves promising results on downstream tasks such as audio reconstruction, speech enhancement and one-shot voice conversion.
+
+### Strengths
+Originality: The paper proposed a codec representation via disentanglement to speech and background noise. This disentanglement is achieved via two novel mechanisms: one is enforcing the speech subspace and background subspace as orthogonal, and second via representation swapping which enforces:
+(a) Speech Invariance to Background: Speech latent $Zs$ must reconstruct correctly regardless of the noise it is paired with. Hence $Zs$ should only encode speech.
+(b) Background invariance to speech:  $Zn$ must reconstruct correctly regardless of speech it is paired with. Hence  $Zn$ should only encode noise.
+
+Quality: The authors conduct thorough experiments for comparing DeCodec against multiple state of the art codec models, such as DAC, SpeechTokenizer for audio reconstruction, and StoRM, SELM for speech enhancement. The authors provide ablations for orthogonal subspace constraint (SOP), representation swapping and semantic guidance. The supplemental material includes visual analysis (cosine similar and singular value comparison), and performance of codec on downstream ASR and TTS tasks to further support the codec's results.
+
+Clarity: The paper is structured well, providing theoretical analysis for factorized audio coding in Section 3, and then providing thorough experimental studies via objective results and visualizations.
+
+Significance: The paper makes a good attempt towards a universal disentangled codec separating the semantic, acoustic and background sounds. It is very relevant to the community working for neural codecs for audio tasks.
+
+### Weaknesses
+The paper aims towards having audio codec as "universal dis-entagled representation learner", though both the theoretical discussion and experimental results point towards the case in which background noise is added to speech. The results on other audio types like singing/music/sound with more language coverage on evaluation other than English speech should be included to call it universal. Singing and background instruments are co-related and this can break the assumption of orthogonal spaces and representation swapping.
+
+The authors used a well-established way of semantic vs acoustic information separation in neural codecs, which is having Hubert loss on semantic tokens, which limits the novelty to only speech vs background sound separation.
+
+For Table 1, DeCodec bitrate is 8kbps, which is higher than all the baselines. For the experimental results, Table 1 and Table 2, the baseline checkpoints are not re-trained with the new training dataset, which makes a comparison a little unfair. 
+
+WER is quite high for Table 3 for Voice Conversion at 50+%, for all baselines, the authors provide an explanation that it's due to different speech segment voicing times, which is not convincing.
+
+### Questions
+Writing and Notation:
+Please increase the font size for the text in the figures. Figure 1 text is really hard to read, especially the ones colored white.
+
+Please consider reducing the amount of new abbreviations like SOP, RST, SG, SRVQ, NRVQ, BRVQ, SDR-O, SDR-B, SDR-S. We could re-name them as Speech-RVQ, Noise-RVQ etc. Background sound vs noise are somewhat used interchangeably in the notation, please keep it consistent to one. The reader should understand tables and figures without referring to the text.
+
+Please have complete headings, instead of ones like "SE", "One-shot VC" in Section 4.
+
+Include the final training loss for the model at end of section 3.6.
+
+Please fix typos such as "representaions", "infered", "aound", "quantinized" etc.
+
+Please use subscript in notation for $Z_s$, $Z_n$, $Z_r$, $Z_c$.
+
+Method and Evaluation:
+Please Provide audio samples to convince readers of disentanglement and reconstruction quality. It will be interesting to see if real world noisy data, rather than synthetically created (speech + noise) can be separated well using DeCodec's approach.
+
+It will be interesting to compare the convolution + transformer based architecture choice as seen in Mimi (https://arxiv.org/pdf/2410.00037)t and TAAE (https://arxiv.org/pdf/2411.19842v1) vs a pure convolution architecture based on DAC.
+
+Please explain or include comparison with another common technique for disentanglement ( gradient reversal + supervision) as mentioned in FACodec (https://arxiv.org/pdf/2403.03100), and why SOP + RST is better. The authors claim there is significant information leakage in FACCodec. It would be good to empirically validate that vs DeCodec on real world data.
+
+Please clarify how the Huber semantic guidance is applied.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+This paper proposes a neural audio codec, DeCodec, designed to disentangle speech and background sounds into separate, orthogonal representations. To achieve this, the authors introduce a subspace orthogonal projection (SOP) module and a representation swap training (RST) strategy.
+
+### Strengths
+The goal of disentangling audio components within a codec is interesting and, if successful, could benefit downstream tasks like speech enhancement, noise-robust voice clone.
+
+### Weaknesses
+1. Fundamental flaw in premise: The core assumption that speech and background must be orthogonal is ill-posed for a reconstruction task. It does not convince me that there is no overlapping in the acoustic phonomenons of speech and background sound. Enforing the codec to encode them into orthogonoal representations could incur significant distortion in the decoupled outputs, as the decoder is deprived of shared acoustic information.
+2. Poor reconstruction quality on individual representation: The audio samples on the demo page substantiate the theoretical concern. Both the extracted speech and background sound in speech enhancemnet scenario are noticeably distorted, indicating that the decoder still relies on the combination of the separate representations to generate high-fidelity audio.
+3. Inefficient representation compression: The decouple of the audio signal in the proposed neural audio codec does not help the compression of encoded representation, but doubles the bitrate to 8kbps instead! Considering that the separate representation is not properly decoupled, the added overhead is not dispensible when applied to various downstream tasks, especially for generation.
+4. The semantic guidance mechanism for the speech encoder has been established in prior work and does not constitute a novel contribution.
+5. There major presentation issues in the paper:
+* In all figures and tables, the VQ module of DeCodec contains a SRVQ and a BRVQ. While in the main text, the BRVQ is never mentioned, but replaced with a NRVQ without any explanation.
+* The font size of Figure 1 is too tiny to read easily. 
+* The Figure 2 is missing a lot of captions and legends to enable a basic understanding of the proposed workflow.
+
+### Questions
+1. It is claimed that the paper proposes a novel neural codec that learns to decouple audio representations into “orthogonal subspaces” dedicated to speech and background sound. Although the authors give some justifications about "orthogonal" on Section 3.4 when introducing SOP module, the experimental results can not give full support that the speech and background sound features are really "orthogonally" disentangled.
+2. What is the theoretical or empirical justification for the requirement of orthogonal representations? Given the shared acoustic properties of speech and noise, how can high-fidelity reconstruction be achieved from a single, artificially constrained representation?
+3. How does this method aid compression, given that it doubles the bitrate? What is the computational overhead (e.g., latency, memory) when integrating DeCodec with a downstream LLM for a task like TTS?
+4. In downstream TTS evalution results, why is the comparison with DAC missing (which is compared in ASR evalution) ? There is only the weak baseline SpeechTokenizer which shows inferior reconstruction fidelity to DAC.
+5. The main text refers to an "NRVQ," while the figures label a "BRVQ." Are these the same component? If not, please clarify the difference.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4

human_reviews/5DpzzTPnJZ.md

@@ -0,0 +1,170 @@
+## Human Reviewer 1
+
+### Summary
+Studies plasticity loss in deep RL under non-stationarity. The theory isolates two mechanisms: (i) NTK rank collapse across sequential warm-starts and (ii) a $\Theta (1/k)$ decay of the initial gradient each round. Motivated by (ii), the paper proposes Sample Weight Decay (SWD)—a lightweight recency-weighted replay scheme—to restore gradient magnitude. Experiments on TD3 (MuJoCo) and SAC with SimBa (DMC Humanoids/Dog) show consistent gains with reliable aggregate metrics (IQM/median/mean, bootstrap CIs) and a reverse ablation (SWA) that up-weights old samples and underperforms, alongside GraMa analyses.
+
+### Strengths
+- Identifies a crisp cause of plasticity loss and links it to a tractable remedy (recency weighting)
+- Very simple algorithm (SWD) with negligible overhead, orthogonal to architectural methods (ReDo, Plasticity injection, etc.)
+- Consistent empirical improvements across TD3 and SimBa-SAC; reverse ablation (SWA) plus GraMa trends support the mechanism.
+- Uses reliable RL reporting (IQM/median/mean + stratified bootstrap CIs).
+
+### Weaknesses
+- Theory scope. Main results are derived for FQI-style/population losses; transfer to fully practical bootstrapped targets with representation drift is not fully established.
+- Breadth. Evaluation is confined to continuous control; adding a pixel-based or sparse-reward task (e.g., DMControl pixels, AntMaze) would test generality.
+- Over-edited text (LLM side-effects). While LLM assistance can improve flow/grammar, several sentences become awkward or semantically off and harm readability—for example, the abstract’s “How plasticity loss arises, dissipates and can be dissolved.” A careful human pass is needed to fix misuses and improve readability.
+- Related work is too narrowly framed (over-emphasis on resets). The section concentrates on reset-style approaches while under-representing other relevant families—particularly **churn-reduction methods** and **auxiliary-loss–based representation stabilisation**. Please discuss these lines of work and clarify how SWD differs or complements them (see, e.g., churn-reduction: [https://arxiv.org/abs/2506.00592](https://arxiv.org/abs/2506.00592); auxiliary losses: [https://arxiv.org/abs/2405.00662](https://arxiv.org/abs/2405.00662)).
+
+### Questions
+- How sensitive are results to **linear vs. exponential** decay and to $((w_{\min}, T))$? Any failure modes with small buffers or rapid behaviour-policy shifts? 
+- Do SWD’s benefits persist with **pixel observations** or **sparse rewards**? (A small add-on task would suffice.)
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper studies plasticity loss in deep RL through an optimization lens, identifying two mechanisms: (i) NTK rank degeneration and (ii) gradient attenuation that scales as Θ(1/k) under non-stationary data/targets with experience replay. Building on this analysis, the authors propose Sample Weight Decay (SWD)—a simple, plug-and-play, age-based sampling scheme for replay buffers that linearly down-weights older samples to counteract gradient decay and restore gradient magnitude. Experiments on MuJoCo (TD3) and DMC (SimBa-SAC) show consistent gains—including strong results on Humanoid—plus robustness across UTD ratios; ablations (including a reverse “SWA” variant) and GraMa measurements support the mechanism.
+
+### Strengths
+1. Clear theoretical framing with actionable takeaways. The paper formalizes how distribution/target non-stationarity yields NTK rank issues and a Θ(1/k) gradient-magnitude decay, then links performance to Bellman-residual control via a suboptimality bound—cleanly motivating data-weighting interventions.
+
+2. Simple, general, and orthogonal method. SWD is an easy drop-in change to replay sampling, compatible with TD3/SAC (and, in principle, other replay-based methods) and positioned as orthogonal to architectural “plasticity-injection” tricks. The paper claims minimal overhead and plug-and-play practicality.
+
+3. Compelling empirical evidence. Consistent improvements across MuJoCo and DMC (including Humanoid), robustness to varying UTD, and reverse validation via SWA plus GraMa analysis strengthen the causal story beyond raw scores.
+
+### Weaknesses
+1. Missing related work / plasticity literature coverage.
+
+The related work should more deeply connect to recent plasticity and replay-weighting literature. Please discuss and contrast with, e.g.:
+
+- Overestimation, Overfitting, and Plasticity in Actor-Critic: the Bitter Lesson of Reinforcement Learning (ICML’24).
+- Disentangling the causes of plasticity loss in neural networks (CoLLA’24).
+- Hyperspherical Normalization for Scalable Deep RL (ICML’25).
+- Mitigating Plasticity Loss in Continual RL by Reducing Churn (ICML’25).
+- A Forget-and-Grow Strategy for Deep RL Scaling in Continuous Control (ICML’25).
+
+2. Overlapping idea; contribution clarity vs ER-decay.
+
+Conceptually, SWD (linear, age-based down-weighting) looks very close to prior ER-decay heuristics, with differences seemingly in coefficients/schedules. The paper does offer more formalism, but the delta in contribution should be crystal clear: which parts are novel theory, which are new algorithmic prescriptions beyond a tuned decay, and what guarantees (if any) distinguish SWD from ER-decay? I’m open to a high score even if the mechanism is similar—provided the theoretical backup and empirical analysis are meticulous and make the case for why this instantiation is principled/non-equivalent.
+
+3. Implementation and efficiency details are thin.
+
+A per-sample weight update naively done every step can be costly. Please:
+- Describe the efficient implementation (lazy updates? piecewise-linear buckets? periodic renormalization?).
+- Report end-to-end wall-clock and GPU-hour costs on standard hardware for SimBa vs SimBa+SWD;
+
+### Questions
+1. Beyond sample-efficient regimes. 
+
+Authors primarily test in sample-efficient settings. What happens with very long training (e.g., 10M env steps with a 1M buffer) where fresh samples repeatedly overwrite older ones? Do you observe more catastrophic forgetting under high churn, and does SWD still mitigate it or saturate? (This is important because small buffers + long horizons can exacerbate plasticity.)
+
+
+
+I’m open to increasing the score if the authors adequately address the weaknesses—particularly by deepening related work discussion, clarifying the novelty relative to ER-decay, and detailing the efficiency and long-horizon robustness experiments.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper investigates plasticity loss in deep reinforcement learning from a theoretical optimization perspective. The authors identify two mechanisms causing plasticity loss: (1) rank collapse of the Neural Tangent Kernel (NTK) Gram matrix, and (2) Θ(1/k) decay of gradient magnitude during training. Based on this analysis, they propose Sample Weight Decay (SWD), a lightweight sampling strategy that assigns higher probabilities to more recent samples in the replay buffer to counteract gradient attenuation. Experiments on MuJoCo and DeepMind Control Suite tasks with TD3 and SAC algorithms demonstrate consistent performance improvements.
+
+### Strengths
+- SWD is remarkably simple to implement with minimal computational overhead, making it easily applicable across different RL algorithms and architectures as a plug-and-play solution, which is practically valuable compared to more invasive methods.
+- The experimental validation is comprehensive, including evaluation on two benchmark suites, comparison with PER, reverse validation through SWA ablation, plasticity measurement using GraMa metrics, and robustness analysis across different UTD ratios, all showing consistent improvements.
+
+### Weaknesses
+1. The linear decay design of SWD (w_i = max(w_min, 1 - age_i/T)) appears somewhat arbitrary without clear theoretical justification for why this particular weighting scheme optimally compensates for the 1/k gradient attenuation. Sensitivity analysis on decay schedules and hyperparameters is insufficient.
+
+2. Recency-based sampling is not novel conceptually, and the paper lacks direct comparisons with recent plasticity-preserving methods (ReDo, ReGraMa, Plasticity Injection). The claimed "SOTA performance" is primarily against uniform sampling and PER, making it difficult to assess the true contribution relative to the current state of the art.
+
+### Questions
+1. Can you provide rigorous analysis showing how the gradient dynamics derived for FQI extend to deep RL algorithms with experience replay and bootstrapping? How does the 1/k decay manifest in TD3/SAC specifically?
+2. How sensitive is SWD to hyperparameters T and w_min? Tables 5-6 show different T values—how were these chosen? Have you systematically compared different decay schedules (exponential, polynomial, etc.)?
+3. Since you claim SWD is orthogonal to existing methods, have you tested combinations with network reset or neuron recycling? Can you provide direct experimental comparisons with ReDo, ReGraMa, and Plasticity Injection to substantiate the SOTA claim?
+
+I will consider increasing the score if the author responds to these questions.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper studies plasticity loss in deep RL and attributes it to two mechanisms induced by non-stationarity: (i) NTK Gram rank collapse, and (ii) attenuation of gradient magnitudes that decays as $\Theta(1/k)$ over training iterations $k$. Motivated by the second mechanism, the authors propose Sample Weight Decay (SWD): linearly down-weight older transitions in the replay buffer so that recent data counteracts the $\Theta(1/k)$ decay and restores effective gradient scale. SWD is plug-and-play for replay-based algorithms.  
+Empirically, SWD is added to TD3 and SimBa-SAC on MuJoCo and DMC tasks, with strong gains on DMC Humanoid. A “reverse” ablation (SWA, which up-weights old data) predictably reduces gradient norms and hurts returns, supporting the causal story.
+
+### Strengths
+1. Tight theory to method link: The $\Theta(1/k)$ gradient attenuation analysis (Theorem 3) cleanly motivates SWD’s linearly decaying replay weights; the SWA “reverse” ablation strengthens causal plausibility.  
+2. Simple and orthogonal: SWD lives at the sampling layer, is easy to add to any replay-based RL algorithm, and should compose with model-level plasticity fixes.  
+3. Compelling results on hard control: Consistent improvements, with standout gains on DMC Humanoid, suggest the effect is meaningful, not a small-n artifact.
+
+### Weaknesses
+1. Unprobed NTK mechanism: The paper posits NTK rank collapse but provides no spectrum or conditioning measurements; half of the causal story remains speculative.  
+2. Missing SOTA positioning: No head-to-head (or composition) with recent plasticity remedies such as ReGraMa, Plasticity Injection, or ReDo, leaving SWD’s incremental or additive value unclear.  
+3. Practical knobs under-explored: No sensitivity for $T$ (linear decay steps) and $w_{\min}$; interactions with PER are not clarified (do they stack or conflict?); runtime or throughput impact is unreported.
+
+### Questions
+1. NTK evidence: Can you track NTK Gram eigenvalues (or condition number) on fixed probe batches across training, ideally under “sequential initialization,” to confirm or quantify rank collapse?  
+2. Comparisons or compositions: How does SWD compare to and combine with ReGraMa or Plasticity Injection on a small DMC subset (including an SWD+X variant) to establish additivity?  
+3. Sensitivity and heuristics: Please add curves for $T$ and $w_{\min}$. Is a simple heuristic (for example, $T$ as a fraction of buffer capacity or tied to a gradient-norm half-life) a reasonable default for new domains?
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/65Ai8mLfjI.md

@@ -0,0 +1,145 @@
+## Human Reviewer 1
+
+### Summary
+This paper explores the impact of modulation-based text conditioning on text-to-image diffusion models. The authors demonstrate that this technique is an important factor in increasing the quality of generated images. They introduce a simple, training-free method that improves performance across various diffusion models without imposing any additional computational burden at inference time.
+
+### Strengths
+- The paper is well-written, presenting its concepts and results with clarity.
+- The proposed approach is easy to apply, computationally inexpensive, and demonstrably improves generation quality.
+- The method is validated through a sound evaluation on state-of-the-art models, confirming its effectiveness.
+- A significant advantage is the method's broad applicability, as it can be used even with models that do not rely on a CLIP text encoder.
+
+### Weaknesses
+- Based on the observation presented in Table 1 that adding CLIP embeddings can increase the quality of images generated from short prompts, it would be beneficial if the authors would also separate their evaluation based on the criteria of prompt length, to demonstrate that modulation can increase generation quality even with long prompts. A more detailed analysis would strengthen the method's reliability.
+- The evaluation lacks a comparison to common, practical methods for quality enhancement. For example, many users simply add phrases like "good quality image" or "very detailed image" to their prompts or use negative guidance, which is a standard feature in many diffusion model interfaces.
+- A potential trade-off between the enforced modulation and prompt fidelity is not explored. If the aesthetic qualities introduced by the modulation contradict the user's explicit request in a prompt, it could negatively impact prompt-following. An exploration of this dynamic would strengthen the submission.
+
+### Questions
+- The quality improvement shown in Figure 5 appears to stem largely from guiding the outputs to look more photographic, particularly with the introduction of blurred backgrounds. While this is often desirable, it raises a question about the trade-off between this aesthetic guidance and prompt fidelity. For example, if a user explicitly asks for a plain background, will the modulation override this request to produce a more "image-like" result with depth of field? Have the authors evaluated this trade-off?
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper investigates the role of the CLIP modulation component within Text-to-Image (T2I) diffusion models. The authors begin by questioning its necessity, demonstrating through ablation that its removal has a minimal impact on overall generation performance. Despite this finding, the authors argue that this component enable controllable shifts in the generation process. They propose a novel method that involves altering the modulation guidance at different blocks of the diffusion transofmer. This claim is supported by a thorough quantitative analysis across several T2I models and is further extended to text-to-video models. The authors show that their method improves generation quality in terms of aesthetics and complexity. They also demonstrate that it mitigates common generation failures, such as incorrect object counts and anomalous finger generation.
+
+### Strengths
+The paper is clearly written and well-structured. A primary strength lies in its comprehensive experimental validation. The experiments are thorough and are conducted on 4 T2I models that are trained with CLIP modulation, and even included additional model that was not using CLIP, training it to incorporate CLIP modulation. Furthermore, they include text-to-video models, thereby broadening the applicability of their findings.
+
+### Weaknesses
+The primary weakness is the limited novelty of the method. This method (with the exception of choosing the dynamic modulation strategies) was already presented in [1] as a naive approach (Equation 2). If the authors disagree, I would be happy to discuss and understand the novelty better.
+
+A second, smaller weakness, concerns the justification for the proposed dynamic modulation strategies. These strategies are heuristically derived from observed attention patterns within the model. This reliance is a potential weakness, as attention weights are not always a reliable or faithful indicator [2] of a model's internal semantic processing at different hierarchical levels. While the authors attempt to validate these attention-driven heuristics through an ablation study, the results appear inconclusive and fail to provide a definitive justification for the chosen strategies.
+
+[1] TokenVerse: Versatile Multi-concept Personalization in Token Modulation Space (Garibi and Yadin et al. 2025)
+[2] Attention is not Explanation. (Jain et al. 2019)
+
+### Questions
+- The analysis suggests that CLIP modulation is more influential for short prompts compared to long prompts. Do the authors have a hypothesis for this observed phenomenon?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper revisits the role of global text conditioning in diffusion transformers. In response to the prevailing trend of abandoning modulation mechanisms in favor of attention-only approaches, the authors demonstrate through analysis that while conventionally used pooled text embeddings contribute limited benefits, repurposing them as a guidance mechanism can effectively adjust the diffusion trajectory toward more desirable attributes. This approach, termed modulation guidance, is training-free, straightforward to implement, and enhances performance across multiple tasks including text-to-image, text-to-video generation, and image editing.
+
+### Strengths
+1. The revisiting and discovery that global text conditioning can be leveraged as a powerful control signal—rather than being merely a passive input—is novel. The proposed dynamic modulation guidance demonstrates a clear ability to address classic and stubborn challenges in T2I generation, such as hand synthesis and object counting, which is a significant finding.
+
+2. The paper is impressive in its extensive experimental scope, demonstrating effectiveness across a diverse set of tasks—including text-to-image, text-to-video, and instruction-guided editing—and model architectures, encompassing transformer-based DMs and the CLIP-free COSMOS model.
+
+3. The paper is well-structured and easy to follow.
+
+### Weaknesses
+I have the following two major questions:
+
+1. I noticed that different hyperparameters are used for different tasks and generation types/styles In Tab.5. Could the authors provide more detailed guidance on the process of selecting the appropriate strategy and its associated hyperparameters for a **new, unseen task**? Is this process largely heuristic, requiring manual search for each new situation, or are there general principles or a methodology that can be derived from the observations in Figure 3 to make this selection more systematic?
+
+2. **Connection and Distinction to h-space Methods:** The work [1] demonstrates that diffusion models possess a semantic latent space (h-space) and that rescaling the difference in latent features ($\Delta h$) can control attribute strength. Could the authors discuss the primary distinction between their method and this prior work? Specifically, is it possible to achieve a similar guidance effect by simply rescaling $\Delta h$, analogous to Equation 3 in this paper, instead of explicitly using the CLIP embedding to compute $y(p^+, t) - y(p^-, t)$?
+
+[1] Mingi Kwon, Jaeseok Jeong, Youngjung Uh. "Diffusion Models Already Have a Semantic Latent Space". *ICLR*, 2023.
+
+### Questions
+Please see my **Weaknesses** part.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper revisits the role of global (pooled) text conditioning in diffusion transformers, which has recently been discarded in favor of attention-only conditioning. The authors first demonstrate through empirical analysis that the pooled CLIP embedding contributes little to generation quality in several state-of-the-art models (e.g., FLUX, HiDream-Fast), especially with long prompts. However, they propose a novel perspective: repurposing the pooled embedding for modulation guidance—a training-free, plug-and-play technique that steers the diffusion process toward desirable visual properties (e.g., aesthetics, complexity, hand realism) by extrapolating in the modulation space using positive/negative prompt pairs. The method is simple, incurs negligible overhead, works with or without classifier-free guidance (CFG), and can be retrofitted into models that originally lack pooled embeddings. Extensive experiments across text-to-image, text-to-video, and image editing tasks show consistent improvements in human evaluations and automatic metrics.
+
+### Strengths
+1. The paper provides a clear and convincing empirical investigation into why global text conditioning appears ineffective in current models, filling an important gap in understanding.
+
+2. Modulation guidance is training-free, easy to implement, computationally lightweight, and broadly applicable across architectures and tasks.
+
+### Weaknesses
+1. The core idea of using pooled embeddings for guidance resembles prior work on semantic directions in GANs (e.g., StyleGAN) and recent methods like TokenVerse or Concept Sliders, though the application to modulation space in diffusion transformers is new.
+
+2. The method introduces new hyperparameters (guidance scale w, layer indices), requiring tuning for different tasks—though ablations help, this adds complexity compared to plug-and-play baselines.
+
+### Questions
+NA
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+4
+
+### Confidence
+3

human_reviews/6P5sAycAQr.md

@@ -0,0 +1,129 @@
+## Human Reviewer 1
+
+### Summary
+This paper organically integrates two lines of zero-shot CLIP enhancement methods—text prompt augmentation and hierarchy-based approaches—into a unified workflow to improve CLIP’s zero-shot classification performance.
+
+### Strengths
+The proposed method demonstrates a clear and rigorous workflow, supported by significant and well-validated experimental results.
+
+### Weaknesses
+The performance of the proposed method may potentially depend on the class structure of the classification task. 
+
+Although the paper emphasizes that taxonomic context is essential for classification, there is limited analysis supporting this claim beyond the reported accuracy of the proposed method. In Table 4, removing the taxonomic context (W-TaxS) outperforms removing the descriptors (TaxCLIP) in several cases, which does not appear to substantiate the assertion that taxonomic context is essential.
+
+### Questions
+Could you please include additional experimental results comparing the proposed method with existing baselines on fine-grained benchmarks, such as Oxford Flowers 102 and Stanford Cars?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper presents a simple approach for prompt adaptation to improve zero-shot classification of Vision Language models such as CLIP. The authors propose constructing prompts with supercategories appended to prompts for each class. Evaluation is conducted over multiple benchmarks. Ablations and improvements over baselines is shown.
+
+### Strengths
+The authors show improvement over baselines for most datasets. Extensive evaluation includes performance of the method over various CLIP architecures as well as ablations over different aspects such as prompt construction choices.
+
+### Weaknesses
+Novelty is very limited. There is very little difference conceptually wrt CHiLS and CGPT-P (cited in the paper). Why have the authors decided to only include one level in their hierarchical tree? This has not been discussed. What happens if you construct a deeper tree? This also brings into concern the fact that simply appending the super category name to the class prompt should be suboptimal as CLIP often suffers with long context without specialized fine-tuning. The performance reported does not match numbers in other papers like CHiLS for some datasets. What could cause this? Table 1 does not state the numbers are for which CLIP architecture. The lack of examples of what the texts look like also leads to confusion about the reason why this approach will work better than CHiLS and CGPT-P.
+
+### Questions
+Please refer to questions in weaknesses.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces DefNTaxS, which addresses taxonomic ambiguity in zero-shot classification by automatically generating LLM-based taxonomic context for class prompts. Building on D-CLIP's approach of adding descriptors to class names, DefNTaxS extends the template to include subcategory information: while D-CLIP uses "[class] which [has/is] [descriptor]", DefNTaxS adds taxonomic context as "[class] which [has/is] [descriptor], [contextual phrase] [subcategory]". The method uses GPT-X to discover subcategories, assign classes (targeting ~20 classes per group), and generate natural contextual phrases, achieving +5.5% average improvement over CLIP and +2.44% over D-CLIP across seven benchmarks without requiring model retraining.
+
+### Strengths
+The work explores an interesting aspect of vision-text alignment by investigating how taxonomic context can help resolve semantic ambiguities in CLIP-style models, addressing a real limitation where classes like "boxer" could refer to dogs or athletes depending on dataset context.
+
+### Weaknesses
+-- marginal extension over existing work: The contribution is an incremental improvement over D-CLIP and related LLM-prompting methods (WaffleCLIP, CuPL, CHiLS), essentially adding one more field to existing prompt templates. The taxonomic discovery process reduces to basic LLM prompting with heuristic post-processing rules, offering limited technical novelty beyond existing descriptor-based approaches.
+
+-- small performance gains: The improvements over D-CLIP are modest (+2.44% average, table 1) and inconsistent across datasets, with some showing minimal gains (+0.16% on Places365) or even degradation (-1.05% on Food101). 
+
+-- limited technical and scientific contribution: the paper lacks architectural insights about vision-language models and proposes no learnable components or model modifications. The study is mainly about prompt manipulation, which doesn't advance our understanding of why CLIP struggles with ambiguity or how to fundamentally improve VLM architectures.
+
+### Questions
+--
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper extends existing works on image-text matching within VLMs such as CLIP by augmenting class descriptors with taxonomical context, achieving a gain of on average 5.5% over CLIP on standard benchmarks.
+
+### Strengths
+The paper is well written and simple to follow, and on standard benchmarks compared to methods such as D-CLIP or WaffleCLIP, consistent performance improvements can be found.
+
+### Weaknesses
+I have several issues with the proposed setup, and would love for the authors to provide more context here:
+* Why would DEFNTAXS actually be a meaningful improvement over D-CLIP, which itself has significant issues, as descriptors proposed by LLMs are not necessary to be found in given query images, see Roth et al. 2023? The problems are the same, it's just that more context is provided that may or may not match the query image, no? I.e., adding "commonly found in kitchen utils" would maybe help to differentate a fork against, say, a motorcycle, but the fail when contrasting between potentially a single fork, and a general image on kitchen utensils, since redundancy is introduced.
+* Importantly, a lot of additional semantic information is introduced into a language encoder that is known to be very weak and often fails to distinguish finegrained or multiples of semantic features. It would be important for the authors to offer some support that the CLIP model actually meaningfully and explicitly leverages the semantic context; and its not just better chosen structured noise which is given the additional performance gains.
+
+### Questions
+See weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4

human_reviews/6i1jVAYbHs.md

@@ -0,0 +1,178 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes AtlasKV, a parametric knowledge augmentation method designed to integrate billion-scale Knowledge Graphs (KGs) into Large Language Models (LLMs) with extremely low VRAM cost (under 20GB). The core contributions are twofold: 1) KG2KV, a method for transforming KG triples into high-quality Query-Key-Value (Q-K-V) data, which aims to improve generalization by increasing query diversity compared to prior work. 2) HiKVP (Hierarchical Key-Value Pruning), a novel inference-time algorithm that organizes keys into a 3-layer hierarchy, enabling sub-linear ($\mathcal{O}(\sqrt[3]{M})$) time and memory complexity. This allows the model to scale to 1B triples while avoiding the linear cost growth seen in methods like KBLaM.
+
+### Strengths
+- Excellent Scalability: The primary strength of this paper is the HiKVP algorithm. It provides a credible path to scaling parametric knowledge augmentation to the billion-fact scale, a task that seems infeasible for prior methods like KBLaM. The experimental results in Figure 4, showing the ability to handle 1B triples with <20GB VRAM while KBLaM fails at 100K triples, are very impressive.
+
+- Strong Generalization and Robustness: The KG2KV method effectively addresses the limited query diversity of template-based synthetic data used in prior work. This is validated by the strong results in Table 3, where AtlasKV (even with pruning) massively outperforms KBLaM on OOD datasets, especially complex ones like ATLAS-Pes20-QKV (e.g., 52.7% vs 5.5% on 10³ triples).
+
+- Addresses a Critical Problem: Efficiently and scalably augmenting LLMs with vast, external factual knowledge is a fundamental goal in AI. This paper tackles this problem head-on with a novel parametric approach that avoids the high inference latency of RAG.
+
+### Weaknesses
+- Major Flaw in Experimental Validation: The paper's core premise is that scaling up KGs to 1B triples is valuable. However, the experiments fail to demonstrate this value. Table 3, the primary knowledge grounding experiment, only shows that as the number of "candidate triples" (i.e., noise) increases, performance decreases. This experiment proves "robustness to noise" but critically fails to prove the benefit of "knowledge coverage".
+The paper is missing the most crucial experiment: a comparison between a small KG and the large KG on a set of questions where the answers only exist in the large KG. Without this, the motivation for the entire paper—scaling up—is left as an unsubstantiated assumption. As it stands, the experiments show that scaling up (the candidate pool) only hurts performance.
+
+- Unaccounted LLM Cost & Unfair Comparison: The KG2KV method (Section 4.1, Appendix H) relies on external, powerful LLMs (GPT-4o-mini, and the authors state in Appendix I that Claude-4-sonnet was used for development) to transform KG relations into natural noun phrases. This introduces a significant, one-off (but potentially massive) computational and/or API cost for preprocessing the 1B+ triples. This cost is completely ignored in the paper's efficiency analysis. This makes the comparison to methods like KBLaM, which uses simple synthetic templates, an unfair, apples-to-oranges comparison. The scalability of AtlasKV is thus partially achieved by offloading significant complexity to an external, unmeasured LLM.
+
+- Loss of Graph Structure: The KG2KV method inherently "flattens" the knowledge graph, treating each triple as an independent fact. This sacrifices all topological/structural information of the KG, limiting the method to one-hop retrieval and precluding multi-hop reasoning. This is a significant trade-off that should be more explicitly acknowledged and discussed as a limitation.
+
+### Questions
+- Could the authors provide an experiment that demonstrates the value of "knowledge coverage"? For example, by comparing a model trained on a 1M-triple KG vs. a 100M-triple KG on a set of questions where the answers only exist in the larger KG? This seems essential to justify the paper's core motivation.
+
+- Can the authors quantify the computational cost (e.g., in GPU hours or total API costs) of the KG2KV preprocessing step for the 1B-triple ATLAS-Wiki dataset? How does this one-off cost compare to the training/inference costs, and how does it affect the overall efficiency claim?
+
+- The performance of AtlasKV with HiKVP is notably lower than AtlasKV without HiKVP (e.g., Table 3, ATLAS-CC-QA, 10³ triples: 74.5% vs 96.4%). This suggests the hierarchical pruning is quite lossy. Have the authors explored strategies to mitigate this precision drop (e.g., different top-k values, as hinted at in Appendix B.2), or is this considered an acceptable trade-off for the achieved scalability?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes a relatively systematic optimization framework to address the high computational complexity of knowledge-augmented large language models (KG-augmented LLMs). It provides both theoretical and experimental analyses of the model’s time and space complexity as well as GPU memory consumption, showing a certain level of exploratory research value. The designs of the hierarchical knowledge base and Rectangular Attention modules are instructive. In addition, the authors present relatively complete complexity derivations and visualization results, which enhance the technical rigor of the work.
+
+### Strengths
+1. **Novel and principled framework:**
+
+The transformation from KG triples to Q-K-V structures is conceptually elegant and theoretically aligned with transformer attention, effectively bridging symbolic and neural representations.
+
+2. **Scalable design and sub-linear efficiency:**
+
+The hierarchical pruning mechanism (HiKVP) achieves sub-linear time and memory complexity, making it feasible to integrate billion-scale KGs in limited VRAM environments.
+
+3. **Strong empirical results:**
+
+Extensive experiments demonstrate clear improvements in both knowledge grounding and generalization, with substantial reductions in memory cost compared to KBLaM and RAG.
+
+### Weaknesses
+1. The methodology section is difficult to follow. The main issue is that many symbols are not clearly defined. For example, the meaning of $S$ on Line 253 and the correspondence of variables in Eq. (3) and Eq. (13). In addition, the paper does not state the model’s optimization objective, which further increases the burden on readers trying to understand the work.
+
+2. The paper’s primary contribution appears to be complexity optimization for KG-augmented LLMs. However, relying solely on time/space complexity analyses and VRAM accounting is not sufficient to fully assess this advantage. For instance, when performing top-k pruning over a hierarchical KB, switching among three levels of keys can induce frequent CPU–GPU I/O, which is a non-negligible hardware latency during real-world training/inference. I recommend discussing this overhead explicitly.
+
+3. Missing parameter sensitivity analysis. To my knowledge, Rectangular Attention in KBLaM is not invoked at every LLM layer; it is typically applied every few layers. I could not find this setting in the main text (it might be in the appendix, which I may have overlooked). Moreover, the paper does not provide any rationale or analysis for such a setting.
+
+4. Why three hierarchy levels? The choice of a 3-level hierarchical key-value cache is insufficiently justified. Please add more detailed theoretical or empirical support for this design choice.
+
+5. Inter-level correlation in top-k selection is not clear. Are the top-k selections across different hierarchy levels correlated? Figure 6 only shows the variation of a single variable. Given that top-k pruning at an upper level can influence the quality of pruning at the next level, an experimental analysis of this inter-level dependency seems necessary.
+
+6. I am curious about the appendix proof of equivalence for Rectangular Attention. It appears to be largely a distributive rearrangement of matrix operators that, in practice, does not change the computed result. Please provide a deeper explanation of its effectiveness. Alternatively, clarify whether this operator manipulation merely serves as an algebraic device to facilitate subsequent complexity analysis for hierarchical key-value computation.
+
+### Questions
+Please See weakness
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes AtlasKV, a parametric method for integrating billion-scale KGs into LLMs without external retrievers, long contexts, or retraining.
+
+It addresses limitations of non-parametric RAG (latency, retrieval bottlenecks) and traditional parametric methods (retraining costs, poor scalability)
+
+Key innovations include KG2KV, which converts KG triples into query-key-value (Q-K-V) data aligned with LLM attention mechanisms, and HiKVP, a hierarchical pruning algorithm for sub-linear time/memory complexity during inference.
+
+Experiments demonstrate superior empirical performance.
+
+### Strengths
+1. AtlasKV shifts from retrieval-heavy RAG to parametric KG injection, enabling training-free adaptation to new KGs. This idea is both novel and practical.
+
+2. The method is well-motivated, with clear formalisms for KG2KV and HiKVP. Such formulation maintains generalization without fine-tuning.
+
+3. Results show strong gains in knowledge-intensive tasks.
+
+### Weaknesses
+1. HiKVP's pruning is efficient, but lacks formal guarantees or error analysis. 
+
+2. Lack comparison with GraphRAG algorithms, such as HippoRAG and Raptor. I understand the focus of this work is on large KGs, where existing graphRAG algorithms can struggle to handle especially for indexing latency and token consumption. But comparing with them on small datasets can benefit the understanding the pros and cons of these methods.
+
+### Questions
+See above.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper introduces AtlasKV, a parametric knowledge integration framework for large language models (LLMs) that enables augmentation with billion-scale knowledge graphs (KGs) under modest GPU memory budgets (<20GB VRAM). The method converts KG triples into query–key–value (QKV) training data aligned with LLM attention structures. It also employs a hierarchical key–value pruning algorithm that reduces computational and memory overhead while preserving grounding accuracy. Extensive experiments on large-scale KG datasets (ATLAS family, Enron, etc.) demonstrate that AtlasKV outperforms both non-parametric RAG methods and parametric baselines (e.g., KBLaM) in terms of scalability, knowledge grounding, and generalization to out-of-distribution (OOD) queries.
+
+### Strengths
+- The combination of KG2KV and HiKVP is elegant, bridging symbolic KG structure with LLM attention in a natural way.
+
+- The proposed method demonstrates integration of 1B triples with <20GB VRAM, a significant improvement over KBLaM (>40GB for 100K triples).
+
+- Experiments across multiple datasets (ATLAS-CC, ATLAS-Pes2o, Enron) with ablations (entity vs. event masking, pruning strategies, encoders) show the effectiveness of the method.
+
+### Weaknesses
+- KG2KV requires relation rewriting (e.g., "because" -> "cause"), which relies on external LLMs. This introduces dependency on external models and may limit reproducibility.
+
+- The baselines are strong (KBLaM, RAG), but newer lightweight RAG variants (e.g., caching-based or hybrid symbolic-neural methods) are not included.
+
+- The claim that only 20K KG2KV samples suffice for generalization is intriguing, but more discussion is needed on why such small-scale training suffices for billion-scale integration.
+
+### Questions
+- How sensitive is AtlasKV to the quality of relation rewriting in KG2KV? Could noisy or ambiguous relations degrade performance?
+
+- How does AtlasKV handle evolving KGs (e.g., streaming updates)? Is retraining or re-encoding required?
+
+- Could AtlasKV be combined with lightweight retrieval (hybrid parametric + non-parametric) for further efficiency gains?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3

human_reviews/7gLNA6nT5d.md

@@ -0,0 +1,185 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes to explicitly add n-gram induction heads to transformer model for in-context reinforcement learning. It can improve the data efficiency, address the simplicity bias and make the model more stable and less sensitive to hyperparameters. Experiments show that the proposed method can outperform the algorithm distillation baseline.
+
+### Strengths
+1. The idea of introducing n-gram induction head is simple and easy to implement.
+2. It can improve the data efficiency and reduce the hyperparameter sensitivity.
+3. The proposed n-gram head can also be applicable to visual environment settings.
+
+### Weaknesses
+1. Section 2.2: the description of the whole model structure is unclear. The formula in line152 needs more detailed explanation, like how the attention score of n-gram head is normalized, what if there are no n-grams in the history, if there is only one head in each NG layer. The authors should better provide a pseudo-code or the architecture figure.
+
+2. What is the architecture of your base transformer model? What is the number of layers and the total size of the model? Do you try using a powerful pretrained model to see if the benefit of 2-gram information still helps? There should also be experiments on different model size to further validate the effectiveness of the n-gram head. 
+
+3. The limited generalizability. The motivation of the 2-gram induction head  is to mitigate the simplicity bias and accelerate the emergence of in-context learning. 
+While effective at small scales and with limited data, this architectural improvement is fundamentally a hardcoded bias for simple, short-term pattern recognition (specifically, 2-gram statistics in the experiments, 3-gram will hurt the performance). In the era of increasingly large transformers and datasets, the attention itself can emerge the complex induction heads and even more sophisticated learning algorithms, which makes the 2-gram information redundant. Moreover, in real-world scenarios, the basic in-context learning capability is already strong, we do not need to train the transformer from scratch for the emergence of in-context RL, and the n-gram information is insufficient for distilling complex RL algorithms that require abstract reasoning instead of the local sequential patterns.
+
+### Questions
+Typos: The title of the paper is wrong.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes N-gram attention in ICRL to reduce the data required in training process but achieve the similar performance of AD. Evaluations on some visual-based scenarios further show its edge.
+
+### Strengths
+1. The paper is easy to follow.
+2. The figure is well-situated and clear.
+
+### Weaknesses
+As I'm not the expert in this area, I cannot provide some professional review here. Below are my comments:
+
+1. The work lacks formal analysis explaining why N-gram heads yield stability gains and reduced data requirement. Maybe the authors should provide some theoretical perspective to elaborate this.
+2. Could the authors add more experiments on some diverse visual tasks to further showcase the superiority ?
+
+### Questions
+See weakness above
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper integrates n-gram introduction heads into transformers for In-Context RL. This integration aims to resolve two challenges in Algorithm Distillation: (1) the low data efficiency in ICRL; (2) the training instability in emerge the in-context capability. The authors apply explicit n-gram introduction heads to the transformer backbone and test n-gram heads in Dark Room and Key-to-Door environments. The authors also propose a scheme to extend the environments to image states. The results show that the n-gram heads can significantly reduce the data requirement and improves hyperparameter robustness.
+
+### Strengths
+1. This paper proposes a very useful improvements to AD by integrating n-gram heads to transformer. This architectural improvement significantly mitigates the disadvantage of AD of data overhead, hyperparameter tuning, and training instability
+1. The empirical study is very comprehensive that answers all my concerns towards the n-gram heads architecture application in AD. Section 4 strengths my confidence in applying this technique to ICRL research.
+1. The VQ technique provides a feasible approach to image-based environments which extends the n-gram heads architecture to broader application scenarios.
+1. The authors give detailed description on how to sweep and choose the hyperparameters. This information is rather important to such empirical papers for reproduction.
+
+### Weaknesses
+1. This paper demonstrates the benifits of n-gram heads by expriments in two environments and their variants in Miniworld. However, these environments are relatively simple. I suggest the authors test n-gram heads in more complicated continuous control tasks (e.g., MuJoCo, Meta world) to show the generalization.
+1. In addition to the experiments, I would expect theoretical analysis for why explicit n-gram heads can help improve data efficiency and traning stability in Algorithm Distillation, as n-gram has been studied for decades.
+1. As VQ and ResNet are used in image-based tasks, I suggest the authors also analyze the training cost sensitivity for these modules. They are important to prove the adaptation and efficiency in pixel-based environment.
+
+Minor issue:
+1. The title is missing in the paper.
+1. The figure positions mismatch the text. For example, I have to jump pages to find Fig. 2 and Fig. 4 when I read the paper. It would be nice to rearrange the paper so that the figures are right in the place where they are referred.
+
+### Questions
+1. Could the authors compare with other baselines, for example, Noise Distillation [1] mentioned in the paper? This will help clarify whether the observed data-efficiency and stability gains stem specifically from the proposed n-gram inductive bias, or whether similar improvements could be achieved through alternative data-centric approaches.
+1. Could the authors test the method in high dimensional continuous control environments, e.g., MuJoCo and Meta world? These are widely used benchmarks to evaluate the generalization and scalability of RL algorithm, and would help demonstrate whether the proposed n-gram heads can extend the scope to more complex, real-world control tasks.
+
+[1] Zisman, Ilya, et al. "Emergence of In-Context Reinforcement Learning from Noise Distillation." International Conference on Machine Learning. PMLR, 2024.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper examines whether augmenting the architecture with n-gram attention heads improves the performance of Algorithm Distillation [17].
+
+### Strengths
+N/A
+
+### Weaknesses
+The scope of this work is narrow, as it focuses solely on improving Algorithm Distillation [17], and it is unclear whether the findings generalize beyond this specific approach. Given the limited experimental scale, its relevance to understanding in-context learning in large transformer models is also unclear.
+
+### Questions
+The title in the PDF file ("FORMATTING INSTRUCTIONS FOR ICLR 2026 CONFERENCE SUBMISSIONS") should be corrected.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+This paper integrates n-gram induction heads into transformer architectures for in-context reinforcement learning (ICRL). Building on Algorithm Distillation, the authors show that explicitly incorporating n-gram attention patterns reduces data requirements (up to 27× fewer transitions) and makes hyperparameter search easier, and works across both discrete and pixel-based environments. The method is validated on Dark Room, Key-to-Door, and Miniworld environments.
+
+### Strengths
+S1: The paper addresses a real practical problem in ICRL: Algorithm Distillation's demand for large datsets, and sometime slow convergence or getting stuck.
+
+S1: The paper tests their method on different setups, ones with more grid based states and another with rgb 64x64 pixel. A diverse set of tasks reinforces the findings.
+
+S3: Clear presentation with intuitive figures.
+
+### Weaknesses
+W1: This is my main concern. This work seems to not disentangle three questions: 1) how does n-gram heads general help the training of transformers 2) why this is espeically important (or not) in ICRL tasks 3) is the gain via n-gram heads really not achievable by other hyperparameter tuning or different architectural tricks.
+
+W2: It is very hard to takeaway generalizable insights on why n-gram heads help and when we should be using these.
+
+W3: Lack of serious comparison. The method is only compared against vanilla AD. There has been many variants on the algorithmic side, and at the same time many improvements on the transformer architecture and initialization.
+
+W4: Overselling of generality. The title and abstract suggest n-grams solve ICRL's data problems broadly, but it is hard to understand whether these gains are precisely attributable to n-gram heads or will be simply addressed by data scale or better optimization.
+
+### Questions
+Q: What is the positional encoding used in the transformer?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4

human_reviews/8J3GTeQmwl.md

@@ -0,0 +1,171 @@
+## Human Reviewer 1
+
+### Summary
+This paper addresses the problem of model selection in graphon estimation by proposing a new cross-validation method for selecting tuning parameters and estimation approaches. The authors prove that their cross-validation score is asymptotically aligned with the estimation error. Extensive numerical simulations and real-data analyses are provided.
+
+### Strengths
+1. I feel the proposed method quite smart. By randomly replacing a portion of the edges, the authors introduce a controllable bias, while obtaining edges that are (conditionally) independent of the training set for validation. The idea is interesting and provides insights for existing research.
+	
+2.	The authors show that minimizing $V_k(M)$ approximately corresponds to minimizing $L(M)$, which provides a theoretical justification for the proposed method.
+	
+3. The authors have provided sufficient numerical simulations and applied their method to real data examples.
+	
+4. I enjoyed reading the paper. The writing is clear and well-structured.
+
+### Weaknesses
+1. It is confused whether $v_i$ in Line 97 is a typo, since it is unusual for a node label to be a random variable. In addition, if $v_i$ is intended to be $\mu_i$, then $p_{ij}$ would be random, see my Questions 1 and 2.
+	
+2. It is unusual that the paper does not impose any assumptions on the graphon function. It is unclear what role the graphon function plays in their framework.
+
+3. The proofs need to be written more rigorously. See my Question 2, 5.
+
+### Questions
+1. Could the authors clarify whether line 560 contains a type error? Should it perhaps be something like $P((v_i,v_j)\in S_k)P(b_{ij}=1, a_{i'j'}=1) + P((v_i,v_j)\notin S_k)P(a_{ij}=1,a_{i'j'}=1)?$ In addition, could the authors explain how the term $P(a_{ij}=1)P(a_{i'j'}=1)$ in line 561 is obtained? If I understand correctly, the edges $(i,j)$ and $(i',j')$ are correlated when $i=i'$ since they share the same $\mu_i$.
+	
+2. Line 597 could be made clearer. More precisely, the statement holds conditional on all $\mu_i$'s.
+		
+3. If all $\mu_i$'s are treated as deterministic, then it is unclear why the paper introduces the graphon function.
+		
+4.  It would be helpful if the authors could comment on the computational complexity of calculating $V_K(M)$, for example when $K\asymp n$.
+		
+5.  Could the authors clarify why Line 610 is correct? It seems that $S_k$ is random.
+		
+		
+6. It would be helpful if the authors could provide more clarification on when  Condition 1 is satisfied. For instance, including a toy example or specifying sufficient conditions would make it clearer. In addition, could the authors comment on whether this condition depends on $w_k$ and $\theta$?
+	
+7. It seems that the theorem holds for any fixed $\theta$. What would happen if $\theta = 0$ or 1? In addition, in the appendix the authors set $\theta$ as a random variable, then how would this affect the validity of the theorem?
+
+
+8. It would be helpful if the authors could provide a theoretical comparison between their method and ECV in terms of implementation time.
+	
+9. It would be interesting to know how the theoretical results behave when $p_{ij}$ tends to zero as $n$ increases.
+	
+10.  Some minor errors: 
+	
+Line 158: $w_k\theta$: $w_k\theta 11^\top$.
+
+Line 249, 251: five: four.
+
+Line 573: equation equation 8: equation 8.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces an improved cross-validation method with random imputation, which is unbiased and efficient compared to existing methods including edge sampling approaches and the matrix completion method proposed by Li et al. (2020). The paper compares against these baselines across multiple graphon estimation methods, demonstrating effectiveness in terms of both accuracy and computational efficiency. The authors provide theoretical justification showing that their cross-validation score is asymptotically correct. I found that this paper is strong and provides a comprehensive analysis based on both theory and numerical evidence. The case studies are insightful and demonstrate practical applicability of the proposed method.
+
+### Strengths
+1. The authors effectively articulate the fundamental challenge in applying cross-validation to network data, particularly how traditional edge sampling destroys network topology.
+
+2.  The random imputation strategy is quite simple and very effective. I could not think of any simpler way.
+
+3. The paper provides extensive experiments across multiple graphon models (varying in density and rank properties) and estimation methods (NS, SAS, USVT, ICE). Section 6's case studies are particularly compelling.
+
+4. Unlike ECV which requires low-rank assumptions and only has theoretical guarantees for specific models (SBM, RDPG), CV-imputation works for the general graphon model class
+
+### Weaknesses
+1. I could not follow the theoretical justification precisely, but I get the intuition behind it which is that the training and test data are independent of each other. Therefore, it doesn't change the distribution of the edge probabilities. It is worthwhile to note that this is valid only when the presence orabsense of an edge is independent of that of other edges.
+
+2.  While the paper tests networks up to approximately 2,600 nodes, many real-world networks contain tens or hundreds of thousands of nodes, yet no discussion addresses computational or memory requirements at such scales. I understand to some extent because the number of pairs for increases quadratically with respect to the number of nodes. But it is nicer to discuss a solution to scale up the method.
+
+### Questions
+I am unclear why the computational time differs across estimation methods (Figure 3), since the cross-validation procedure itself should be independent of the estimation method being evaluated. I suspect the authors calculated the total CPU time, which includes both the cross-validation procedure and the estimation method execution, rather than isolating the cross-validation overhead alone. This make it unclear the actual speedup specifically attributable to CV-imputation compared to ECV. The authors should clarify what operations are included in their reported CPU times and, ideally, provide a breakdown separating cross-validation overhead from estimation costs.
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes a new cross-validation approach for graphon estimation, introducing a random-imputation strategy to handle dependency structures in networks and to tune hyperparameters from a single observed graph. The authors argue that existing edge-sampling methods lead to bias by degrading network connectivity, and they provide a careful theoretical treatment to show consistency of their approach. The work is technically sound and well written, but the contribution represents a relatively narrow conceptual advance within the specific context of graphon models rather than a broader methodological step forward for network machine learning. The empirical results, while generally positive, do not convincingly demonstrate consistent or substantial gains over edge-based cross-validation (ECV). The paper would be strengthened by a deeper analysis of when and why ECV fails in practice, clearer discussion of how the proposed method could generalize beyond the graphon setting, and a more comprehensive evaluation showing robust and meaningful performance improvements across a wider range of network types.
+
+### Strengths
+The paper addresses a legitimate technical challenge: how to form cross-validation sets in under dependence in networks. The proposed CV-imputation scheme is clearly described, mathematically justified, and accompanied by consistency proof. The implementation appears efficient, and the experiments consider several graphon estimators (NS, SAS, USVT, ICE), as well as both synthetic and real-world graph datasets. The manuscript is generally clear and thorough.
+
+### Weaknesses
+(1)	Scope and generality. The method applies specifically to graphon models and depends on smoothness and exchangeability assumptions. It is unclear how the proposed theoretical framework or random-imputation idea would extend to broader classes of network models (e.g., latent-space, temporal, or sparsified networks). Thus the paper’s claims of general applicability thus seem overstated.
+(2)	Empirical impact. The experimental gains over ECV are modest. In Table 1, SAS and ICE show no meaningful improvement, suggesting that differences arise mainly from estimator robustness rather than from the proposed validation method. In real-world datasets, results are essentially tied on three of four networks. Only the drug–disease network shows a visible advantage, which is not analyzed in depth. Without more in-depth analysis demonstrating the impact of ECV bias, it is difficult to conclude that the differences are practically important.
+(3)	Efficiency gains. Although Figure 3 shows a computational speed-up relative to ECV, the paper does not provide an asymptotic complexity comparison. Moreover, the statistical efficiency results in Figure 5 are averaged across all graphon models and may be dominated by the weaker performance of NS and USVT, making it hard to assess whether the gains are consistent or meaningful across settings.
+
+### Questions
+See weaknesses.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper proposes CV-imputation, a cross-validation method for selecting graphon-based models on network data. The method is model-agnostic, computationally efficient, and supported by theoretical guarantees of asymptotic consistency. Empirical results show that it outperforms or matches existing methods across various networks.
+
+### Strengths
+1. CV-imputation does not assume a specific form of the graphon, allowing it to be applied across diverse network structures without model restrictions. 
+2. The method avoids expensive singular value decomposition (SVD), reducing runtime and enabling scalability to large networks. 
+3. It is supported by a convergence result showing that its validation criterion aligns with mean squared error minimization in the asymptotic regime.
+
+### Weaknesses
+This work assumes the data comes from a graphon and the goal is to assess graphon-based estimators. However, would such a method be extended beyond this assumption? Discussion about the limitations of applicability would be a nice addition to the paper.
+
+Additional discussion on the order complexity of the method vs the baselines would strengthen the paper.
+
+### Questions
+1. How sensitive is CV-imputation to violations of the graphon assumption? Are the theoretical guarantees still meaningful outside the graphon framework?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+2

human_reviews/8KcjEygedc.md

@@ -0,0 +1,166 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces a principled theoretical framework for data curation, trying to figure out when less data actually works better than more, to challenge the classical "more is more" scaling laws with comprehensive experiments. It derives exact scaling laws for test error under label-agnostic and label-aware curation rules, showing that strategically pruned small datasets can outperform full datasets under conditions like strong generators and abundant data, while also mitigating model collapse.
+
+### Strengths
+1.  The first paper tries to theoretically explain when and why LIMO happens in machine learning by developing a principled theoretical framework for data curation, enabling practitioners to design data-centric pipelines instead of blindly scaling datasets.
+2. Also provides a unifying explanation for contradictory curation strategies.
+3. The writing and presentation look perfect to me.
+
+### Weaknesses
+1. While the paper links theory to LLM math reasoning (AIME benchmark in Table 1&2), it would be better to discuss other LLM-critical tasks (e.g., code generation, natural language understanding, or multi-step reasoning benchmarks) where curation is also widely used. Additionally, the authors do not investigate how the pruning ratio or oracle quality affects performance across different model sizes (e.g., 7B vs. 70B LLMs), leaving uncertainty about whether the LIMO principle holds consistently across LLM scales.
+
+### Questions
+1. The theory is framed around binary classification, but multi-class tasks are definitely more common in practice. Could you discuss whether the optimal pruning strategies could extend to multi-class settings?
+2. In the ImageNet experiment, the generator and pruner are the same pre-trained ViT, which ensures alignment, but real-world CV pipelines may use separate generators (e.g., a synthetic data model) and pruners (e.g., a human-in-the-loop verifier). Could you discuss how misalignment between generator and pruner (e.g., low ρ₊ relative to ρ) would impact the optimal curation strategy, and whether this differs from the LLM case (where generator and pruner are often separate)?
+3. I have another question for the authors, but not necessarily to be answered: LLM training encompasses pre-training, post-training, and stages such as SFT and RL within the fine-tuning process, with distinct training data used for each stage. For the LIMO method discussed in the paper, in which of these training stages is its principle more likely to be reflected?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper develops a teacher–student, high-dimensional model of learning from curated data in order to reconcile two empirically observed regimes in modern LLM/vision training: "More is more" regime (classical scaling laws) in Kaplan et al. (2020) and Hoffmann et al. (2022), where performance improves monotonically with dataset size vs. "Less is more" regime (LIMO/s1–style curation), when the generator is strongly aligned with the target, then selectively keeping only the most informative/hard examples yields strictly lower test error than training on the full set. Formally, the authors setup a theoretical model with Gaussian inputs and labels generated by a linear teacher with known alignment, and the learner as a ridge (square-loss) estimator trained on a pruned sample whose selection is governed by an oracle with its own alignment parameter. Using random-matrix/deterministic-equivalent techniques, the authors derive closed-form high-dimensional limits for the test error after curation. The authors show numerical simulations of their theory and corroborated with empirical simulations.
+
+### Strengths
+* The paper is clearly written and well-organized. The paper has both solid theoretical side and empirical evidence.
+* The paper has deep theoretical building blocks with high dimensional asymptotics of linear models and classification problems. Using both RMT techniques and tools from Feng et al. for classification problems to study the effects of pruning by the oracle. The conclusions are sound as far as I have checked, and the results appear novel to me.
+* The paper's conclusion provides an alternative to neural scaling laws: it is possible to perform with higher accuracy with fewer data with an expert pruner. This also gives a potential remedy to the neural collapse phenomenon where the training procedure seeks to scale with self-generated data---perhaps by scrutiny using the synthetic data (instead of training on them) and picking only harder/easier problems we can avoid model collapse from synthetic data.
+
+### Weaknesses
+Major comments:
+* I think the main inconsistency between the main message and theory is that: both "KE" and "KH" strategies are effectively "less is more". It is just what data should be picked. I think it is a bit of stretch to say "KE" is equivalent to "more is more", while the neural scaling laws scenario should correspond to a very small $\phi$ instead of the "KE" strategy. 
+* The theory relies on linear model with Gaussian covariates, and is limited to the binary classification setup with proportional asymptotics. This is, definitely limited, as is acknowledged also by the authors in the limitation discussions, a good starting point for study the pruning and data curation in large datasets. I think this is only a minor weakness.
+
+Minor comments:
+* A general style suggestion: use \eqref instead of \ref for equations.
+* Page 19, "Corollary xyz". I think this is a placeholder or typo.
+
+### Questions
+* My main questions are in the weaknesses section. I would especially like to see more discussions between "KE" and "more is more". 
+* The paper uses deterministic equivalence to obtain risk asymptotics for the problem (2). While in general, when l is a convex loss function, using the CGMT techniques the similar asymptotics can also be obtained. See Karoui (2013). What do authors think the behavior in this setup? Would the same behavior also be expected?
+
+Karoui, Noureddine El. "Asymptotic behavior of unregularized and ridge-regularized high-dimensional robust regression estimators: rigorous results." arXiv preprint arXiv:1311.2445 (2013).
+
+### Soundness
+4
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This work seeks to develop a theoretical framework for understanding data curation. In other words, they seek a theory that can predict when it is best to use less data (filtered/curated) versus all data (no filtering/curation). The authors are motivated by the fact that there are cases where small datasets have been shown to be beneficial, like LIMO and s1. Their theoretical results assume a Gaussian feature model under a binary classification setting.
+
+### Strengths
+1. The authors use their theory to show that it can predict empirical results to a surprising degree (mean relative error of 1.8% according to App. B), and within these empirical results live the settings they sought to understand: when all data is needed ("more is more") and when curation is needed ("less is more")
+2. Authors consider both label-agnostic and label-aware data curation in their theory
+3. Significant extra detail for figures and proofs are provided in appendix, answering a few of my early questions
+
+### Weaknesses
+1. Arguably the most important feature of the paper is that their theory is predictive of empirical results in a very realistic dataset (ImageNet training of a ViT), but it is not clear from the writing how to use this in practice. In particular: how does someone measure the quality of the generator $\rho$ for a dataset you are given? No commentary on this is given. A bonus would be if the authors gave a step-by-step appendix section for a practitioner on how to use their theory (willing to raise score for clarity on this)
+2. "the base LLM is a strong generator" based on what? how is this claim measured? 
+3.  Why is $n=5000$ considered the "large data" setting in Figure 1? How was this number chosen? That $n$ is smaller than the CIFAR-10 training set...
+
+Minor: 
+- capitalize PDF and CDF on L184
+
+### Questions
+1. Why is the Gaussian feature model noted as a limitation? Wouldn't we expect gaussian distributions of features according to central limit theorem? The authors mention real-world data is "structured" but I find that vague.
+2. How does someone measure the quality of the generator $\rho$ for a dataset you are given?
+3. Is it possible to know the quality of pruning? For example, suppose I curate a dataset of text by removing duplicates, how would a "quality" be assigned to that?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper introduces a rigorous theoretical framework to analyze the efficacy of data curation strategies in high-dimensional learning. It addresses a central paradox in modern ML: When does aggressive data pruning ("Less is More," exemplified by methods like LIMO and s1) outperform training on the full dataset ("More is More")?
+
+The paper model this scenario using high-dimensional binary classification and regression with Gaussian covariates. The setup captures the interactions between a data generator ($w_g$), the ground truth ($w_*$), and a pruning oracle ($w_o$), allowing for label shift. They analyze both label-agnostic (based on difficulty/margin) and label-aware (based on difficulty and correctness) curation.
+
+Using RMT, the paper derive exact asymptotic expressions for the test error in the proportionate scaling limit ($d/n \to \phi$). The central theoretical finding (Theorem 2) identifies a precise phase transition: "Less is More" (specifically, keeping hard examples) is optimal iff data is abundant AND the generator is strong. In all other regimes (scarce data or weak generator), using more data or keeping easy examples is superior.
+
+While the paper is primarily theoretical, it tries rounding itself off with some simulations and Imagenet experiments.
+
+### Strengths
+- **Theoretical Rigor and Exactness:** The primary strength is the sophisticated mathematical analysis. By leveraging RMT, the paper derive *exact* asymptotic formulas for the generalization error (Theorems 1, 3). This allows for a precise characterization of the interplay between data quality ($\rho$), oracle quality ($\rho_*$), and data scale ($\phi$), moving far beyond bounds or heuristics. The derivation of how pruning "deforms" the Marchenko-Pastur law governing the data spectrum is technically impressive and commendable. 
+
+- **Good Resolution of the "Less vs. More" question:** The paper provides a clear and elegant resolution to conflicting empirical observations regarding data scaling. Fig. 1 beautifully illustrates the four key regimes of learning. The insight that the optimal strategy is fundamentally tied to the generator's strength relative to the task difficulty is a major conceptual contribution.
+
+- The framework successfully unifies several recent lines of work, including data pruning strategies and model collapse mitigation. The analysis of label-aware curation (Section 3.2) is highly relevant to modern SFT/RL*F pipelines where reward models filter policy outputs. The demonstration that strategic pruning prevents model collapse (Fig. 3) is practically significant for the stability of iterative training loops.
+
+- The authors provide some practical validation. Synthetic experiments show strong agreement with the theory (App. B, Fig. 4). Crucially, the ImageNet experiments confirm that the qualitative insights transfer to deep learning settings (ViT), and the analysis in Sec. 4.2 provides a compelling explanation for recent contradictory LLM results.
+
+### Weaknesses
+**Linear Models:** The theoretical results are derived under the assumption of linear models and Gaussian covariates. This is a known standard simplification for RMT analysis but limits the direct quantitative applicability to deep neural networks on structured data. However, the empirical results do point towards potential qualitative transfer, but would warrant a more complete exploration. 
+
+**Fixed Oracle:** The framework assumes a fixed pruning oracle ($w_o$). In practice, oracles (e.g., reward models) are often learned and may evolve adaptively during training (e.g., iterative RLHF), which is not captured in the current analysis. Extending the theory to adaptive oracles would be an important future direction.
+
+
+I am **highly skeptical** that this constitutes the explanation for "Less is More" in modern deep learning. The assumptions are too strong, the dynamics of optimization (like SGD) and feature learning are ignored, and the empirical validation is not sufficiently robust to confirm that the mechanisms identified by the theory are the ones driving results in practice.
+
+### Questions
+- There is an inherent balance/tradeoff between optimal strategy vs model collapse mitigation. Could the authors further clarify the contradictory observation between Theorem 2(b) and Fig.3 ? To be more specific, does the optimal strategy shift from Keeping easy back to Keeping Hard when moving from label-agnostic curation to label-aware curation? Maybe an equivalent Theorem 2 for label-aware setting could help.
+
+- An important practical consideration is to identify which of the four regimes in Fig. 1 is one operating in. Practically where ground truth is scarse and strength of model is difficult to estimate, how would the paper suggest reliably diagnose this? I guess there needs to be a detector for this crossover point when KH > KE?
+
+- The RMT analysis relies on linear models, while experiments elude to ImageNet teasing the qualitative transfer, but the theoretical boundaries remain tied. Can the paper describe the mechanisms they believe enable this transfer?
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3

human_reviews/8ZQ0HjBOEc.md

@@ -0,0 +1,176 @@
+## Human Reviewer 1
+
+### Summary
+This paper studies the NTK regime of neural networks in the limit of increasing depth $L \to \infty$, i.e., the regime obtained by first taking the infinite-width limit $n \to \infty$ and then the infinite-depth limit $L \to \infty$. The analysis is restricted to the *ordered phase*, where the initialization variance $W_{ij} \sim \mathcal{N}(0, 1/n)$ leads to vanishing gradients as depth increases. In this setting, the NTK $\Theta(X,X)$ approaches a singular matrix as $L \to \infty$. The authors study the behavior of the *mean predictor*, defined as $\Theta(x, X)\Theta^{-1}(X, X)$, which corresponds to the expected kernel regression solution under the NTK. The main result appears to be a proof that, despite the convergence of $\Theta(X,X)$ to a singular matrix in this vanishing-gradient regime, the mean predictor itself admits a well-defined limiting form.
+
+### Strengths
+- **Relevance:** In general, scaling limits of neural networks beyond the standard infinite-width regime are in important and relevant topic in deep learning theory, and the paper aims to advance this line of work.
+
+### Weaknesses
+- **Lack of novelty:**  The presented results have appeared in various forms in prior work. For instance, Xiao et al. (2020) [1] show the existence of a well-defined limit for the mean predictor in Eq. (16). Related results and discussions of the limitations of the "$n \to \infty$ then $L \to \infty$" limit also appear e.g. in Seleznova & Kutyniok (2022) [2] and Hayou et al. (2022) [3]. This regime has long been considered trivial, as the covariance between distinct inputs converges to one, leading to effectivelly constant predictions for any input. Moreover, it seems like the paper uses unnecessarily complex machinery of rough differential equations to prove convergence of the mean predictor, which is a result that can be obtained directly through standard linear-algebraic arguments.
+
+- **Restricted initialization setting:**  The analysis considers only the initialization $W_{ij} \sim \mathcal{N}(0, 1/n)$, corresponding to the *ordered phase*, in which gradients vanish with depth. Prior works cited above studied a broader class of initializations and show that the behavior of the "$n \to \infty$ then $L \to \infty$" regime depends critically on this choice. The present paper does not acknowledge this and, in fact, never explicitly defines the initialization variance.
+
+- **Lack of relevant literature discussion:**  Existing work on the regime $n \to \infty$, $L \to \infty$, $L/n \to 0$ is not discussed. The paper does not reference or contrast its results with any prior analyses of this limit.
+
+- **Misrepresentation of prior results:**  The paper incorrectly summarizes the results of Hanin & Nica (2019) [4], which concern the *double-scaling* regime where $L, n \to \infty$ with $L/n \to \lambda > 0$. The text repeatedly claims that Hanin & Nica study the case where "the ratio of depth to width is unbounded" or where $L \gg n$, which is incorrect.
+
+- **Lack of clarity and technical precision:** There are multiple problems with clarity and technical precision in the text. A few examples are:
+  - The initialization distribution is never specified, so the source of randomness in the concentration results is unclear.  
+  - "Case a" (lines 221–222) states that "if data points lie on a unit sphere, the NTK is invertible." This is trivially false, with a counterexample of taking the same point on a sphere multiple times. Moreover, the text references "Proposition 2 of Jacot et al (2018)" here. However, Jacot et al (2018) [5] does not contain Preposition 2.
+  - "Case c" (lines 227–228) refers to "stereographic projection from $\mathbb{R}^{n_0}$ to $\mathbb{S}^{n_0-1}$." Such a projection does not exist.
+
+## References
+
+[1] Xiao, L., Pennington, J., & Schoenholz, S. (2020). *Disentangling trainability and generalization in deep neural networks.* ICML.  
+[2] Seleznova, M., & Kutyniok, G. (2022). *Analyzing finite neural networks: Can we trust neural tangent kernel theory?* MSML.  
+[3] Hayou, S., Doucet, A., & Rousseau, J. (2022). *The curse of depth in kernel regime.* NeurIPS Workshop.  
+[4] Hanin, B., & Nica, M. (2020). *Finite depth and width corrections to the neural tangent kernel.* ICLR.  
+[5] Jacot, A., Gabriel, F., & Hongler, C. (2018). *Neural tangent kernel: Convergence and generalization in neural networks.* NeurIPS.
+
+### Questions
+- What is meant by “stereographic projection from $\mathbb{R}^{n_0}$ to $\mathbb{S}^{n_0-1}$” mentioned several times in the paper?
+- Why is it necessary to rely on rough differential equations to obtain the main result?
+- How do the presented results differ from or improve upon the mentioned existing work?
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper analyzes the effect of increasing depth on the Neural Tangent Kernel (NTK) for infinitely wide, overparameterized ReLU networks. While prior work established that wide networks behave like kernel methods, the role of depth remains less understood. The authors prove that as depth increases (with depth growing slower than width), the normalized NTK converges to a matrix of ones, and the kernel regression solution approaches a fixed limit for data on the sphere. They characterize this convergence theoretically using rough path theory and identify key properties enabling generalization to other kernels. Experiments show that while some kernel components converge quickly, the full normalized NTK converges extremely slowly.
+
+### Strengths
+1. By characterizing the deterministic limit of the NTK and its corresponding solution as depth goes to infinity (under a specific regime), this paper provides novel theoretical insights into the fundamental properties of deep, overparameterized networks, moving beyond the established infinite-width paradigm.
+
+2. The work demonstrates high technical quality through its use of advanced mathematical tools, such as rough path theory, to prove its central convergence theorem (Theorem 3). This sophisticated approach allows the authors to handle the challenging technical obstacle of the kernel matrix becoming singular in the limit, showcasing a rigorous and robust analytical framework.
+
+3. By identifying a list of key properties that lead to the observed limiting behavior, the authors  provide a valuable template for analyzing other kernels and architectures. Furthermore, their empirical evaluation clarifies the practical implications, notably highlighting the extremely slow convergence rate of the normalized NTK, which is a crucial observation for connecting theory to practice.
+
+### Weaknesses
+1. The empirical validation is minimal, using only synthetic, uniformly distributed data on a sphere for very shallow depths ($L=1$ to $10$). This is insufficient to support the theoretical claim of "extremely slow" convergence, as the chosen depth range is too small to visually demonstrate the asymptotic behavior. To be more compelling, experiments should include benchmarks on real-world datasets and probe much larger depths to empirically quantify the convergence rate and its impact on test performance.
+
+2. While the paper successfully characterizes the existence of a limiting solution, it provides little discussion on what this limit implies for generalization. Does convergence to a fixed limit imply a loss of representation power or a tendency towards simpler functions? A deeper analysis connecting the specific form of the limiting kernel solution to known generalization behaviors would significantly strengthen the significance of the theoretical results.
+
+### Questions
+1. Your theoretical results establish an "extremely slow" convergence of the normalized NTK to 1, yet your empirical results only show depths up to $L=10$. Could you provide a quantitative estimate of the depth $L$ required for the kernel solution $\kappa_x\kappa^{-1}$ to be within a small $\epsilon$ of its limit on, for instance CIFAR-10? This would greatly help in assessing the practical relevance of your limit for modern deep architectures.
+
+2. Your analysis crucially relies on the depth growing slower than the width to maintain a deterministic NTK. Is this a fundamental requirement for your rough path theory technique to hold, or is it an artifact of the current proof? Could you comment on the feasibility and potential challenges of extending your analysis to the stochastic NTK regime described in Hanin & Nica (2020)?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper examines the role of network depth through the lens of Neural Tangent Kernel (NTK) theory. Within this framework, this paper shows that the normalized NTK converges to the all-ones matrix as depth increases for data on the sphere (with extensions via projection for more general data), under the assumption that depth grows more slowly than width. Empirical evidence illustrates that this convergence is very slow.
+
+### Strengths
+The background of NTK theory is clearly presented. Theorems appear to be rigorously argued, and the experiments are conducted to validate them.
+
+### Weaknesses
+Although the paper considers a different width and depth scaling than Hanin & Nica (2020), I do not think novelty of this result meets the bar for acceptance at this venue. The scope is limited to CNNs, omitting Transformers, now widely used in practice. 
+
+The experiments explore depths up to only 10 layers, which is not particularly deep by modern standards.
+
+### Questions
+If I understand correctly, normalized NTK converge to the all-ones matrix and if so, it would become a deterministic kernel independent of the input data. Is this due to the normalization? And does it imply that the neural network just reduces to a kernel method with a trivial kernel?
+
+Could the authors emphasize the key technical difficulties in the proof of Theorem 3 and explain how they overcame them in plain language?
+
+Missing reference: Please include and discuss “Neural Tangent Kernel Analysis of Deep Narrow Neural Networks” (Lee et al., 2022) in the Related Work section.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper provides a theoretical and empirical analysis of how increasing depth affects the Neural Tangent Kernel (NTK) in infinitely wide, fully-connected ReLU networks. The authors show that as depth $L \to \infty$, the normalized limiting NTK $\bar{\Theta}^{(L)}$ converges to a matrix of ones, implying that all inputs become perfectly correlated in the infinite-depth limit. Despite this kernel degeneracy, the closed-form NTK predictor $f(x) = f_0(x) + \Theta_x^{(L)} [\Theta^{(L)}]^{-1} (y - y_0)$ converges to a finite, well-defined limit due to a cancellation effect between the numerator and denominator terms. Theoretical results are supported with toy experiments that visualize convergence rates and verify monotonic correlation growth.  
+Overall, the paper clarifies the limiting behavior of NTKs with depth and its implications for overparameterized models trained in the kernel regime.
+
+### Strengths
+-- Rigorous theoretical grounding, with a clean progression from correlation dynamics to kernel limit to predictor stability.
+
+-- Extends NTK theory by isolating the effect of depth, complementing prior focus on width.
+
+--  Elegant use of normalization and rough differential equations to handle singular kernel limits.
+
+-- Empirical plots and numerical validation reinforce theoretical claims (Fig. 1, lines L432--L446).
+
+-- Results implications: very deep, infinitely wide ReLU nets become less expressive, converging to constant mappings.
+
+### Weaknesses
+-- Experimental section is minimal; only small-scale synthetic tests are shown. A broader empirical sweep would strengthen conclusions.
+
+-- Intuition behind the RDE-based boundedness could be expanded—currently highly technical and somewhat opaque to non-specialists.
+
+-- The convergence rate discussion could quantify how “extremely slow” the approach to 1 is (as noted in L446--L454) using asymptotic bounds.
+
+-- The experimental section and conclusions feel somewhat underdeveloped, leaving the reader wishing for more explicit explanations or additional insights into the implications and experimental findings as per ICLR standards.
+
+### Questions
+-- Can the authors clarify how the convergence rate of $\phi^{(L)} \to 1$ depends on activation nonlinearity—e.g., would smoother activations yield slower or faster kernel collapse?
+
+-- Since the NTK converges to a constant kernel, does this imply that in the deep infinite-width limit, gradient descent loses any data-dependent inductive bias?
+
+-- Minor comments: 1) L051, L052: Proposition 4 and, Theorem 3 do not have hyperlink (which makes it easy to naviagte) and also does not give much insight as what the actually contributions are by just reading the contributing section, can be written better. 2) same issue with L218, no hyperlinks, poorly written ==>
+This, in turn, allow us to aim at characterizing the output of such neural network,as done in the rest of the section. Indeed, from Proposition 2 and Proposition 2 from Jacotet al. (2018), we can immediately observe a few facts regarding the input data:
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/9h2c2rMSY1.md

@@ -0,0 +1,130 @@
+## Human Reviewer 1
+
+### Summary
+Spatiotemporal PDEs are being increasingly solved with neural operators. Uncertainty quantification on such predictions, therefore, is not immediately possible as a result of distribution shifts that occur as temporal windows slide. In this vein, previous works have studied conformal prediction under covariate shift. This paper then studied the application of such ideas to this setting, first in the full functional form and then in the discretized setting.
+
+### Strengths
+The paper presents an interesting problem of study: the study of spatiotemporal PDE coverage seems like a worthwhile direction. The exposition of the paper is also very clear, making the paper easy to follow and engage with. The paper also nicely discusses both the functional and the discretized forms of the problem statement.
+
+### Weaknesses
+In the first part of the paper, the authors establish that the TV distance between consecutive steps is maximal in the functional space. I believe the implication of this statement is supposed to be that, therefore, we cannot consider the functional space coverage using ideas from covariate shift. However, the approach of Barbet et. al only establishes that $\mathcal{P}(Y\in\mathcal{C}(X))\ge 1-\alpha-\sum_i w_i d_{TV}(z, z^i)$, which is a lower bound. How can we say anything more strongly than just making claims of this lower bound here?
+
+Moreover, this then motivates studying this problem in a finite-dimensional discretized form. This, however, then just becomes a standard, finite-dimensional time series problem, so I do not see how this becomes any different from a typical finite-dimensional conformal time series problem. Also, Theorem 4.2 appears to be a standard proof from SDEs, so I do not see why it and its proof are provided in the main body of the paper.
+
+Finally, the experimental results appear to be not very compelling: many times, WCP appears to produce extremely conservative prediction regions, having a coverage of 1.0 or close to it at desired coverage levels well below this. The other methods, in contrast, appropriately hit 0.90 initially.
+
+### Questions
+1. How does this particular setting differ from a general time series conformal prediction problem after the PDE has been discretized? Are there particular structures that you can uniquely leverage to develop a method in this setting that cannot be more generally applied? This reduced scope would likely yield more useful methods.
+
+2. Why are the coverages so conservative in this setting? The empirical results seem to indicate over-coverage even before the time steps progress very far.
+
+3. Even though the TV in the functional case is 1.0, does that necessarily mean there is no other form of covariate-shift type adjustments that could fix the miscoverage?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper develops weighted conformal prediction (WCP)  for applying conformal prediction to time-dependent PDE surrogate models.  The authors prove that in function spaces with time-dependency, distributions at different times can be mutually singular (TV distance = 1), breaking the exchangeability assumption and hence CP coverage guarantees. As a solution, they propose reweighing the nonconformity scores with density ratios to enable exact coverage guarantees. Experiments on three PDEs (univariate, linear, with Gaussian initial conditions) demonstrate that WCP maintains coverage where two baselines, naive CP and LSCI, fail.
+
+### Strengths
+- The problem is significant and well motivated. The authors make a good point in laying out the challenge of covariance shifts in time-dependent PDEs that are common in physics and engineering. Theorem 4.1 was also a clear illustration of this challenge and how existing methods can fail.
+
+- The paper was over-all well organized and easy to follow.
+
+### Weaknesses
+- No real algorithmic or theoretical contribution. Theorem 4.1 proves that for the heat equation with Gaussian initial conditions, the TV distance between solution distributions at any two distinct time points is maximal (equals 1). Theorem 4.2 shows that for discretized linear PDEs with Gaussian initial conditions, the solution distribution at any time t remains Gaussian, with explicitly computable mean and covariance. These two results are not particular novel for PDEs. The authors' method, then, directly applies Barber et al. 2023's weighted conformal prediction algorithm using the likelihoods. The coverage guarantees are from Barber 2023. I fail to see how this algorithm is a novel contribution. 
+
+- The scope is limited. The algorithm only applies to _linear_ PDEs with _Gaussian_ (and location-scale) initial conditions, but most physical systems of interest of UQ have nonlinear dynamics (Navier-Stokes, reaction-diffusion with nonlinear terms, etc.). The proposed algorithm require explicit probabilities which makes it not extendable to the more general case. 
+
+- Significant weakness and ambiguity in the experiment section. This includes:
+    - What is the target coverage? Why is the target in Fig 3 set at 90% but the target for WCP said to be 99%? Why does WCP's line disappear half way through the horizon? What is $n_\infty$ in Table 1 and how should we interpret it?
+    - Limited experiment setups: only linear, univariate, synthetic experiments. This is a result of the nature and assumption made by the method, but with this set of experiments it's not very convincing that this algorithm is useful in practice.  
+    - Lack of comparison to baselines, and discussion of literature. From the PDE literature, there are works that provide probabilistic forecasts (for example [1,2], and the baselines in LSCI). From the CP literature, there are a swath of works that handles time series distribution shift that does not require explicit covariates or local exchangeability [3-6]. They should be compared to show what are the advantages of your algorithm, or at least discussed w.r.t. why they are not applicable to your setup. 
+    - It is also questionable that WCP over-covers by so much, often achieving 1.0 coverage with infinite bands. While I respect the authors' honest reporting, this is a significant limitation on the usability of the UQ bands. UQ's goal is calibration and the tendency to over-cover is not desirable. Can you show results of both 90\% and 99\% target coverage?
+    - Some qualitative plots of UQ bands (similar to figure 1) will also help the presentation of results.  
+
+[1] Yang, Liu, Xuhui Meng, and George Em Karniadakis. "B-PINNs: Bayesian physics-informed neural networks for forward and inverse PDE problems with noisy data." Journal of Computational Physics 425 (2021): 109913.
+
+[2] Bülte, Christopher, Philipp Scholl, and Gitta Kutyniok. "Probabilistic neural operators for functional uncertainty quantification." arXiv preprint arXiv:2502.12902 (2025).
+
+[3] Gibbs, Isaac, and Emmanuel Candes. "Adaptive conformal inference under distribution shift." Advances in Neural Information Processing Systems 34 (2021): 1660-1672
+
+[4] Angelopoulos, Anastasios, Emmanuel Candes, and Ryan J. Tibshirani. "Conformal pid control for time series prediction." Advances in neural information processing systems 36 (2023): 23047-23074.
+
+[5] Xu, Chen, and Yao Xie. "Sequential predictive conformal inference for time series." International Conference on Machine Learning. PMLR, 2023.
+
+[6] Auer, Andreas, et al. "Conformal prediction for time series with modern hopfield networks." Advances in neural information processing systems 36 (2023): 56027-56074.
+
+### Questions
+See Weaknesses.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper studies how conformal prediction (CP) fails under temporal non-stationarity common in time-dependent PDEs, where calibration and test data are not exchangeable. The authors first prove that, in function-space settings, even simple PDEs (e.g., the 1-D heat equation) produce mutually singular solution distributions across time, making exact coverage guarantees impossible. They then propose Weighted Conformal Prediction (WCP) for discretised PDE surrogates, deriving closed-form Gaussian densities for the evolving solutions and using likelihood-ratio weights to re-establish exact coverage. Experiments on fractional diffusion, backward heat, and reaction–diffusion equations show that WCP maintains nominal 90 % coverage over long horizons, whereas naïve CP and local-exchangeability (LSCI) methods quickly under-cover.
+
+### Strengths
+- Mathematical clarity. The analysis of mutual singularity in function spaces (Theorem 4.1) exposes a genuine limitation of applying CP directly to PDEs.
+
+- Principled solution. The likelihood-weighted approach is a suitable approach to time-dependent distribution shift, avoiding unverifiable “local exchangeability” assumptions.
+
+- Clear writing and sound experiments. The presentation is precise, proofs are rigorous, and empirical results clearly support the claims.
+
+### Weaknesses
+- Limited scope. Theoretical guarantees hold only for linear PDEs with Gaussian initial conditions, where analytic densities exist.
+
+- Incremental novelty. Weighted CP under covariate shift is known (e.g., Barber et al., 2023); the contribution is mainly its application to PDE surrogates, not a new CP framework.
+
+- Narrow experiments. Tests on 1-D synthetic PDEs demonstrate correctness but not scalability or real-world complexity.
+
+### Questions
+- How sensitive is WCP to deviations from Gaussianity or linearity?
+- Can the framework be extended to nonlinear or stochastic PDEs without closed-form densities?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/A4Us8jxVGq.md

@@ -0,0 +1,196 @@
+## Human Reviewer 1
+
+### Summary
+This paper study how transformers learn semantic associations during training on natural language data. The authors develop a theory based on training dynamics, specifically using a leading-term approximation of the gradients to analyze the model's weights in the early phase of training. Their central finding is that the weight matrices, like the output $W_O$, the value $V^{(\ell)}$, and query-key $W^{(\ell)}$ can be expressed as compositions of three basis functions derived directly from corpus statistics. These basis functions are bigram maping, that captures the next token co-occurrence ($\bar{B}$); the interchangeability maping, that fnds tokens with similar preceding token distributions (like synonyms or words in the same grammatical class) ($\sum_{\bar{B}}$); and context mapping, that encodes long range co-occurrance between a token and the tokens in its prefix ($\bar{\Phi}$).
+The authors validate this theory first on a 3 layer attention only model trained on TinyStories, and they show that the learned weights have a very high cosine similarity (often >0.99) to their theoretical characterizations. Then, they extend the analysis to the Pythia-1.4B checkpoints, showing that these foundational associations are also learned in larger models, especially during early training and in the later layers.
+
+### Strengths
+- The leading-term gradient approximation is a clever technical innovation that makes an otherwise intractable analysis feasible while maintaining interpretability. The decomposition into three basis functions provides clear mechanistic insight into how different components capture semantic structure.
+- Unlike much prior theoretical work that relies on synthetic data or heavily simplified models, this analysis is grounded in a more realistic setup. 
+- The paper provides closed-form expressions for all major weight matrices uniformly across layers, showing how they compose the basis functions differently.
+
+### Weaknesses
+- The theory applies to attention-only transformers, but validation on Pythia requires indirect analysis through covariance matrices of activations rather than direct weight comparison. This introduces an additional layer of approximation and makes it unclear how MLPs and multi-head attention affect the theoretical predictions.
+- The theoretical guarantees only hold for $\mathcal{O}(1/\eta)$ steps in early training. While empirical results show features persist longer, the paper lacks analysis of how and why these features evolve during later training, limiting practical applicability for understanding fully trained models.
+- To achieve theoretical tractability, the analysis relies on certain assumptions, such as "sufficiently small" initialization and bounds on the learning rate and number of steps (as seen in Theorem 4.1). The error bounds for the approximation, while derived, grow over time. This means the exact closed-form expressions are an idealization, and their accuracy naturally decreases as training progresses and the model's state becomes more complex.
+- Experiments use truncated vocabularies (3K for TinyStories, 20K for Pythia) which may not capture the full complexity of semantic associations in real LLMs with vocabularies of 50K+ tokens. The computational tractability argument doesn't fully justify this limitation.
+
+### Questions
+Your leading-term approximation brilliantly explains how associations emerge in the very early stages of training. The empirical results with Pythia suggest these structures persist for some time but also evolve, especially in the earlier layers. What are your thoughts on the mechanisms that drive this evolution away from the initial basis functions? Is it simply the accumulation of higher-order gradient terms, or do you suspect qualitatively new structures, like logical reasoning circuits, are being built on top of this initial associative foundation?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper presents a theoretical and empirical analysis of how semantic associations emerge in transformer-based language models during early training. By analyzing the leading-term approximation of training gradients, the authors derive closed-form expressions for the model's weight matrices. They show that these weights can be characterized by the composition of three interpretable basis functions—bigram, interchangeability, and context mappings—which collectively explain how transformers acquire semantic structure from data. The theoretical claims are rigorously validated through experiments on both the controlled TinyStories benchmark and large-scale models like Pythia-1.4B, demonstrating strong agreement between the theoretically predicted features and the empirically learned weights.
+
+### Strengths
+1. The paper starts from the perspective of gradient flow and utilizes the contextual distribution in the text to explain the implied semantic associations within the parameters. This perspective is very novel and provides an in-depth way of understanding the distributional characteristics of the parameters of transformer-based language models.
+2. The paper is trained on real text, and the experimental results are highly consistent with the theoretical analysis, demonstrating the reasonableness of its results.
+
+### Weaknesses
+1. The paper is validated only on the TinyStories dataset, which I consider insufficient for comprehensive verification.
+2. The training loss on TinyStories remains very high (greater than 5), indicating that the model's learning on this dataset is inadequate and arguably unsuccessful. The value of discussing parameter characteristics under such inadequate fitting conditions is highly questionable.
+3. The description of how theoretical and experimental results are compared in Section 5.2 is difficult to understand. The authors should revise this section and provide more detailed methodological descriptions in the appendix, preferably in an algorithmic format.
+4. The paper uses one-hot encoding instead of the more commonly used embedding encoding. The authors should discuss how this distinction affects their theoretical results, particularly whether the presence of an embedding layer under zero initialization would invalidate their theoretical findings.
+5. Paper [1] also discusses from a gradient flow perspective how the embedding layer is influenced by semantic distributions under small initialization, and its findings share some similarities with certain results in this paper. I believe the authors should discuss the distinctions between their work and [1].
+
+[1] An Analysis for Reasoning Bias of Language Models with Small Initialization, Forty-second International Conference on Machine Learning.
+
+### Questions
+See weakness.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper studies how semantic associations first emerge in attention-based language models by analyzing early training dynamics. Using a leading‑term expansion of the gradients, the authors derive closed‑form approximations for all major weight classes in a self attention only transformer with causal masking and learned relative positional bias. 
+shows very high cosine similarity between learned weights and their leading‑term predictions, and an analysis of Pythia‑1.4B on OpenWebText shows strong early training alignment that gradually drifts laterr. Qualitative examples illustrate that the basis functions recover intuitive relations like "fish" <--> "pond/lake".
+
+### Strengths
+* The three basis functions (bigram, interchangeability provide an intuitive, corpus‑linked explanation for what each weight class is learning and how attention and values cooperate early in training
+
+
+* The writing is very clear
+
+* Near perfect cosine agreement on TinyStories over many steps and reasonable agreement in a non-toy sized model (pythia) support external validity. heatmaps/diagrams are clear
+
+### Weaknesses
+* Quantitative validation centers on cosine similarity and selected qualitative token lists. Some broader behavioral evaluations or ablations (e.g., other corpora, tokenizations, or stronger baselines) would further strengthen the claims
+
+* Unless I'm missing something, this doesn't hold well for later stages of training because of drift. That's fine, but I figured I'd raise it. I am not entirely sure how much this is a problem for the scope of the paper. Are there analyses on later stages of training that would be interesting but are blocked by this constraint?
+
+* I would like to see some interventional results showing some practicality: E.g., show that the leading term predictors enable training diagnostics or interventions (e.g., early phase monitoring that forecasts later perplexity or feature formation).
+
+* The theoretical result is clean and realistic for early training, and the empirical results are suggestive. I think the thing holding this paper back right now are scope
+
+### Questions
+N/A
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper studies how semantic associations between tokens emerge during the early training of language models. Focusing on a self-attention-only architecture, the authors derive the leading-order gradient updates and show that the learned weights can be decomposed using three quantities that reflect the statistics of the training corpus: bigram mapping (correlations between consecutive tokens), intercheangiability mapping (correlations induced by similarity of tokens’ previous-token context distributions), and context mapping (average contextual features of tokens across their occurrences). They validate the theory on small attention-only models - where the predicted leading terms match the learned parameters in the early steps - and then analyze embeddings in Pythia models, finding qualitative alignment with the theory in the initial training phase.
+
+### Strengths
+- The paper presents an interesting and, to my knowledge, novel blend of mathematical analysis, numerical experiments, linguistic insight, and mechanistic interpretability. The bottom-up approach of deriving representations from first principles of training dynamics is significant.
+- The paper analytically derives a leading-order approximation of the weight evolution and provides clear, intuitive interpretations for each component.
+- The theory does not rely on synthetic data models. Rather, the basis functions are derived directly from the statistics of the language corpora.
+- The paper is well-written and its assumptions are clearly outlined.
+
+### Weaknesses
+- The primary weakness is that the theory still relies on strong architectural assumptions. It is derived only for self-attention-only transformers. This analysis does not account for the impact of MLP layers, which are important components of present models. While this is a limitation, it is an understandable simplification necessary to make the analysis tractable, and future work might build upon this to relax these assumptions.
+
+### Questions
+- Th. 4.1 provides formal bounds on the error between the learned weights and their leading-term approximations. How tight are these error bounds in practice? For instance, in the TinyStories experiment.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+This paper addresses a critical gap in transformer interpretability: how semantic associations (e.g., "bird"-"flew", "country"-"capital") emerge during training of attention-based language models (LLMs) on natural language data. Unlike prior work that relies on synthetic data, simplified architectures, or non-standard training, the authors ground their analysis in realistic settings (natural text distributions, standard transformers with positional encoding, and standard next-token prediction loss). Their key technical innovation is leveraging a gradient leading-term approximation to derive closed-form expressions for transformer weights in the early training stage.
+
+The authors show that all transformer weights (output, value, query-key, positional encoding) can be characterized as compositions of three interpretable basis functions: (1) bigram mapping (captures next-token dependencies), (2) token-interchangeability mapping (reflects functional similarity, e.g., synonyms), and (3) context mapping (encodes long-range prefix-suffix co-occurrence). They validate this theory empirically: on a 3-layer transformer trained on TinyStories, learned weights maintain cosine similarity ≥0.9 with theoretical predictions even beyond early training; on real-world LLMs (Pythia-1.4B trained on OpenWebText), early-stage activations and attention weights strongly align with the proposed basis functions.
+
+### Strengths
+1. Realism of the theoretical setup: By retaining critical components of practical transformers (positional encoding, causal masking, residual streams) and using natural text, the paper avoids the over-simplifications that limit the generalizability of prior work.
+
+2. Mechanistic interpretability: The basis functions provide a causal explanation for semantic association (e.g., the value matrix combines context and bigram mappings to encode long-range semantics), rather than just correlational observations.
+
+3. Strong empirical validation: The experiments are comprehensive, testing the theory on both small controlled models (TinyStories) and large real-world LLMs (Pythia-1.4B), with quantitative (cosine similarity) and qualitative (token correlation examples, Figure 5) evidence.
+
+4. Foundational value for future work: The closed-form expressions and basis functions can serve as a starting point for further research (e.g., diagnosing bias in LLMs by analyzing deviations from the theoretical bigram/context mappings, or designing more interpretable architectures).
+
+### Weaknesses
+1. Limited analysis of later training stages: While the paper notes that theoretical features persist beyond early training (e.g., cosine similarity ≥0.7 after 100 epochs in TinyStories), it does not explore why or how weights drift from the leading terms. Understanding this drift (e.g., whether it corresponds to higher-level semantic learning) would strengthen the theory’s completeness.
+
+2. Multi-head attention and MLP layers are understudied: The Pythia experiment adapts the analysis to multi-head attention by averaging attention heads, but it does not explore how individual heads or MLP layers interact with the proposed basis functions. For example, do some heads specialize in bigram mappings while others focus on context?
+
+3. Lack of causal interventions: The empirical validation relies on correlational measures (cosine similarity, covariance), not causal tests (e.g., ablating the bigram mapping component of weights to see if next-token prediction degrades). Causal interventions would more strongly confirm that the basis functions are necessary for semantic associations.
+
+### Questions
+1. Drift in later training stages: You note that weights drift from leading terms as training progresses (Figure 6). Can you characterize this drift? For example, does it correspond to learning higher-level semantics (e.g., syntax, pragmatics) that build on the initial basis functions, or does it reflect overfitting to idiosyncrasies in the data?
+
+2. Multi-head attention specialization: In Pythia, you average attention heads to compute token correlations, but prior work shows heads specialize in different tasks. Do individual heads align with specific basis functions (e.g., some heads focus on bigram mappings, others on context)? If so, how does this specialization emerge?
+
+3. Causal validation: Your analysis uses correlational measures (cosine similarity) to link learned weights to theoretical terms. Have you tried causal interventions (e.g., modifying the bigram mapping component of the output matrix and measuring changes in next-token prediction accuracy for bigram-dependent pairs like "bird"-"flew")? Such tests would strengthen the claim that the basis functions are functional, not just correlational.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+4

human_reviews/AB3jM3iMaW.md

@@ -0,0 +1,166 @@
+## Human Reviewer 1
+
+### Summary
+The paper proposes an approach for future link prediction in temporal graphs (called link forecasting). It finetunes an LLM (Qwen3-4B) for this task via reinforcement learning. All textual data is removed from the graphs and replaced by IDs.
+
+### Strengths
+- Original idea of using LLMs for link forecasting
+- Strong results compared both to other LLMs and compared to TGNNs
+- Explanation traces are provided and evaluated in a user study
+- Paper is well-structured
+
+### Weaknesses
+- The motivation for using an LLM remains unclear (apart from some success in preliminary work)
+
+### Questions
+- What is the motivation of using an LLM if all textual data is removed?
+- Do you have a hypothesis why your approach works so well despite removing all textual data?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper constructs temporal graphs in natural language and finetunes LLMs with reinforcement learning on the dataset. The finetuned model outperforms larger LLMs under the proposed pMRR and LLM-as-a-Judge metrics.
+
+### Strengths
+The paper presents that the reasoning ability of LLMs on temporal graphs can be enhanced by RL training, which is not explored by previous works of LLM on temporal graphs, which mainly focus on evaluation and ICL. The full pipeline, from data construction to training and evaluation, is reasonable and easy to follow.
+
+### Weaknesses
+The main weakness is a lack of innovation and further insights. The idea that the graph reasoning ability of LLMs can be incentivized through reinforcement learning has been illustrated in previous work. Although the authors claim that previous works focus on static graphs while they target temporal graphs, they don't show what difference it will bring to the reasoning or further insights (e.g., what is special for temporal graphs but not static graphs). In the whole pipeline, the innovation is quite limited. For example, the algorithm is the well-established GRPO, and the LLM-as-Judge evaluation is also commonly used.
+
+### Questions
+1. The base models only include Qwen3-4B and Qwen3-0.6B. What's the performance of the model finetuned from a larger Qwen3-8B?
+
+2. What's the performance of the ReaL-TG models on OOD benchmarks, e.g., unseen temporal graph benchmarks?
+
+3. Does the training harm or enhance the general reasoning ability of base models, e.g., the math benchmarks?
+
+4. How does the RL training increase the reasoning of LLMs on temporal graphs? Can you show some intuitive examples?
+
+5. The reason for proposing pMRR is to evaluate the over-generation of LLMs. Can the recall in F1 play a similar role?
+
+6. Why don't you use the proposed metrics as rewards during RL training?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper introduces Real-TG, which employs reinforcement learning with verified rewards (LRVR) like GRPO to fine-tune an LLM for more effective explainable link forecasting on anonymized temporal graphs. It formulates the link forecasting as the QA tasks and designs a specific prompt to instruct LLMs to identify the destination node. The authors also propose a metric (i.e., pMRR) to evaluate the prediction capability of LLMs (as LLMs can provide discouraging over-generation). Extensive experiments show that Real-TG-4B achieves good results across different vanilla LLMs and traditional link forecasting methods.
+
+### Strengths
+* The idea of introducing LRVR to LLM-based temporal graph models is interesting.
+* The proposed fine-tuning paradigm, Real-TG, is simple and effective.
+* The paper is well-written and easy to follow.
+
+### Weaknesses
+* The novelty of the proposed Real-TG is limited and lacks technical innovation. In Real-TG, the QA formulation [1], the GRPO pipeline [2], the $\alpha$-temporal random walk [3], and the LLM-as-a-Judge evaluation [4] are introduced from existing works. All these mechanisms resemble existing works and do not introduce principled advances in LLM-based temporal graph learning or RLVR.
+* The motivation of the model component design is unclear. 
+	* Why is RL chosen instead of SFT? In the literature on large reasoning models, RL is typically used to optimize preferences or human-aligned objectives; however, for tasks with a single correct answer, SFT may be more appropriate. The authors should clarify why RL is necessary and what concrete benefits it confers in this setting.
+	* Why adopt $\alpha$-temporal random walks rather than simpler k-hop neighborhoods?
+	* Why use F1 as the reward metric rather than other ranking/IR metrics (e.g., MRR)? In equation 1, do the authors compute this F1 over a single pair of samples, since F1 fundamentally requires aggregation over multiple samples? An illustrative example may help with a better understanding of the paper.
+* The GRPO procedure is known to be sensitive to hyperparameters. Does Real-TG also suffer from such instability?
+* Plots showing reward vs. training step would greatly strengthen confidence in the RL training procedure.
+* Why not conduct experiments in the recently proposed benchmark [5] and compare the models that also utilize LLMs for temporal graph learning [6]?
+
+[1] Are Large Language Models Good Temporal Graph Learners?
+
+[2] DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models
+
+[3] Zebra: When temporal graph neural networks meet temporal personalized pagerank
+
+[4] Judging Llm-as-a-judge with Mt-bench and Chatbot Arena
+
+[5] DTGB: A Comprehensive Benchmark for Dynamic Text-Attributed Graphs
+
+[6] Unifying Text Semantics and Graph Structures for Temporal Text-attributed Graphs with Large Language Models
+
+### Questions
+No
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper tackles link forecasting on temporal graphs, emphasizing explainable predictions using large language models that generalize to unseen graphs without retraining. It introduces ReaL-TG, a reinforcement learning framework that fine-tunes LLMs with Grouped Regularized Policy Optimization and an outcome-based F1 reward to encourage self-exploration of reasoning strategies from graph structures. Core contributions include the Temporal Context Graph Selection algorithm for relevant subgraph extraction, a QA formulation for the task, and a new evaluation protocol featuring penalized mean reciprocal rank alongside an LLM-as-a-Judge system assessing faithfulness, logical consistency, and answer-explanation alignment. Main results demonstrate that the fine-tuned ReaL-TG-4B model outperforms larger frontier LLMs and traditional temporal graph neural networks on prediction metrics and reasoning quality across seen and unseen datasets from the Temporal Graph Benchmark.
+
+### Strengths
+- The paper introduces ReaL-TG, a complete RL-based framework designed to fine-tune LLMs for explainable link forecasting on temporal graphs. This is a novel approach to a task where prior methods lacked explainability.
+- The experimental results are impressive. The fine-tuned 4B model (ReaL-TG-4B) significantly outperforms its base model (Qwen3-4B) and, notably, much larger frontier LLMs like Llama3.3-70B and GPT-5 mini on overall prediction accuracy.
+- The framework demonstrates excellent generalization. ReaL-TG-4B achieves the highest MRR and PMRR on the two "unseen" TGB datasets, outperforming all baselines.
+- The paper successfully demonstrates that ReaL-TG not only improves prediction accuracy but also reasoning quality.
+
+### Weaknesses
+- The paper provides an insightful observation of "reward hacking" in the smaller ReaL-TG-0.6B model, which "justifies its predictions by claiming '(uq, vq, tq) has already been seen'". This is a critical finding for outcome-based RL methods. Despite its importance, this phenomenon is not systematically measured. The paper provides no evidence or quantification to demonstrate it.
+- The approach injects graph context purely as text; while explainable, it may drop important structural cues vs. learned embeddings.
+- Relies solely on six TGB datasets, which may share similar characteristics like social or transaction networks, limiting generalization claims.
+- Training data sampled from only four datasets with 1,000 queries, potentially insufficient for diverse temporal patterns.
+- Unseen datasets (uci, enron) are smaller, with fewer involved nodes, possibly underrepresenting complex scenarios.
+
+### Questions
+- Your definition in Section 4 states PMRR penalizes over-generation, resulting in a lower score than MRR. However, in Table 2, several large baseline models (e.g., Llama3.3-70B, GPT-5 mini) show a PMRR score that is higher than their MRR score. Could you please confirm if the description of PMRR in Section 4 is correct, or explain the mechanism by which over-generation leads to a higher PMRR score.
+- Could you please briefly elaborate on the path that makes $(e_2, t_2)$ a 3-hop neighbor in the context of Figure 2? 
+- Can you clarify how the termination probability $\alpha$ and the decay factor $\beta$ interact in the random walk?
+- Why was the F1 score chosen as the reward function? Did F1 prove uniquely robust to this hacking behavior?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/BDNctVKwuD.md

@@ -0,0 +1,140 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes a interpretability framework to improve feature alignment in CNNs. The authors theoretically demonstrate that HiResCAM explanations are not unique because of the softmax activation, and adding a shift via a common matrix 
+ can distort attention maps. To solve this, they propose ContrastiveCAMs, which remove this redundancy by producing shift-invariant, attention maps that more accurately highlight discriminative image regions. Experiments show that models depend on non-core, irrelevant regions, specially when regions occupy small portions of images.They then develop Core-Focused Cross-Entropy, a modified loss function that suppresses attention to non-core areas while also making target features more relevant. The method performed good across multiclass and binary classification tasks and some segmentation benchmarks.
+
+### Strengths
+- The work is theoretically grounded. They theoretically show that HighResCAM explanations are not uniquely determined because of arbitrary shifts via softmax invariance, and provide formal proofs. 
+- The paper proposes a solution by connecting interpretability output with training objective to improve feature alignment. Such as actionable interpretability technique is impressive.
+
+### Weaknesses
+- This paper investigates CNNs, not ViTs. I think the whole field has shifted to ViTs long time ago. I am not saying CNNs are obsolete, but having ViT variants (along with CNN variants) are a must now. What about CLIP models? 
+- The paper investigates one (out of many) variation of CAM (here, HighResCAM ), and HighResCAM is not even a published work. I am not sure whether this problem is worth investigating in the first place. What about other CAM methods? What about other explanation methods (saliency-based)?
+- The finding is not novel, in essence similar works have observed the same thing, see [R1]. This work is not even cited. 
+- There are many ways of explanation-guided learning. See [R2]. The authors did not consider these works in their comparisons. Is their method better to [R1, R2 methods]? Does it work on other explainability methods? 
+- There are no results on ImageNet. The authors report on Hard-ImageNet but not ImageNet. Does the method improve performance on the ImageNet? 
+- The introduction lacks motivation. It lists some works and their importances, and then jumps directly to "In this work, we develop...".
+- The related work section provides some similar works but the most important part of that section is missing; the authors should clearly state the difference between the related works and their own work, how they tackle the problems and the limitations of those works, and why their method is better.
+
+
+[R1] Consistent Explanations by Contrastive Learning
+[R2] Studying How to Efficiently and Effectively Guide Models with Explanations
+
+### Questions
+I think this paper is not ready for ICLR due to the weaknesses mentioned. In particular, it investigates only CNNs, and a very special case (HighResCAM) of a very special case (CAM) of the wide variety of explainability methods. Furthermore, the paper lacks motivation, it has problems in how the related work is written, and there are many comparisons with key works missing. My decision will therefore be a reject.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper investigates the reliability of HiResCAM, a commonly used deep learning interpretability method that generates attention maps highlighting image regions important for a model’s predictions. The authors theoretically demonstrate  that HiResCAMs are not uniquely determined and admit arbitrary, spurious
+
+To overcome this, the authors propose ContrastiveCAM, a new visualization technique that is invariant to the spurious shift and provides class-versus-class contrastive explanations, offering more precise insight into what differentiates one class from another.
+
+Using these improved explanations, they discover that networks frequently attend to irrelevant image regions. To correct this, they introduce Core-Focused Cross-Entropy, a modified loss function that encourages attention on core (label-relevant) image regions and suppresses attention elsewhere, thereby improving feature alignment between visual regions and class semantics.
+
+Experiments on Hard-ImageNet and Oxford-IIIT Pets show that ContrastiveCAM produces more faithful and robust attention maps, and that Core-Focused Cross-Entropy leads to better predictive performance derived from semantically meaningful regions.
+
+### Strengths
+The authors identify two main issues in existing methods. From a theoretical perspective, they reveal a limitation of HiResCAM, showing that its attention maps are not uniquely determined. To overcome this, they propose ContrastiveCAM, which eliminates this ambiguity. They also observe that networks often focus on irrelevant image regions, and to address this, they introduce Core-Focused Cross-Entropy, a loss function that encourages attention on label-relevant areas.
+
+Experiments conducted on three different datasets demonstrate the effectiveness of their approach, particularly the Core-Focused Cross-Entropy loss, in improving interpretability and performance.
+
+### Weaknesses
+The authors discuss the limitation of HiResCAM only from a theoretical perspective, without providing experimental evidence to validate their claim.
+
+Furthermore, they do not convincingly demonstrate that ContrastiveCAM outperforms other modern CAM-based methods. Their comparisons are limited to GradCAM, which is relatively outdated, while many recent and more advanced CAM variants could have been included for a more comprehensive evaluation. In addition, the paper lacks experiments directly comparing ContrastiveCAM and HiResCAM, making it difficult to assess the claimed improvements.
+
+Finally, although the authors claim to have improved feature alignment in convolutional networks, their experiments are conducted only on ResNet. Evaluating their approach on a broader range of convolutional backbones would provide stronger and more generalizable evidence for their conclusions.
+
+### Questions
+Number of examples used per dataset for quantitative results is unclear — needs clarification.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This work introduces an improvement over HiResCAM methodology through contrastive alignement. Experiments are conducted on Hard ImageNet. and Pets and Pascal VOC. It is mostly compated to GradCAM and HiResCAM. ResNet50 backbone is only used for validation.
+
+### Strengths
+The description of the work is clear. And the attribution-based explanations are important research field.
+
+### Weaknesses
+This work has a lot of issues that I am worried about:
+
+- first claim is that HiResCAM are popular. Unfortunately I could not find that even this method was published. Only an arxiv version from 2021. And even so, they are cited by 208 works on Google Scholar, and when we will compare it to other CAM methods such as GradCAM (over 30k) and GradCAM++ (over 4k). 
+
+- Lack of baselines, GradCAM is a really outdated method, and recent attribution method are LeGrad [1], OMENN [2] or CheferLRP [3].
+
+- Lack of other backbones in experimentations, especially ViTs. Other work do have them. 
+
+- Lack of contextualization, broader discussion about LRPs, other CAMs and other attribution-based methods such as B-Cos [4] are missing. 
+
+- Visualizations of explanations are not convincing, and there is no experimentation to prove it is better. 
+
+- There is no user study to showcase that users better perceive those explanations. 
+
+- No XAI benchmarks such as FunnyBirds [5]
+
+[1] Bousselham, Walid, et al. "Legrad: An explainability method for vision transformers via feature formation sensitivity." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2025.
+
+[2] WrĂłbel, Adam, MikoĹ Janusz, and Dawid Rymarczyk. "OMENN: One Matrix to Explain Neural Networks." arXiv preprint arXiv:2412.02399 (2024).
+
+[3] Chefer, Hila, Shir Gur, and Lior Wolf. "Transformer interpretability beyond attention visualization." Proceedings of the IEEE/CVF conference on computer vision and pattern recognition. 2021.
+
+[4] Böhle, Moritz, Mario Fritz, and Bernt Schiele. "B-cos networks: Alignment is all we need for interpretability." Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 2022.
+
+[5] Hesse, Robin, Simone Schaub-Meyer, and Stefan Roth. "Funnybirds: A synthetic vision dataset for a part-based analysis of explainable ai methods." Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023.
+
+### Questions
+Seeing missing baselines, general not backed up by the literature statements and poor comparisons, I do not got into much methodological details as I do not that see this work can improved during rebuttal period enough to be ready for publishing. 
+
+I put more details in the weaknesses section showcasing the limitations of the work, especially contextualization, validation of the method and strong statements.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5

human_reviews/BzqkbJTsbq.md

@@ -0,0 +1,166 @@
+## Human Reviewer 1
+
+### Summary
+This paper redefines style transfer as a weak-label guidance problem and image super-resolution / deblurring as degraded-label guidance. It argues that these existing methods are typically tailored to specific tasks, lack generalization and transferability, and often rely solely on loss-based optimization while neglecting underlying domain priors. To address this limitation, the authors propose a unified framework that leverages data knowledge inherent to each task and process knowledge extracted from the reverse diffusion process to solve image restoration and style transfer tasks jointly. They claim that unifying these two tasks is inherently challenging because style transfer depends on abstract style information, whereas restoration tasks use strong, explicit conditioning. Furthermore, they critique loss-gradient-based guidance methods, emphasizing that loss values may not accurately reflect image fidelity and are vulnerable to error propagation. Their method interpolates between the noised label prediction and the generated image from the reverse diffusion process at each step, while simultaneously enforcing label information through a loss computed between the predicted $x_0$ and the target $y$
+
+### Strengths
+- This paper addresses an important and timely problem: unifying weak-label (style transfer) and degraded-label (image restoration) tasks under a single diffusion-guided framework.
+- The proposed approach, integrating data knowledge and process knowledge, is conceptually innovative and could provide new insights for handling tasks with imperfect or indirect supervision.
+
+### Weaknesses
+- The motivation for unifying two inherently distinct tasks, style transfer and image restoration, is not sufficiently substantiated. The authors (L99-L111) suggest potential knowledge transfer between the two, but no empirical evidence or references are provided. In fact, these tasks possess fundamentally different supervision properties: style transfer relies on weak, abstract cues, whereas restoration requires high-fidelity alignment with the reference. This discrepancy raises concerns about negative transfer when attempting to integrate both under a single framework. Including empirical evidence or prior works demonstrating positive transfer between these tasks would strengthen the argument.
+- The inclusion of task-specific components, such as the label-processing function $M$ and hyperparameters like $\eta_1$, appears to contradict the paper’s claim of a unified framework. If these elements must be individually tuned per task, the true level of unification remains ambiguous.
+- In L263, the authors briefly compare their approach with SDEdit, but the described mechanism, mixing the noised conditioning image during the diffusion process, closely resembles the method proposed in ILVR [a]. A proper citation and discussion of ILVR should therefore be included for completeness.
+
+Minor Weaknesses
+
+- The deblurring setup assumes a Gaussian blur kernel, which may be an overly restrictive assumption and limit general applicability.
+- The paper seems to inconsistently use \citet and \citep throughout; all citations should be revised for consistency according to the chosen bibliography style.
+
+[a] Choi, Jooyoung, et al. ILVR: Conditioning Method for Denoising Diffusion Probabilistic Models., CVPR 2021
+
+### Questions
+See weaknesses
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper proposes DPG, which is a unified guideline framework for tasks with incomplete labels, covering both weak label settings (e.g., style transfer) and degraded label settings (e.g., super-resolution, denoising). The core idea is (1) to incorporate “data knowledge” by diffusing the (possibly processed) label itself and incorporating it into the early steps of the reverse diffusion process, and (2) to enforce “process knowledge” through a stepwise constraint, whereby each step of denoising improves the label alignment compared to the previous step. Experiments on style transfer and inverse problems show comparable qualitative and quantitative results.
+
+### Strengths
+- The proposed DFG can be used for stylization, SR, and de-noising, especially without changing the architecture. In addition, DFG can be easily applied to the existing reverse diffusion process.
+- The paper distinguishes between “what to inject" (signal derived from the label), “when to inject” (early steps), and “how to keep progress" (stepwise constraint). This modular presentation makes the pipeline easy to understand.
+
+### Weaknesses
+- The authors redefine stylization as “weak label guidance” and classic inverse problems (e.g., super-resolution, denoising) as “degraded label guidance,” which differs from established usage in the computer vision literature. Using the available measurement/observation $y$ as a guide during the reverse diffusion process is already common practice in existing algorithms for both stylization and inverse problems. Therefore, this work cannot be considered the first to "unify" these approaches; such a formulation risks confusing the readership rather than clarifying the contribution.
+- To capture the measurement/observation inforamtion, $z_{0|t}$ is refined  using eq. (9). This objective function $\ell_1 (z_{0|t},  y)$ is already frequently used to improve the performance of stylization and inverse problems. This cannot be considered a novelty of the proposed DPG.
+- For process knowledge integration, $\alpha_\text{margin}$ would strongly influence the performance of tasks. Depending on other hyperparameters such as $T_1$, $\eta$, and $\gamma$, the quality of results may change. However, the authors do not provide any ablation studies. They just offer simple guidelines in the supplementary material. Based on the currently available information, the algorithm cannot be reproduced. 
+- The qualitative comparisons presented in the paper cannot confirm that the proposed DPG clearly outperforms existing algorithms. Furthermore, the image quality is not sufficient to compare the algorithms.
+
+### Questions
+- For the data knowledge integration, why are two linear interpolations fundamentally required in eq. (7) - (i) mixing the latent ($z_t$ and $\hat{c}_t$) and (ii) mixing the predicted noise (${\epsilon}(z_t, \cdot)$ and $\epsilon(c_t, \cdot)$? In other words, what stability principle makes both convex combinations necessary together, rather than one of them (or another mechanism) being sufficient?
+- How high are the runtime and memory requirements of DFG compared to the standard reverse diffusion processes and existing algorithms?
+- The ablation studies in relation to the hyperparameters
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+DPG is a training-free guidance method for “imperfect-label” problems: both weak labels (like a style image) and degraded labels (like low-res or blurred inputs). It first applies a simple task-specific pre-process to the label (e.g., upsampling or Wiener deblur). In the early denoising steps only, DPG injects “data knowledge” by blending that noised label with the current latent and by mixing its predicted score with the model’s normal score, giving the sampler scale-matched structural hints without hard constraints. At each step, a task loss nudges the current clean prediction toward the label (e.g., VGG/CLIP losses for style; re-degradation consistency for SR/deblur). Across the whole trajectory, DPG adds “process knowledge” via a margin term that enforces step-to-step improvement in the task loss, which suppresses zig-zagging and error accumulation.
+
+### Strengths
+- Novelty on label integration: forward-diffusing the label and blending it for additional integration of the label seems interesting. While I am not fully convinced why this should work well, it seems novel. 
+- Performance: compared to baselines, DPG seems to be superior on various tasks.
+
+### Weaknesses
+- Unfortunately, the paper format doesn't follow ICLR guidelines yet. The guidelines say as follows: ```Citations within the text should be based on the \texttt{natbib} package
+and include the authors' last names and year (with the ``et~al.'' construct
+for more than two authors). When the authors or the publication are
+included in the sentence, the citation should not be in parenthesis using \verb|\citet{}| (as
+in ``See \citet{Hinton06} for more information.''). Otherwise, the citation
+should be in parenthesis using \verb|\citep{}| (as in ``Deep learning shows promise to make progress
+towards AI~\citep{Bengio+chapter2007}.'').``` To follow this guideline, the first sentence of the introduction section should start with "In recent times, diffusion models (DMs) (Ho et......)". Unfollowing this ICLR format guideline makes it hard to read the paper. 
+- DPG introduces many hyperparameters, because it should determine 1) mixing weights, 2) early-step window, 3) margin, and 4) inner step sizes. If this parameter should be changed according to tasks, the author's main claim that DPG is a generalizable framework for weak-label and degraded-label tasks should not hold, because this means not a generalizable framework but an adjustable framework according to tasks.
+- Justification depth: While intuitions are clear, the paper lacks a deeper analysis of line 250: "By adding noise and applying guidance, we let the model select the most relevant information for the task." My question is whether just mixing the noisy label is enough to change the model's behavior to select the most relevant information. In the rebuttal and discussion phase, I would like to discuss this with the author and revise the paper to correctly understand the reader. In current form, I am more towards that just mixing is not enough to enforce the model to choose relevant information only.
+
+### Questions
+- In Appendix, I found that the hyperparameters should be changed according to tasks. Is DPG sensitive to hyperparameter changes?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper introduces DPG (Diffusion Process Guidance), a training-free framework for improving diffusion model guidance in imperfect-label tasks such as style transfer, super-resolution, and deblurring.
+DPG combines data knowledge, by injecting diffused representations of imperfect labels into early denoising steps, and process knowledge, by enforcing progressive consistency across timesteps.
+This unified approach refines diffusion trajectories without task-specific tuning, achieving better perceptual quality and fidelity across diverse restoration and translation tasks.
+
+### Strengths
+The paper presents a unified and training-free framework that creatively combines data- and process-level knowledge for diffusion guidance.
+The idea of injecting diffused label information and enforcing progressive alignment offers a novel and generalizable formulation across multiple imperfect-label tasks.
+Experiments are comprehensive and consistent, showing clear improvements in perceptual quality.
+The paper is well-written and easy to follow, with solid motivation and clear methodological structure.
+
+### Weaknesses
+1. **Clarity and originality of the “unified framework” claim.**
+   While the paper frames DPG as a unified, training-free guidance framework, several prior works have already explored similar directions. For example, *Manifold Preserving Guided Diffusion* (ICLR 2024) formulated a single training-free framework that covers super-resolution, deblurring, face-ID, CLIP, and style guidance—all without an explicit forward operator. Thus, the novelty of unification alone may be limited. The introduction dedicates considerable space to emphasizing potential advantages of such unification, but the paper would benefit from demonstrating concrete benefits (e.g., improved generalization across tasks or reduced hyperparameter tuning) rather than discussing them conceptually.
+
+2. **Task-specific dependency of the mapping function ( M ).**
+   Although the framework is described as general, it still requires defining a mapping function ( M ) to transform task-specific labels ( y ) into an image-like form. This step is not always straightforward and may vary considerably across tasks. In some inverse problems (e.g., phase retrieval or Fourier-domain tasks in *Diffusion Posterior Sampling*), such a mapping is not directly feasible, which somewhat limits the claimed generality of the approach.
+
+3. **Missing discussion and comparison with closely related work.**
+   The *Manifold Preserving Guided Diffusion* paper—arguably the most methodologically similar prior work—is not cited or discussed. That work also introduces an ( x_0 )-alignment loss with target information, conceptually close to Equation (9) here. Including a citation, a brief discussion, and possibly a quantitative comparison (e.g., in Fig. 4 or Table 1) would clarify the distinction between DPG and prior frameworks and make the contribution positioning clearer.
+
+4. **Need for deeper analysis of the core components.**
+   The main novelty appears to lie in the *Data Knowledge Integration (DKI)* and *Process Knowledge Integration (PKI)* modules, yet their effects are not analyzed in depth.
+   – For **DKI**, injecting label information during early denoising is an intuitive and appealing idea, particularly since early predictions are often blurry and unstable. However, the ablation table (Table 2) seems inconsistent—entries (b) and (c) report significantly lower PSNR than the “w/o D” or “w/o P” variants, which might be a reporting error.
+   – To help readers better appreciate DKI’s role, it would be valuable to visualize several denoising trajectories, comparing predicted ( x_0 ) with and without DKI. Such qualitative evidence would make the mechanism’s effect clearer and strengthen the empirical analysis.
+
+### Questions
+- Basically, all points mentioned in the Weaknesses section.
+- The role of ( $\alpha_{\text{data}}$ ) is intuitively clear—it controls how much information is drawn from the label—but the meaning of ( $\gamma_{\text{data}}$ ) is less so. It seems analogous to a CFG-style guidance term interpolating between predictions from label-mixed and original latents. Why is this additional guidance needed if prediction from the label-mixed latent alone is already available? What new information is obtained from extrapolating between the two predictions? Given that this doubles the computational cost, please justify the necessity of this step and, if possible, provide qualitative or quantitative results across several key tasks for different ( $\gamma_{\text{data}}$ ) values.
+
+If I have misunderstood this mechanism, I would welcome clarification. My rating could increase accordingly.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3

human_reviews/CrCPeM5fRB.md

@@ -0,0 +1,155 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces the Signal Dice Similarity Coefficient (SDSC), a structure-aware metric for time-series self-supervised learning. SDSC extends the Dice–Sørensen coefficient, widely used in image segmentation, to signed continuous signals by (1) gating overlaps with a (smoothed) Heaviside on sign agreement between prediction and target, and (2) accumulating pointwise intersections via a $\min(\|E\|,\|R\|)$ operation normalized to $[0,1]$. The paper further proposes a hybrid loss that combines SDSC with MSE, weighted by uncertainty-based coefficients, and evaluates both as reconstruction objectives within a SimMTM pretraining framework.
+
+### Strengths
+- **Conceptual originality:**: Extending Dice coefficient to continuous, signed time-series is intuitive yet non-trivial. The proposed formulation yields a bounded, symmetric, and interpretable metric within the range $[0,1]$.
+- **Design rationale**: Using $\min(\|E\|,\|R\|)$ with a Heaviside gating directly targets polarity issues and reduces pure amplitude bias, which are key limitations of conventional MSE/MAE.
+- **Efficiency and simplicity**: SDSC avoids explicit temporal alignment or complex dynamic programming, making it lightweight and easy to implement compared to SoftDTW or DILATE.
+- **Hybrid loss formulation**: The uncertainty-weighted combination of SDSC and MSE provides a balanced approach that couples structure-awareness with amplitude precision.
+
+### Weaknesses
+1. **Narrow definition of "structure"**: SDSC captures only pointwise magnitude overlap under sign gating, overlooking broader structural properties such as local waveform shape and phase alignment. Consequently, it lacks time-shift/warping tolerance—small temporal lags can significantly reduce the score, contradicting the intended “structure-aware” characterization.
+2. **Offset bias**: When both signals share the same polarity or a strong DC offset, $H(E\cdot R)\approx1$ holds broadly, leading to inflated similarity even when shapes differ.
+3. **Zero-crossing noise and stability**: Around near-zero amplitudes, the Heaviside gating becomes highly noise-sensitive, causing unstable or vanishing gradients. Although the authors acknowledge gradient vanishing and describe it as robustness, this more likely indicates blindness to misalignment rather than genuine robustness.
+4. **Evaluation mismatch**: If MSE is said to fail at capturing structure, the evaluation should include structure-aware metrics such as Pearson correlation, spectral coherence, or STFT/Mel-cosine similarity. Relying solely on MSE/MAE weakens the empirical claim of structural fidelity.
+5. **Limited baselines**: The work omits direct comparisons with established structure-aware objectives. Simply stating these are "computationally heavy" is insufficient. An explicit time-vs-accuracy trade-off or short-sequence comparison would strengthen the argument.
+6. **Gradient analysis limitations**: Reporting only gradient norms does not characterize optimization behavior. The analysis should also consider gradient direction alignment, variance, or loss-landscape smoothness to assess stability.
+
+### Questions
+1. Could the authors include experiments with timing shifts or mild time warping, and compare SDSC against MSE, SoftDTW, and DILATE?
+2. Beyond pointwise overlap, could you report additional structure-sensitive metrics to substantiate the "structure-aware" claim both at pretraining and downstream evaluation stages?
+3. Have you explored mean-removal preprocessing or frequency-domain SDSC to mitigate offset bias?
+4. Could you provide a practical guideline summarizing when SDSC, MSE, or the hybrid loss is preferred (i.e., amplitude-critical vs. structure-critical regimes)?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces the Signal Dice Similarity Coefficient (SDSC), a structure-aware metric for time series self-supervised representation learning that addresses limitations of distance-based objectives like MSE.  SDSC extends the Dice Similarity Coefficient from image segmentation to continuous temporal signals by quantifying structural agreement through signed amplitude intersections.
+
+### Strengths
+1. The paper is relatively simple and easy to understand.
+
+2. Experiments span multiple tasks (forecasting, in-domain classification, cross-domain classification), settings (frozen encoders, fine-tuning), and datasets, demonstrating broad applicability and providing nuanced insights into when each approach works best.
+
+3. The paper acknowledges that SDSC models achieve higher structural alignment at the cost of increased MSE, and that dataset characteristics influence which approach works better, showing intellectual honesty.
+
+### Weaknesses
+1.  The paper uses only SimMTM as the backbone "for architectural simplicity," which severely limits generalizability claims. Without validation on diverse architectures (Transformers, CNNs, RNNs, recent foundation models), it's unclear if SDSC benefits are architecture-specific or truly general.
+
+2. The paper only compares against MSE, PCC, and SI-SNR. Recent structure-aware losses for time series (e.g., shape-based losses, spectral losses, contrastive losses) are not included, making it difficult to assess whether SDSC represents state-of-the-art for structure-aware objectives.
+
+3. Despite strong motivation, the actual improvements are modest: hybrid loss achieves 0.4783 vs. 0.4852 MSE in forecasting (1.4% improvement), and Table 6 shows MSE sometimes outperforms SDSC in fine-tuning scenarios. The gains don't match the strength of the conceptual contribution.
+
+4.  Computing SDSC requires element-wise min operations, Heaviside evaluations, and additional summations compared to MSE. No analysis of training time overhead or memory consumption is provided, which is critical for practical adoption.
+
+### Questions
+1. There seems to be an incorrect line break at the title.
+
+2. Why does SDSC underperform MSE in cross-domain settings (Tables 5, 6)? If SDSC provides better "semantic representations," shouldn't it generalize better across domains? What explains this contradiction?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes the Signal Dice Similarity Coefficient (SDSC), a novel, structure-aware metric for self-supervised learning (SSL) of time-series representations. Existing methods such as MSE is overly sensitive to signal amplitude and scale, while being insensitive to waveform structure, phase, and polarity. DTW and other metrics are also having their weakness. To overcome teh limitations, SDSC is adapted from the Dice Similarity Coefficient (DSC), a metric widely used for overlap in image segmentation. SDSC compares two signals at each time step. It calculates a score based on the minimum amplitude of the two signals, but only if both signals have the same sign (e.g., both are positive or both are negative). If the signs are different, that time step is heavily penalized (it contributes zero to the similarity score).Properties: This approach makes SDSC robust to amplitude scaling while being highly sensitive to polarity mismatches. The resulting metric is bounded between [0, 1] (making it interpretable) and is computationally efficient (linear $O(T)$ complexity), unlike other alignment-based metrics (like SoftDTW, which is $O(T^2)$). The author also uses a smooth sigmoid approximation to make the function differentiable. In experiments, they also propose to combines MSE and SDSC, to capture both structural and amplitude information.Experiments on forecasting and classification tasks show that pre-training with SDSC or the hybrid loss is competitive or superior to models pre-trained with MSE.
+
+### Strengths
+1. The proposed method is efficient, linear time complexity, much better than other methods like DTW, which performs similar structure-awareness measurement.
+2. The metric is bounded from 0 to 1, which provides better interpretability.
+3. The paper is clean written, with illustrative examples to show numbers using different metric, under different structure changes.
+
+### Weaknesses
+1. No clear definition of "structure", still related to alignment or warping.
+2. The backbone model is not widely tested. With more powerful models, we don't know if the advantage of SDSC still exists.
+3. The imrpovement on various tasks, are very marginal. For example, in the fine-tuned classification task, the SDSC approach is not showing better results in either in-domain or out-domain experiments.
+
+### Questions
+If switching to other transformer-based backbone model, would the proposed method performs consistenly better?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper addresses the limitations of conventional distance-based metrics (e.g., MSE) in time-series SSL by introducing a novel structure-aware metric, the Signal Dice Similarity Coefficient (SDSC). The method reframes the signal reconstruction problem as measuring the overlap of the areas under the respective curves. Through the introduction of a signed amplitude intersection term, it ensures that overlap is computed only when signal polarities align, thereby effectively addressing the noted deficiency of MSE in being insensitive to phase inversions.
+
+### Strengths
+1.	Adapting the DSC from the segmentation domain to time-series signals is a novel perspective. Using the signed amplitude intersection as a proxy for waveform structure similarity is an interesting idea.
+2.	The O(T) linear complexity of SDSC is computationally efficient, which is a practical advantage.
+3.	The mathematical definition is intuitive, and the experimental design is well-structured.
+
+### Weaknesses
+1.	A motivation for the paper is that SDSC serves as a lightweight alternative to O(T^2) metrics (e.g., SoftDTW, DILATE). However, a direct comparison against them is missing. Currently, we only know that SDSC is faster, but we do not know how much performance is lost (or gained) compared to SoftDTW.
+2.	The α parameter in the Sigmoid function significantly influences the gradient shape. The paper lacks a sensitivity analysis on how α affects the performance of downstream tasks.
+3.	The "alignment-free" description may be somewhat misleading. While SDSC does not perform explicit temporal warping like DTW, it does enforce strict temporal and polarity alignment through its point-wise comparison.
+
+### Questions
+I will reconsider my score during the rebuttal phase based on the authors' response to following issues.
+
+1.	Could the authors include an experiment in the appendix that compares SDSC with SoftDTW (as a loss function) on a downstream task (e.g., forecasting), using at least one small-scale dataset?
+2.	Regarding the hybrid loss in Equation (8), how is the uncertainty-based tuning strategy specifically implemented to determine the weights λsdsc and λmse? How does this adaptive strategy compare to using simple fixed weights (e.g., λsdsc = λmse = 0.5)?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+3

human_reviews/DB2KJKFX0d.md

@@ -0,0 +1,171 @@
+## Human Reviewer 1
+
+### Summary
+To address limited availability of 7T data, this paper proposes a novel method for enhancing resolution and SNR of 3T BOLD fMRI to approximate 7T data quality. It aligns both 3T and 7T data from different subjects and datasets into the same space with shared parameters. By using an unpaired Brain Disk Schrodinger Bridge (BDSB) diffusion model, the spatio-temporal resolution and SNR of 3T data could be enhanced. Experiments were performed on three public fMRI datasets and a synthetic dataset. The proposed method showed comparable results of SNR and the goodness-of-fit of the pRF in the enhanced 3T to 7T data quality.
+
+### Strengths
+1) The motivation of the paper is clear and straightforward. Given the scarcity of large-scale paired 3T-7T fMRI datasets, research for unpaired learning frameworks is necessary, and this paper handles this challenge. 
+
+2) The method is validated on both real and synthetic datasets to demonstrate generalizability. Moreover, the results show that the proposed method outperforms GAN-based and diffusion-based approaches across 4 evaluation metrics, including $R^2$, which assesses neuroscientific interpretability.
+
+### Weaknesses
+1)	The exact and detailed formula of the regularization terms (PatchNCE and BD-SSIM) is missing. Which distance metrics were used for the regularization losses? It would be helpful to add the formulations in the supplementary material.
+
+2)	It is mentioned that the NCE regularization loss is designed to maintain consistency with the LQ inputs in order to preserve structural details. However, how can you guarantee that this term actually preserves ‘structural ‘ information? Are there any ablation studies regarding this term with qualitative analyses and/or brain regional comparisons that support this claim? In addition, if my understanding is correct, this term reduces the discrepancy between the LQ and generated data. In that case, it seems to potentially conflict with the role of $ L_{Adv}$, which enhances resolution (i.e., the main goal of this paper). How does the proposed method address this issue?
+
+3)	The paper does not include qualitative comparisons and neuroscientific comparisons with baseline models.
+
+### Questions
+1)	Given the small data size, how could you handle the overfitting issue in the experiments? Did authors adopt cross-validation and/or perform multiple trainings to validate the generalizability of the methods (including baselines)? 
+
+2)	How were the hyperparameters of baseline models tuned? Could you clarify the tuning strategy and selection criteria?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This work proposes BDSB, an Schrödinger Bridge application for 3T-to-7T cross-dataset fMRI enhancement. Experiments on synthetic dataset and real-world datasets demonstrated superior performance by BDSB over GANs and DDPM.
+
+### Strengths
+1. The paper is clear and well-organized. The way fMRI data fits into the Schrödinger Bridge framework has been clearly described, and readers can learn from it.
+
+2. Performance on synthetic and cross-dataset experiments has shown a promising boost over baselines. While on paired data, the improvement is marginal, BDSB can still hold top-2.
+
+### Weaknesses
+1. Notations need a double check, e.g., in Fig 3 caption, shouldn't the approximating distribution $\hat x_{1|t_i}$ be derived from the neural generator $q_\phi$ instead of the joint distribution $p$? Please correct me since I'm not an expert on diffusion models, but I also think it should be the approximated $\hat x_{1|t_j}$ in $p(x_{t_{j+1}}|x_{t_j},x_{1|t_j})$ in Fig 3.
+
+2. The first attempt of learning 3T-7T generation from unpaired data is desirable, but the real-world evaluation should be focused on paired data since the final purpose is generating ground truth 7T fMRI from 3T fMRI. In this regard, BDSB performs similarly to OTT-GAN for PSNR and even worse for SSIM.
+
+3. The contribution of a better fMRI enhancement via learning across datasets has not been evaluated. As mentioned above, there are no experiments of training with unpaired data and testing on paired data. The ability of learning across datasets makes BDSB more fundamental than existing models, and it should lead to a model scaled from multiple datasets. However, the BDSB is trained separately for unpaired and paired experiments.
+
+4. Technical innovation is limited. Why don't other optimal transport methods fit into the proposed framework?
+
+### Questions
+1. How's the performance of super-resolution methods on your data?
+
+2. How's cross-session fMRI prediction scientifically sound? There are various compounds affecting the BOLD signal aside from cognition and visual stimuli, such as scanner settings and test-retest variations [2]. How to ensure the model learning from SNR differences rather than other compounds?
+
+[1] Ding, Jiaqi, et al. "Machine Learning on Dynamic Functional Connectivity: Promise, Pitfalls, and Interpretations." arXiv preprint arXiv:2409.11377 (2024).
+
+### Soundness
+2
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes a method for enhancing 3T blood-oxygenation-level-dependent (BOLD) functional magnetic resonance imaging (fMRI) by leveraging an unpaired Brain disk Schrödinger bridge (BDSB) model. The authors map 3D brain surfaces into a shared parametric domain via conformal mapping and apply an unpaired BDSB diffusion model to approximate the higher resolution of 7T scans using 3T scans. The framework is evaluated across three public datasets, of which one is synthetic, another containing only 3T scans, and a paired 3T/7T dataset. 
+
+Experiments indicate meaningful improvements on the synthetic data and Cross-Dataset Real, while performance is essentially in line with OTT-GAN for the paired TDM Real dataset.
+
+### Strengths
+Originality. The fMRI enhancement pipeline consisting of conformal parameterization, brain disk schrödinger bridge and resample & pRG analysis appears novel. The use and application of the Schrödinger Bridge for unpaired samples is also novel in this domain.
+
+Quality. The method is mathematically motivated and the experiments include relevant metrics (SSIM, PSNR, R^2). Evaluation across synthetic, cross-dataset, and paired data provides a reasonable spread of conditions.
+	
+Significance. Enhancing 3T data using unpaired 7T examples is an important and practically relevant problem. The approach could, in principle, enable higher-quality analyses without costly high-field scans. Additionally, the method performs well on synthetic and cross-dataset real compared to the provided baselines. 
+
+Clarity. The paper is clearly written and well-structured overall. Figures 1–3 effectively illustrate the architecture and training process.
+
+### Weaknesses
+- While the method is framed as unpaired learning, the implementation and experiments do not convincingly show that unpaired samples are leveraged for learning. A meaningful test would involve partial or full training on unpaired data and evaluation on paired data to quantify the benefit of the unpaired setup.
+- On the TDM Real (paired) dataset, BDSB performs similarly to OTT-GAN, contradicting the claim of superior performance “across all real and synthetic experiments.” This discrepancy should be discussed explicitly.
+- Frechet Inception Distance (FID) typically relies on an ImageNet-trained network and is not meaningful for fMRI-like data, whose statistics differ drastically from natural images.
+- Related methods are only briefly mentioned in the appendix; a dedicated section would improve context and clarify how baselines are chosen.
+- When the authors or the publication are not included in the sentence, the citation should be in parenthesis using \citep{}, as outlined in the formatting instructions.
+
+### Questions
+- What do the authors believe explains the discrepancy between performance gains on synthetic/Cross-Dataset Real data and the parity with OTT-GAN on TDM Real?
+- Why are results from fast-DDPM missing for the Cross-Dataset Real setting?
+- How did the authors decide which baselines to include and which not to include?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper presents a flow-based approach to synthesize 7T fMRI from the counterpart 3T data. The Schrodinger bridge technique is used to train the model based on the 2D disk of surface projection map. While the idea of leveraging ultra-high-field fMRI for improved spatial modeling is potentially valuable, the current paper lacks methodological rigor, critical validation experiments, and sufficient theoretical justification to establish technical credibility. Strengthening these aspects would substantially improve the impact and trustworthiness of the work.
+
+### Strengths
+The paper is technically ambitious (given a limited number of training samples and high dimensional data) and conceptually original, combining **geometric brain mapping, diffusion modeling, and validation** into a cohesive pipeline. These contributions make it a practical solution for enhancing fMRI resolution for neuroscience studies.
+
+### Weaknesses
+1) **Main Concerns**. The primary issue with this submission lies in its scientific rigor and validation design. The authors claim that BOLD signals obtained from 7T MRI can be effectively mapped to or synchronized with evolving time series from 3T scanners. However, this assumption is not convincingly supported. It remains highly questionable whether (1) temporal synchronization between 3T and 7T acquisitions can be achieved with sufficient precision, and (2) whether the spatial correspondence of voxel- or surface-level activations can be meaningfully aligned across different field strengths. From a machine learning standpoint, however, the 3T and 7T datasets should ideally be paired or co-registered at the subject level to enable cross-modality learning. Without a clear justification or evidence supporting this assumption, the technical soundness of the proposed framework is undermined.
+
+2) **Projection Method**. The use of 2D conformal mapping to reduce fMRI data complexity is conceptually interesting, yet it raises serious concerns about potential information loss during the projection from high-dimensional cortical surfaces to a 2D representation. This approach may oversimplify the spatial geometry of the brain and distort the topological structure of functional activations. The authors should consider or at least discuss alternative representations, such as spherical harmonics (see the well-established approach by Anderson et al., NeuroImage, 2010), which preserve surface geometry while enabling efficient spectral decomposition.
+
+3) **Replicability and Test-Retest Validation.** Given that this is a neuroimaging study, replicability testing is a crucial standard for evaluating robustness. The current manuscript lacks experiments on test–retest datasets, which are commonly used to assess the reliability of functional signals and model generalization. The authors should include such analyses or provide clear justification for their omission.
+
+4) **Temporal Resolution Limitation**. Finally, the authors should acknowledge the inherent limitation that the proposed method primarily enhances spatial resolution and signal-to-noise ratio (SNR), but does not address the true temporal resolution gap in fMRI. The fundamental challenge remains the coarse temporal sampling (on the order of 1 s) compared to neuronal timescales (milliseconds). The paper would be stronger if it explicitly discussed this limitation and clarified whether the proposed method could, in principle, be extended to improve temporal fidelity.
+
+5) **Limited novelties**. This work is a combination of existing components such as conformal mapping and Schrodinger bridge model.
+
+### Questions
+1. The paper projects the cortical surface into a 2D disk before applying the Schrödinger bridge model. How much geometric distortion is introduced by this mapping, and how might it affect spatial correspondence between the 3T and 7T fMRI data?
+2. Why was a 2D conformal mapping chosen instead of a spherical or spectral representation (e.g., spherical harmonics)? Would the latter preserve global topology more faithfully?
+3. How are the boundary conditions handled in the 2D disk representation, given that cortical manifolds are not naturally disk-like?
+4. The model learns a flow between 3T and 7T BOLD signals, but are these data temporally synchronized or spatially paired? Without strict pairing, how can the model distinguish physiological differences from scanner-induced variability?
+5. How does the model ensure that synthesized 7T fMRI signals retain biologically meaningful temporal and spectral properties (e.g., frequency content, signal-to-noise ratio)?
+6. Although an ablation study is presented (Table 3), it primarily focuses on different surface mapping strategies and regularization terms. Was any ablation performed to isolate the contribution of the Schrödinger bridge formulation itself—e.g., by comparing with a baseline model without the bridge constraint or with a standard flow-based mapping?
+7. The model is trained on a small dataset of paired 3T–7T scans. Can it generalize to other sites, scanners, or acquisition protocols?
+8. What is the computational cost of solving the Schrödinger bridge compared to simpler flow-based mappings?
+9. How would the approach scale to whole-brain volumetric data rather than surface-based 2D projections?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+5

human_reviews/EbSkBZQF9g.md

@@ -0,0 +1,165 @@
+## Human Reviewer 1
+
+### Summary
+This paper applies Mechanistic Interpretability (MI) to examine why a single-layer transformer fails to grok the 0–1 knapsack problem. The analysis reveals that the model relies heavily on the capacity token and the MLP layer, while its embedding space is poorly structured and resembles random initialization. Based on these observations, the authors argue that transformers have fundamental difficulty with NP-complete problems due to combinatorial explosion. They further hypothesize that a transformer with klayers can only generalize tasks with time complexity up to O(n^k), suggesting that large language models may face inherent limitations for high-stakes computational decision-making.
+
+### Strengths
+-The paper addresses an underexplored research area by examining how Transformers generalize on NP-complete problems.
+
+### Weaknesses
+-The paper arrives at the broad conclusion that transformer-based models struggle to generalize on NP-complete problems, yet supports this claim using only a single case study (the 0–1 knapsack problem) and only with a single-layer transformer. This limits the strength of the conclusion.
+-The analysis is restricted to a synthetic dataset and a shallow model. To convincingly argue general limitations of Transformers on NP-complete tasks, the study should include additional NP-complete problems and architectural variants (e.g., deeper or pre-trained models).
+-While the paper includes many figures, the accompanying explanations are insufficient, making it difficult to interpret the key findings from the visualizations.
+-The work would benefit from a more formal theoretical examination of Transformer generalization in NP-complete settings, similar in spirit to prior analytical studies of model expressivity and computational limits [1, 2].
+-The authors state that state-of-the-art models were excluded due to computational constraints. However, validating a claim with such broad implications requires experimentation on more capable models; otherwise, the conclusions may not generalize beyond the minimal architecture tested.
+
+[1] Zaheer, M., Guruganesh, G., Dubey, K. A., Ainslie, J., Alberti, C., Ontanon, S., ... & Ahmed, A. (2020). Big bird: Transformers for longer sequences. Advances in neural information processing systems, 33, 17283-17297.
+[2] Peng, B., Narayanan, S., & Papadimitriou, C. (2024). On Limitations of the Transformer Architecture. First Conference on Language Modeling.
+
+### Questions
+-What is the rationale for choosing the 0–1 knapsack problem as the case study? There are numerous NP-complete problems (e.g., n-SAT, Traveling Salesman, Hamiltonian Cycle) that may exhibit different structural characteristics. Clarifying this choice would help contextualize the scope of the conclusions.
+-At which training stage(s) were the attention patterns visualized, and how do these patterns evolve over the course of training? Showing the progression could provide valuable insight into the model’s learning dynamics.
+-The paper states that probing analysis shows the model “perfectly stores up to half of the weights and prices,” but detailed results or visualizations of this probing are not provided. Could the authors include these results to substantiate the claim?
+-The hypothesis that a k-layer Transformer can only generalize tasks with time complexity up to O(n^k) is central to the paper, but the theoretical and empirical basis for this claim is not fully articulated. To make the hypothesis more testable and convincing, one possible experiment would be to vary model depth and examine generalization performance on a known polynomial-time task such as sorting—where a one-layer model would be expected to fail, while a two-layer model should succeed. This would provide direct support (or counter-evidence) linking model depth to computational complexity.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper investigates why a small, single-layer transformer fails to solve NPC problem. The authors first construct a dataset of four-item 0-1 knapsack problem. Using the TransformerLens library, the authors trained a one-layer transformer. Through mechanistic interpretability methods—such as attention visualization, singular value analysis, logit lens, probing, and activation patching—they discovered that the model overemphasizes the knapsack’s capacity token, encodes information weakly, and lacks structured internal representations. These findings suggest that transformers struggle with NPC problems.
+
+### Strengths
+The research topic is important. Most of the difficult reasoning questions in the real world are NP-problems, like the Puzzle of 24.
+
+The authors applied various mechanistic interpretability methods to interpret the model.
+
+### Weaknesses
+1. The paper may need to improve its presentation. For example, (1) explain the figures, (2) visualize the method and data, (3) make subsections, (4) split methodology and experiment.
+
+Besides, to improve the flow, a general idea is not to use "we use xxxxx and we find xxxx", or "xxx is xxx, we use it and find xxx". A more easy-to-follow flow is first introducing the motivation, or the overview sentence of the paragraph or the subsection.
+
+2. The author may need to add the discussion of related works and the discussion of the comparison with existing works. This paper is very relevant to LLM reasoning, LLM search/exploration.
+
+3. The author may want to reconsider the scope of this work. I acknowledge that computation is a common issue, but considering more diverse tasks of NPC problems, or considering small-scale LLMs (e.g., 300M, 2B), is reasonable. The claim in the paper is strong, but the evidence is weak.
+
+### Questions
+See weaknesses.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper studies whether a single-layer Transformer (via TransformerLens) can “grok” 0-1 knapsack when trained on an algorithmically generated dataset with 4 items per instance. The model is trained up to 100k epochs to predict the optimal total price given tokenized weights, prices, and capacity. It does not grok: train/test curves do not exhibit delayed generalization. The authors then run a suite of mechanistic interpretability probes, attention visualizations (capacity token receives most mass), logit lens (MLP dominates), linear probes (some weights/prices are recoverable; capacity poorly encoded), activation patching (capacity-focused neurons drive loss), and SVD/PCA on embeddings (spectra resemble random; unlike modular subtraction). They conclude that shallow Transformers struggle to form robust circuits for NP-complete tasks and hypothesize a depth-time-complexity link, i.e., k layers ≈ O(n^k ) algorithms.
+
+### Strengths
+1. Clean MI toolkit application (attention maps, logit lens, probes, activation patching, spectra). 
+3. Transparent reporting that the model fails to grok and that capacity dominates attention. 
+3. Honest limitations section and compute-aware motivation.
+
+### Weaknesses
+1. Fixed 4-item knapsack; no scaling across #items, model width/depth, or data size. 
+2. No MLP/RNN/DeepSets or symbolic/DP baseline; no depth>1 ablations. 
+3. Unclear train/test split; no out-of-distribution tests (e.g., new capacities, price/weight ranges). 
+4. Predicting only the optimal value sidesteps combinatorial structure (subset selection); may bias learning signals.
+5. NP-complete generalization and O(n^k) depth hypothesis are unsupported by the presented evidence. 
+6. SVD/PCA on embeddings is suggestive but not decisive; no quantitative eval of probe accuracy vs layer/position.
+
+### Questions
+1. Why predict only the optimal value instead of the optimal subset (or value + subset)? Does predicting the subset change the behavior of grokking? 
+2. Report curves across items (4→6→8), depth (1→2→4), width, and data size. Where (if anywhere) does grokking emerge? 
+3. Add DP (optimal), greedy heuristics, MLP/DeepSets, and a 2–4-layer Transformer to separate depth vs optimization effects. 
+4. Specify train/test generation, show IID vs OOD generalization, and formal grokking metrics (e.g., Power et al.). 
+5. Provide causal tracing for capacity vs price/weight pathways and quantify probe R^2 /accuracy by token & layer. 
+6. Either theorize the O(n^k) connection or tone it down; a single 4-item, 1-layer experiment cannot support it.
+7. Try factored inputs (set encoders / permutation-invariant layers) to test if failures stem from sequence order, not task hardness. 
+8. Compare classification over all feasible values vs regression vs program-of-thought target; report which helps.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper applies mechanistic interpretability to a single-layer transformer trained on a small 0-1 knapsack dataset, finding it fails to generalize. It analyzes internals via visualizations and techniques, hypothesizing transformers struggle with NP-complete problems and are limited by layer count in computational complexity. Contributions aim to highlight LLM limitations in planning tasks.
+
+### Strengths
+The paper explores mechanistic interpretability on an NP-complete toy problem, extending beyond simpler prior tasks. It uses established tools to probe model failures, raising relevant safety concerns for AI agents in complex domains.
+
+### Weaknesses
+The claims about broad transformer limitations may be somewhat overstated given the constrained single-layer setup and small scale; exploring multi-layer models or larger instances could strengthen them. The methodology would benefit from additional baselines, clearer grokking criteria, and more ablation experiments. Presentation could be improved by addressing typos, providing more details, and better integrating figures. To improve, the experiments would have to be much more extensive, and any claims would have to have stronger backing from empirical evidence. 
+
+The mechanistic interpretability techniques are from previous literature, and don't offer any compelling evidence for the claims in the paper. It isn't strictly necessary for the paper to introduce new mechanistic interpretability techniques, but I'd recommend looking at [2] for an example of a stronger algorithm-oriented approach to interpretability. 
+
+It may be helpful for the authors to review work on tool-use. For instance, an argument can be made that even if LLMs cannot learn algorithms robustly, they can call tools to solve sub-problems, e.g. a knapsack solver. I personally think this argument has some flaws, but I'd encourage the authors to look into tool-use to strengthen their paper and claims. 
+
+The paper ignores work on transformer expressivity and algorithm learning, e.g. [1], which would make it clear that NP-complete algorithms shouldn't be able to be represented in a tiny transformer. Claims that transformers can learn simpler algorithms (abstract sentence 1, line 11), aren't really true [3], so it isn't necessary to jump to NP-complete problems. For example, transformers struggle to learn the parity algorithm effectively [2]. 
+
+[1] Merrill, W., Sabharwal, A., & Smith, N. A. (2022). Saturated transformers are constant-depth threshold circuits. Transactions of the Association for Computational Linguistics, 10, 843-856.
+
+[2] Shaw, P., Cohan, J., Eisenstein, J., Lee, K., Berant, J., & Toutanova, K. (2024). Alta: Compiler-based analysis of transformers. arXiv preprint arXiv:2410.18077.
+
+[3] Markeeva, L., McLeish, S., Ibarz, B., Bounsi, W., Kozlova, O., Vitvitskyi, A., ... & Veličković, P. (2024). The clrs-text algorithmic reasoning language benchmark. arXiv preprint arXiv:2406.04229.
+
+### Questions
+I will use this section for a comment: While I do not recommend publication of this paper, I do not want to discourage the authors. I believe that their original intended idea is important to study, and hope to provide them with constructive options to explore for a future paper. My main comment is that it may be more valuable to study simpler algorithms, and connect the failures in representing these algorithms to common AI-agent approaches, e.g. to back up their claims in the abstract (lines 18-20).
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5

human_reviews/FEqZmGl36g.md

@@ -0,0 +1,211 @@
+## Human Reviewer 1
+
+### Summary
+The authors propose ESS-Flow, a method to guide sampling from probability flow generative models without gradient-based guidance. Their method samples with MCMC along ellipsoids in the source space (Gaussian prior), and takes a step based on the potential /target property evaluated in the data space. They motivate their method with the limitations of gradient-based guidance and evaluate their guidance on protein and material design tasks.
+
+### Strengths
+- Guidance without gradients is a nice benefit of the proposed method
+- The method is motivated well theoretically and with prior work
+- The experiments suggest that ESS-Flow is able to effectively guide samples compared to gradient-based approaches
+
+### Weaknesses
+- Does guidance for certain properties improve the estimation of other properties/observables not used in guidance?
+- Given that one of the benefits of the proposed method is not relying on gradients, it would strengthen the paper to show guidance for discrete target properties.
+- The discussion of challenges about challenges with gradient-based guidance (Fig. 2) is interesting, and I think further discussion / experiments along these lines would further strengthen the paper. I wonder if there is a way to show something similar with more realistic data by artificially creating 2 disconnected modes (i.e. removing data in a transition region or something like that).
+- "Limits ESS-Flow’s effectiveness when the prior does not well inform the target distribution": while I understand that things like image inpainting might be challenging for the method, have the authors evaluated ESS-Flow on other image guidance tasks?
+- There are newer and more accurate models compared to something like CHGNet that might give more accurate metrics (MACE, eSEN, UMA, etc.)
+
+### Questions
+Please see above.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposed a plug-and-play controllable generation method for flow-based generative models, which is based on sampling the controlled prior distribution through elliptical sliced sampling (ESS), and then run the ODE to convert the prior samples to samples from the tilted target distribution. The method avoids the need to compute the Jacobian of the flow map $T _ \theta$, and does not require the tilt to be differentiable. The authors further proposed to use a coarse discretization of the flow ODE as a proposal for the transition kernel and reweight the samples with the evaluation of the tilt on the final samples to reduce the computational cost.
+
+### Strengths
+Unlike most existing controllable generation methods such as CFG and DPS that involve the computation of an extra term in the SDE (often requires approximation), the proposed method just treat the flow ODE as a black-box and only modifies the prior distribution, and instead of doing gradient-based optimization, it directly samples from the controlled prior distribution. This makes the method simple and concise, and I personally appreciate this idea.
+
+The paper also provided comparisons of the proposed method with existing controllable generation methods based on optimization. I'm not familiar in the domain of material design and protein structure prediction, but the experimental results look nice.
+
+### Weaknesses
+The authors provided limited background information about the ESS algorithm, which I believe most of the readers in the ML community should not be familiar with. We know that there are lots of zeroth-order sampling methods for sampling from $\pi(z)\propto p(z) g(T _ \theta(z))$, such as rejection sampling, the MH algorithm, and proximal sampler (https://proceedings.mlr.press/v134/lee21a.html). What's the insight behind ESS, and why is the Gaussian prior important for it to work? (Also I'm not quite in favor of the use of the abbreviation ESS in this paper, since it is also used to refer to "effective sample size" in the literature of Monte Carlo methods, which may cause some misunderstanding. But it's fine if you keep it.)
+
+For the experiments, I think it would be more convincing if the authors can also consider some image generation tasks, which are more familiar to most ML researchers, and can better demonstrate the effectiveness of the proposed method. Also, as the authors have introduced the multi-fidelity version of the proposed method through a coarse discretization of the flow ODE, it would be interesting to see some ablation studies on the choice of the discretization step size and its impact on the generation quality and computational cost.
+
+I'm happy to raise the score if these issues can be addressed during rebuttal.
+
+### Questions
+1. What's the typical number of rejections needed in one iteration of ESS in order to get an accepted sample?
+
+2. We know that diffusion model and flow matching allow one-step prediction of $\mathbb{E}[x _ 1|x _ t]$. For faster sampling, instead of still doing discretization of the flow ODE, do you think it is possible to replace the flow map $T _ \theta$ with a one or few step predictor in the proposed method? If a consistency model is available, we can even directly predict the final sample from the prior sample in one step, which may further reduce the computational cost.
+
+3. The main text of the paper assumes the flow model is trained for Euclidean data, but in the experiments it turns out that the whole framework can also be applied to manifold data as long as we have a flow-based generative model, which I believe is an advantage of the proposed method, and should be highlighted more in the paper. I'm not quite familiar with the literature of manifold, but is there anything that we need to pay attention to when applying the proposed method to manifold data? For example, do we need to modify the ESS algorithm in any way?
+
+4. In table 2 and 3, it would be better to highlight the best results in bold font for better readability.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces ESS-Flow, a training-free and gradient-free guidance method for flow-based generative models. The key idea is to do Monte Carlo inference directly in the source space, where the prior is Gaussian, by sampling from a distribution of the form $\pi(z) \propto g(T_\theta(z))p(z)$. This makes the method applicable to quantized/materials settings and to non-differentiable observation or reward functions. The authors further propose a multi-fidelity variant where the authors sample using a coarse ODE discretization and reweight using a fine discretization to reduce computation. Experiments on materials and protein structure prediction show that ESS-Flow achieves nearly SOTA results on all metrics for materials and comparable performance on protein generation metrics.
+
+### Strengths
+1. The paper is clearly written, with a coherent structure and well-motivated.
+
+2. The proposed method is simple to implement and achieves strong results on the material generation task.
+
+### Weaknesses
+1. The paper lacks experiments on standard image-domain tasks (e.g. inpainting, deblurring), which are commonly used in related work such as D-Flow and PnP-Flow.
+
+
+2. While the empirical contribution is solid, the theoretical contribution is relatively modest.
+
+3. One of the main points of the contribution is having this apply to non-differentiable rewards/potentials yet there are no experiments demonstrating this capability.
+
+### Questions
+1. What are the quality metrics in Table 3 for the competing methods? It would be helpful to report the same set of metrics so we can compare ESS-Flow directly to the baselines.
+
+2. How does multi-fidelity ESS compare to standard ESS on the metrics reported in Table 2 and Table 4.
+
+3. What are the acceptance rates of the sampling?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The authors introduce ESS-Flow, a gradient-free method for controlled generation in the setting of generative modelling with flow matching models. The authors perfoming Bayesian inference in source space using Elliptical Slice Sampling, which enables conditional generating without requiring gradients. The authors demonstrate their approach on various applications ranging from materials to proteins.
+
+### Strengths
+[S1] The gradient-free nature of this approach is appealing. There have been other works for source space sampling such as , but this still required gradients, wheras this approach here circumvents this via the Jacobian cancellation and the ESS approach.
+
+[S2] One common with gradient-based optimisers in diffusion samplers are the multitude of often brittle hyperparameters like guidance scales or other schedulers; this approach here seems to be less reliant on these.
+
+[S3] Good motivation of the different components introduced with formal justifications.
+
+### Weaknesses
+[W1] The authors demonstratet that they can avoid the Jacobian computation, but ESS-Flow still requires many evals (>1000 MCMC steps) of the transport map, hurting the efficiency of the approach
+
+[W2] The authors openly describe the limitation of ESS-Flow in cases where the target is constrained on a lower-dim manifold, but claim that in scientific domains the target distribution is not overly collapsed. However, in many applications like in protein design the target distribution lies exactly on such a lower dim manifold with most of the target space being invalid sampels; some more explanation why the authors think that this is not the case in many scientific applciations would help here.
+
+[W3] In many scientific applications, people have circumvented the non-differentiability of categorical sequences via soft relaxations similar to the atom relaxation the authors use for their comparisons. In approaches like BindCraft (Pacesa et al, 2025 Nature), this works remarkably well, so the authors should potentially try to tune that baseline to see if it as strong as it can be.
+
+[W4] While some of the baselines in the protein structure prediction case of Figure 4 look unrealistic, ESS-Flow also seems to have biophysical implausibilities, and the ELBO of the model only partially captures these things. A more fundamental evaluation like counts of clashes could demonstrate how good the structures actually are; the RMSD values above 10 suggest that all baselines seem pretty far off.
+
+### Questions
+[Q1] The case studies all have quite low dimensionality, how does the approach scale to high dimensional problems?
+
+[Q2] Given the authors say their method does not work well with priors that poorly inform the target distribution, can this statement be made more exact? ie is there a quantity that one can look at to see if the approach will work or not?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+The paper introduces ESS-Flow, a training-free and gradient-free approach for controlled generation using pretrained diffusion or flow-based models. The method reframes inference in the Gaussian latent source space and applies elliptical slice sampling (ESS) instead of gradient-based updates, using a change of variables to apply updates in the data space. The main goal is to enable preference alignment when gradients are unavailable or unreliable, such as in cases involving quantized or simulator-based objectives.
+
+### Strengths
+- The paper introduces a new and practically useful application of elliptical slice sampling to flow- and diffusion-based generative models. While ESS itself is a known MCMC method, its use within the latent Gaussian space of pretrained flow-based models is novel and interesting.
+- The algorithm is designed such that jacobian determinants of transport maps don’t need to be computed, enabling a scalable and efficient algorithm.
+
+
+- The theory is sound, and the algorithm preserves theoretical guarantees from the original ESS method, making interpretability for downstream researchers easier.
+Reported results compared to baselines show a non-negligible improvement in matching a new target energy function.
+
+
+- Overall, the paper connects ideas from generative modeling, Bayesian inference, and MCMC in a coherent and insightful way. It is an interesting paper that I believe would bring a net positive to the research community, which would be strengthened by addressing the weaknesses below.
+
+### Weaknesses
+- While 0th-order methods can benefit greatly in settings with unreliable gradients, they can also face severe scaling issues in high-dimensional spaces. Having an experiment showing performance as problem dimension scales would help further inform future readers when they should use ESS-flow vs. a gradient-based method.
+
+
+- While results on the provided experiments show that the 0th order methods are outperforming the gradient-based methods, none of the experimental settings to my understanding actually fall under the non-differentiable setting that the paper proposes to address. It would help to either (a) include a relevant experiment setting where gradient information is truly intractable to retrieve, or (b) demonstrate that in the reported settings, the gradient structure is highly unideal for gradient-based methods, e.g. high Lipschitz constant (also, for line 290, maybe better to call them gradient-based rather than optimization-based, since there are many 0th order optimization algorithms).
+More statistical details on the experimental setup are needed, e.g. some std’s in table 2 are higher in magnitude than the mean.
+
+
+- There are some missing baselines/ablations (e.g. adjoint matching [Domingo-Enrich et al.], non-ESS-based source space MCMC) that would help clarify whether the main benefit comes from being gradient-free or from the specific ESS mechanism.
+
+### Questions
+- How does ESS-Flow scale with increasing latent dimension? Does the acceptance rate or effective sample size drop significantly in higher dimensions?
+
+
+- What is the runtime cost compared to the baselines?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/FF9QVQduAu.md

@@ -0,0 +1,169 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces CrowdFM, a graph neural network-based "foundation model" for crowdsourced label aggregation, pretrained on synthetic data to enable retraining-free, cross-dataset generalization. The model is evaluated on 22 real-world datasets and shows strong empirical performance and efficiency. However, the main weakness of this paper is its failure to acknowledge and position itself with respect to highly relevant prior work—specifically, the hyper label model (Wu et al., ICLR 2023). The methodology, motivation, and even the technical design of CrowdFM are very similar to those of the hyper label model, with the only substantive difference being the application domain (crowdsourcing vs. programmatic weak supervision) and the technical details in training data generating and the graph neural network architecture design. The paper overstates its novelty and does not provide a meaningful discussion or comparison to this existing work. As a result, the contribution is incremental and the claims of being the "first" to propose such a foundation model are not justified.
+
+### Strengths
+- The paper addresses a real and important challenge in crowdsourced label aggregation: the need for scalable, retraining-free aggregation methods that generalize across datasets.
+- The paper provides a synthetic data generation process for training such models.
+- CrowdFM achieves fast inference times, comparable to simple methods like Majority Voting, while outperforming more complex dataset-specific models.
+
+### Weaknesses
+1.  Lack of proper acknowledgement of closely related work. The most significant issue with this paper is its failure to properly acknowledge and position itself with respect to existing work that is nearly identical in methodology and motivation. In particular, the ICLR 2023 paper[1] that proposes a hyper label model for programmatic weak supervision that is, in essence, the same as the approach in this paper:
+  - Both solve the label aggregation problem. 
+  - Both aim for cross-dataset generalization and retraining-free inference.
+  - Both papers propose a GNN-based model that is pretrained on synthetic data to learn a generalizable label aggregation function.
+  - Both use size-invariant initializations and GNN architectures to handle variable numbers of annotators (workers/LFs) and tasks.
+  - Both demonstrate that their model, once trained, can be applied to new datasets in a single forward pass, outperforming dataset-specific methods in both accuracy and efficiency.
+
+While the paper claims to be the "first foundation model for label aggregation," this is not accurate. The hyper label model[1] is a direct precedent, and the technical similarities are substantial. The only notable difference is the application domain: The hyper label model[1] focuses on programmatic weak supervision (labeling functions), while this paper focuses on crowdsourcing (human workers). However, the mathematical setup is essentially the same (a bipartite label matrix with noisy annotators) in these two domains.
+The lack of discussion, comparison, or even citation of this highly relevant prior work is a serious omission. This gives the misleading impression that the proposed approach is novel, when in fact it is a straightforward adaptation of an existing method to a closely related domain.
+
+2. The paper repeatedly claims to be the "first" to propose a foundation model for label aggregation, and to introduce a new paradigm. Given the existence of[1], these claims are overstated.
+
+3.  Given the existence of the hyper label model approach[1], the main contribution of the paper seems to be modifications to the synthetic data generator and GNN architecture.
+
+[1] Wu et al., "Learning Hyper Label Model for Programmatic Weak Supervision" (ICLR 2023)
+
+### Questions
+1. Are the authors aware of the hyper label model[1]? Can the authors clarify the technical and conceptual differences between CrowdFM and the hyper label model, beyond the application domain? Please explicitly discuss this prior work in the related work section, and provide a detailed comparison (both conceptual and empirical) to clarify what is novel in CrowdFM.
+2. How does CrowdFM perform compared to the hyper label model?
+3. What new insights does CrowdFM provide for the crowdsourcing setting on top of what is discussed in[1].
+4. Is CrowdFM able to provide any theoretical guarantees as in[1].
+
+[1] Wu et al., "Learning Hyper Label Model for Programmatic Weak Supervision" (ICLR 2023)
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes CrowdFM, the first foundation model for crowdsourced label aggregation. It replaces traditional per-dataset parameter estimation with a pretrained bipartite graph neural network trained on a vast, domain-randomized synthetic dataset. By leveraging a size-invariant initialization and attention-based message passing, it learns universal principles of collective intelligence and generalizes to new, unseen datasets. Experiments on 22 real-world datasets show that CrowdFM matches or outperforms state-of-the-art methods in both accuracy and efficiency, while its learned representations also support worker assessment and task assignment.
+
+### Strengths
+1.The proposed CrowdFM is the first foundation model for label aggregation, enabling cross-dataset generalization without retraining and marking a key step toward universal label aggregation.
+
+2.CrowdFM further exhibits strong versatility, as its pretrained representations can be directly adapted to multiple downstream tasks such as worker assessment and task assignment without additional training.
+
+3.Extensive experiments on 22 real-world benchmarks demonstrate that the proposed single, fixed model consistently matches or surpasses dataset-specific methods in both accuracy and efficiency.
+
+### Weaknesses
+1.The paper lacks a systematic quantitative analysis of synthetic and real-world crowdsourced data. Although the authors validate generalization on real datasets, they do not analyze the distributional differences of key parameters (e.g., worker ability θ, task difficulty β, task guessing rate c) between synthetic and real data, making it difficult to assess how well the learned patterns reflect real human annotation behaviors. Besides, how to obtain these sampling ranges?
+
+2.Although the parameters in the synthetic data generator are randomly sampled, the paper lacks a sensitivity analysis of key parameters. Random sampling alone cannot fully verify the model’s robustness across different crowdsourcing conditions. It is recommended to include such an analysis to strengthen the credibility of CrowdFM’s generalization claims.
+
+3.With regard to the comparison results, statistical tests are needed in the comparison results. The detailed description about statistical tests for comparisons of multiple algorithms on multiple datasets can be found from the papers such as Statistical comparisons of classifiers over multiple data sets.
+
+4.The paper lacks ablation studies to verify the individual contributions of key components (e.g., attention-based aggregation, size-invariant initialization, and synthetic data diversity). Without such analysis, it is difficult to determine which design choices are most critical to the model’s performance gains.
+
+### Questions
+The same as the weaknesses above.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+Traditional methods (e.g., Dawid–Skene, GLAD, EBCC) typically adopt a dataset-specific modeling paradigm: they require training model parameters from scratch for each new dataset, resulting in poor generalization. Meanwhile, Majority Voting (MV), widely used in industry, requires no training but ignores differences in annotators' capabilities, leading to limited accuracy.
+To bridge this gap, the authors propose a universal aggregation paradigm: by pre-training a bipartite GNN on large-scale, domain-randomized synthetic datasets, the model learns transferable "collective intelligence" aggregation rules. During the inference phase, the model can be directly applied to any real crowdsourced dataset without fine-tuning or retraining.
+
+### Strengths
+1.	Strong paradigm innovation: For the first time, the foundation model concept was introduced into crowdsourced label aggregation, breaking the dataset-specific paradigm that has persisted for decades.
+2.	The problem definition has practical significance: Shifting the aggregation of crowdsourced labels from the "dataset-by-dataset modeling" paradigm to the "unified fundamental model" paradigm aligns with the urgent demand of the industrial sector for scalable and retraining-free systems.
+3.	The experiments were thorough: 22 real datasets covered text, images, and audio, and included extreme scale/density scenarios, verifying the generalization ability.
+4.	The accuracy rate and running time were reported, and it was pointed out that some methods failed due to insufficient memory
+
+### Weaknesses
+1.	How can we ensure that the distribution of synthetic data is sufficient to cover the feature space of real crowdsourcing tasks? Is there any performance degradation under certain types of tasks (such as extremely unbalanced ones)?
+2.	Can CrowdFM be regarded as the specialization of GraphFM in the field of crowdsourcing? Has the pre-training objective of the existing GFM been borrowed?
+3.	The performance for extremely large-scale data (such as Senti and Fact) slightly decreases, but the reasons are not discussed.
+4.	The absence of ablation experiments makes the contributions of each module unclear
+5.	Although it includes recent works such as EBCC, GOVERN, TiReMGE, etc., some new methods for 2024-2025 are missing (such as Zhang et al., KFNN and IWBVT of NeurIPS 2024 are cited but not used as baselines);
+6.	The superparameters such as the number of layers L and dimension d of GNN were not discussed in detail.
+7.	The "Sim-to-Real Gap" between synthetic data and real data has not been fully quantified
+8.	Has the model truly learned the "principle of collective intelligence", or has it merely fitted the statistical patterns of synthetic data? How to prove.
+
+### Questions
+Same as weaknesses
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper aims at building a foundation model for crowdsourced label aggregation that generalizes across heterogeneous datasets without dataset-specific training. It first generates a synthetic crowdsourced data generator to produce diverse synthetic datasets, and then trains a bipartite graph neural network which can be generalizable to other data and tasks. Experiments show the foundation model is efficient and achieve performance comparable to or superior to state-of-the-art aggregation methods.
+
+### Strengths
+1. Generating the synthetic data and then train a generalizable GNN model to approximate it is a smart idea to unify the data-specific crowdsourcing models. I think it is essentially a type of knowledge distillation, which distills the knowledge from the data generator to a GNN with less input features (e.g. task difficulty/worker ability, which are generally regarded as latent parameters in traditional probabilistic models). 
+2. The method is sound, and the presentation is clear.
+3. Experiments show good accuracy with more efficiency.
+
+### Weaknesses
+1. The backbone model is relatively simple and not new, but the design is a good fit for crowdsourcing, e.g. the size-invariant initialization for worker/task emebddings. I do not see it as a weakness, but just not a novel contribution. I still like the whole framework of building synthetic data and train a simple domain-indepedent model as foundation models.
+
+2. The possible limitation of the model is it may lack the flexibility to adapt to the case when workers/tasks have their attributes, since these nodes are initialized with fixed pretrained embeddings.
+
+3. Another possible weakness is the limitation of the data generator. Although it is a quite general scheme, there might be other graphical models that it cannot cover; and also the prior distribution of the parameters might cause some bias for the pretrained foundation model.
+
+### Questions
+See weaknesses.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+3

human_reviews/GE0b5aQAvm.md

@@ -0,0 +1,118 @@
+## Human Reviewer 1
+
+### Summary
+This paper studies the suboptimality of nonlinear and ensemble policies in continuous control problems. The authors formulate their analysis in the context of linear systems with quadratic costs and compare the performance of nonlinear neural policies and ensembles of neural policies against linear policies and linear ensembles. They define a nonlinearity measure quantifying the deviation of a function from being affine and show that, for linear-quadratic systems, any nonlinear (or non-affine) policy must be suboptimal relative to the optimal linear policy. Theorem 1 and Theorem 2 establish that nonlinear policies cannot be globally optimal, and Theorem 3 extends this argument to ensemble policies, suggesting that non-convex mixtures of policies are also suboptimal compared to convex-mixing policy ensembles. Empirical evaluations are conducted on synthetic linear control systems, comparing the performance of linear, nonlinear, and ensemble policies. The results show that neural policy ensembles perform worse than linear policies and ensembles.
+
+### Strengths
+- The paper tackles a clear and well-posed theoretical question: under what conditions are neural or nonlinear ensembles suboptimal compared to linear policies? This is question, I believe, is of theoretical interest.
+- The mathematical results are formally presented with explicit assumptions and definitions (e.g., the "nonlinearity measure" and the LQR cosntraints). The formalism appears internally consistent and logically valid given the assumptions.
+- The paper attempts to connect classical control theory insights (LQR optimality) with modern neural network and ensemble formulations, a potentially valuable contribution.
+
+### Weaknesses
+**Presentation**:
+
+- I found the paper extremely difficult to parse, both linguistically and conceptually. The writing is often ambiguous, and key terms are not defined well. For example, the opening statement of the abstract ("neural policy ensembles are sub-optimal with respect to linear policy ensembles") is already unclear and leaves the reader guessing what kind of optimality this is referring to. 
+- The central theoretical claim almost seems self-referential given the assumptions. Since the problem setup is fully linear with quadratic costs, and the class of linear policies already contains the global optimum, it follows by construction that any non-linear function cannot be better. If I understood correctly, this result is highly unsurprising and, arguably, trivial within the stated framework of linear problems with quadratic costs. 
+- The authors seem to assume that neural policies are nonlinear by construction and therefore suboptimal. However, this neglects the fact that, by the universal approximation theorem, a neural network with ReLU activations can represent a linear function exactly. In other words, a neural policy can approximate the optimal linear controller arbitrarily well. The suboptimality is in my view a statement about nonlinear functions in general, not necessarily about neural policies.
+- In the treatment of ensembles, the authors reference, for example, Burke et al. for non-convex mixing with neural networks, but these works consider state-dependent mixtures. In the author's notation, the weights $w_i$ seem to be static rather than state-dependent. As before, the authors exclude the probability simples from the support of non-convex mixing by construction.
+
+### Questions
+- What motivates the focus on linear dynamics and quadratic costs for the theoretical exposition? It seems clear that LQRs solve these optimally. 
+- In Theorem 3, how are mixture weights $w_i$ defined as statis scalars or as state-dependent functions? The notation suggests static weights, but the text refers to neural mixing networks.
+- How does this result translate to systems that are not linear?
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper presents a theoretical study that argues that non-linear neural policy ensembles are sub-optimal and less stable compared to linear policy ensembles in Linear Quadratic Systems in optimal control. It also presents a theoretical result that shows non-convex policy mixing is sup-optimal to convex mixing of policies. The theoretical results are verified in a control environments which demonstrate that even well-tuned neural ensembles can underperform by 2 orders of magnitude compared to equivalent linear ensembles in Linear Quadratic Systems.  
+
+
+**Recommendation:**\
+The content of this paper falls quite far outside my area of expertise, as I am an expert in neural ensembles in deep reinforcement learning, not in control systems. However, from my perspective, I recommend to reject this paper, as it seems to lack in the following areas: overly general claims and statements, and a general lack of clarity.
+
+### Strengths
+- The question of whether neural policy ensembles exhibit the same benefits as previously found for ensemble classifiers is very relevant. 
+- The theoretical results are interesting.
+
+### Weaknesses
+- The paper overclaims its contributions. The abstract and introduction have no mention of the assumptions made in both the theoretical and empirical studies. 
+	- For example, the abstract states: _"We develop a theoretical framework to formally prove that (non-linear) neural policy ensembles are sub-optimal with respect to linear policy ensembles."_ However, from different parts of the text I have pieced together that this results only holds for: a Lipschitz continuous transition function, deterministic polices, in a linear quadratic system, and doesn't hold for neural policies in general, but instead for policies with an additional assumption of a minimum level of non-stationarity of each individual ensemble member (neural networks are universal function approximators, so they can also learn linear policies). Furthermore, the result seems to only guarantee the existence of one state for which the non-linear neural ensemble is suboptimal, which might not be encountered by the controller from the given starting state, which would make it irrelevant for its performance.
+- The theoretical results are difficult to follow. The theorems are stated rather clinically, and are lacking in clear structure regarding assumptions, how they are proven, and what they imply.
+
+### Questions
+- The paper claims that (non-linear) neural policy ensembles are inherently suboptimal compared to the individual policies. How do you square this claim with the studies that show improved performance of policy ensembles compared to single policies in RL (some examples, [1,2,3,4])?
+- Line 39: _"In contrast, nonlinear policy ensembles face temporal coupling: the ensemble’s actions affect future states, creating feedback loops that may amplify rather than cancel errors. The temporal dependence breaks the mathematical foundation of ensemble methods."_
+	- Which part of the results or experiments in this paper explain or verify this intuition?
+- Section 3.3: There seems to be a one-to-one correspondence implied between neural ensembles and non-convex mixing. Why could a neural network not produce a convex mixing function? 
+
+
+**References:**\
+[1] SUNRISE: A Simple Unified Framework for Ensemble Learning in Deep Reinforcement Learning. (Lee et al. 2021)\
+[2] Swarm Behavior Cloning. (Nüsslein et al. 2024)\
+[3] Why Generalization in RL is Difficult: Epistemic POMDPs and Implicit Partial Observability. (Ghosh et al. 2021)\
+[4] How Ensembles of Distilled Policies Improve Generalisation in Reinforcement Learning. (Weltevrede et al. 2025)
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper studies the ensembling of policies where each sub-policy is solving a different LQR problem. The authors provide several results that highlight the fact that ensembling of non-linear policies might be sub-optimal and unstable. Specifically, the authors show that given a set of tasks (defined by different reward functions), ensembling non-linear controllers learned on each task is worse than ensembling the best linear controllers for each task, that the resulting ensemble of non-linear controller might be unstable even when each individual controller is stable and that a non-convex averaging of the linear controllers is always worse than the convex averaging of linear controller with the same weight as the weight averaging the reward functions. The authors then provide a set of experiments that illustrate each of their theoretical findings.
+
+### Strengths
+The paper might provide interesting theoretical results on ensembling of non-linear policies for LQR problems.
+
+### Weaknesses
+- There is a clear mismatch between the claims and the theoretical results. The introduction and related work cite deep RL methods using ensembling of policies (notably the work of Yang et al. 2022) in a completely different way than what the theoretical results consider. Introduction mentions that their theory would explain why ensembling of neural policies is doomed to be sub-optimal because of temporal coupling and non i.i.d setting compared to standard classification but the theoretical results show none of that. Instead all the theory is built around ensembling non-linear policies that are solving distinct control problems and then averaging them with no further learning. However, taking a quick look at the paper of Yang et al. 2022, there are large difference in the setting, as in the latter work every sub-policy is solving the same task, and the ensemble policy is also optimized using what they call an ensemble aware loss. Can the author clarify the relation between their theoretical results and the setting of Yang et al. 2022, or any other deep RL setting using ensembles of neural policies? 
+
+-  Theorem statements, assumptions and proofs lack clarity. In the main theorem, it is not clear how the non-linear policies are obtained. The optimal policy is linear for each LQR sub-problem, but by the assumption in line 140, the neural policies are forced to be non-linear and thus these policies are sub-optimal by design. It is then not surprising then that the ensembling of these policies is also sub-optimal. Studying problems where linear policies are optimal and making the assumption that neural policies cannot be optimal is in my opinion not a setting reflective of practical applications that use neural policies. Moreover, the proof could also gain some clarity: where do equation 21 and 22 come from? What even is $R$ in these equations, do you perhaps mean $R_{\text{ens}}$? Is $\lambda_{\min}$ the smallest eigenvalue? For the proof of Theorem 3, could you please provide a reference to the claim that $K_\lambda$ is optimal? As it does not seem trivial to me from the equations since there doesn't seem to be a linear relation between gain matrices and cost matrices. Thank you.
+
+### Questions
+Please see the questions in the weakness section.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+2

human_reviews/H1s6SBhzlg.md

@@ -0,0 +1,178 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes two methods—Optimal Weight (OW) and Inverse Surprisingly Popular (ISP)—to improve aggregation of answers from multiple LLMs beyond simple majority voting. OW leverages first-order information (model accuracies) and is provably Bayesian-optimal when accuracies are known. ISP uses second-order information (prediction correlations) without requiring ground-truth labels and outperforms both majority voting and the classic Surprisingly Popular method.
+
+I have some concerns about the paper at this stage, so my evaluation may be biased towards the negative. If the author can address these issues in the rebuttal stage, I will consider raising my score.
+
+### Strengths
+1. The concepts proposed are novel.
+
+2. Achieved good performance on multiple datasets
+
+### Weaknesses
+1. Related work is an important part of the paper and should not be put in the appendix.
+
+2.  Why is the LLM's output unaffected by the order of the options? Is there any literature or experiment demonstrating this? Since different letters have different statistical probabilities, the LLM should have different preferences for different letters.
+
+3. Why didn't the authors conduct experiments on widely used mathematical and common sense reasoning datasets, such as GSM 8K? This might be more helpful in verifying the effectiveness of the proposed method.
+
+4. I'm sorry that I can't find any relevant literature, but in my impression, optimal weight aggregation should not be the first one proposed in this paper. It has been proposed in other works before.
+
+### Questions
+1. Are the concepts of first-order information and second-order information first defined in this article?
+
+2. How to obtain the expected accuracy？
+
+3. The sizes of the several models selected in the experiment are obviously unbalanced. GPT-4o is much larger than the other models. Will this affect the experimental results?
+
+### Soundness
+3
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper investigates how to aggregate answers from multiple Large Language Model (LLM) agents to overcome the limitations of standard Majority Voting (MV). The authors note that MV ignores heterogeneity and correlation among models. To address this, the paper proposes two new algorithms that leverage higher-order information, namely Optimal Weight (OW) and Inverse Surprising Popularity (ISP). Since OW only need unknown true accuracies, the authors further propose two practical unsupervised methods (OW-L and OW-I). The experimental section validates these methods on synthetic datasets, UltraFeedback, MMLU and ARMMAN.
+
+### Strengths
+How to effectively aggregate the response from several LLMs is important tasks. This paper goes beyond simple heuristics, formally modeling the problem from an information-theoretic perspective is interesting.
+
+### Weaknesses
+1. The paper describes MoE as "only one expert gets a non-zero weight", which is too simplification. Modern MoE combines outputs from multiple experts via weighted averaging. The authors should more accurately describe the distinction from their work (e.g., aggregating final answers vs. internal representations).
+2. The authors should explain why in Table 3, the 'Single Best' model on the MMLU dataset outperforms the proposed aggregation methods.
+3. The proposed method optimizes the aggregator on several large datasets. But does it generalize effectively to other, unseen datasets?
+Questions: Please refer to weakness.
+
+### Questions
+Theorem 3 shows that the dataset size (M) needs to be large to reduce the error. However, considering that N LLMs are required, is the cost of this algorithm still prohibitively high?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper studies the response aggregation problem of multi-agent large language models (LLMs), arguing that simple majority voting ignores the heterogeneity and correlation between models. The authors propose two new algorithms: (1) Optimal Weight, which uses first-order information, i.e., the accuracy of each LLM, for Bayesian optimal weighted aggregation; and (2) Inverse Surprising Popularity, which uses second-order information, i.e., the correlation between the answers provided by each agent, to improve the traditional surprising popularity method. Theoretical analysis proves the Bayesian optimality of OW and the advantage of ISP over MV. Experiments validate the methods on synthetic data, UltraFeedback, MMLU, and ARMMAN datasets, showing an improvement of 0.5-2 percentage points compared to MV. However, the problem studied in this paper suffers from insufficient importance, and the practical value of the methods is quite limited.
+
+### Strengths
+S1. The core theoretical results of the paper, including the Bayesian optimality of OW (Theorem 1) and the expected advantage of ISP over MV (Theorem 2), are mathematically rigorous.
+
+S2. The authors found that the traditional surprising popularity method performed worse than simple MV in LLM scenarios. The authors' explanation makes sense: SP is designed to correct systematic biases in human judgment, but LLM has less systematic bias, so SP can only take advantage of a smaller margin of error.
+
+S3. The experimental design of this paper is comprehensive, including three levels: (1) synthetic data that fully conforms to the theoretical assumptions to verify the theoretical predictions; (2) standard LLM benchmarks to verify practicality; and (3) real-world medical and health applications (ARMMAN) to demonstrate potential social value. The authors also used eight models from four families including GPT, Qwen, Llama, and Phi to form 16 combinations that cover models of different scales and capabilities.
+
+S4. Appendix B.2’s relaxation of the conditional independence assumption is a positive attempt.
+
+### Weaknesses
+W1. The fundamental problem is that the research question itself is not important enough, and its practical value is limited. Specifically, the paper devotes a large portion to studying how to optimally aggregate votes from multiple LLMs, but this problem has very limited importance in the current LLM application ecosystem. 
+
+First, the closed-set classification (K fixed options) that the paper focuses on is only a small part of LLM applications; the more mainstream applications are open-ended generation, multi-turn dialogue, and complex reasoning, and the paper's method cannot be extended to these scenarios. 
+
+Second, the improvement shown in experiments is extremely small (0.5-2 percentage points), not worth increasing the system complexity. 
+
+Third, in practice, there are often simpler and more effective solutions like directly use the strongest single model (Table 3 shows that a single model is sometimes better than ensemble), or invest in training a better model, rather than ensemble multiple weak models. 
+
+Fourth, the core issue in multi-agent LLM reasoning is how to design an effective debate protocol and how to enable models to truly exchange and integrate information and the final voting is actually the most trivial part. The paper is essentially optimizing a marginal problem.
+
+W2. The OW algorithm is a beautiful theoretical contribution, proving its Bayesian optimality. However, this method relies on knowing the accuracy of each LLM, and in the unsupervised aggregation scenario considered in the paper, obtaining these accuracies requires a large amount of labeled data. Since it is unsupervised, how do we obtain this data?
+
+W3. Theorem 2 shows that the advantage of ISP decays at a rate of 1/K, meaning that for tasks with many options or open-ended functions, ISP offers virtually no advantage. The paper acknowledges this limitation in its conclusion but does not propose a solution. This may further restricts the applicability of the method.
+
+W4. The paper does not adequately discuss the limitations and applicability of the method.
+
+### Questions
+Q1. The paper focuses on the final aggregation step, but is this really the most important bottleneck in multi-agent LLM reasoning?
+
+Q2. Table 3 shows that the single strongest model achieves 91.02% on MMLU, while the ensemble only achieves 90.37%. This seems to suggest that the value of ensemble is questionable. Could you provide a cost-benefit analysis comparing `ensemble of 4 intermediate models` vs. `using only one strongest model` vs. `using the budget of ensemble to train/fine-tune a better single model`?
+
+Q3. How can these methods be extended to open-ended tasks?
+
+Q4. What are the theoretical guarantees of the OW-L and OW-I methods?
+
+Q5. ISP requires estimating O(N²K²) conditional probabilities, which becomes a bottleneck when N and K are large. How can this be scaled?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This study addresses the fundamental challenge of **aggregating responses from multiple Large Language Models (LLMs)** to produce a single, reliable collective decision, aiming to significantly improve upon simple **majority voting (MV)** or reliance on a single LLM.
+
+
+The paper proposes two key aggregation schemes:
+- The **Optimal Weight (OW)** algorithm generates the most accurate collective decision by assigning a **vote weight ($\omega_i$)** to each LLM $i$ based on its **known expected accuracy** ($x_i$).
+- A practical limitation of OW is that the true expected accuracies ($x_i$) are often **unavailable in real-world scenarios**. To overcome this, the study introduces the **Inverse Surprising Popularity (ISP)** algorithm. ISP bypasses the need for ground-truth labels by leveraging **predictions correlations** among LLMs to infer the correct prediction. ISP functions as an aggregation method in its own right and, critically, provides a mechanism to **estimate the optimal weights needed for OW** in label-free settings.
+
+
+The algorithms were validated across simulated datasets, LLM benchmarks (UltraFeedback, MMLU), and a real-world healthcare dataset (ARMMAN). Both **OW and ISP consistently outperformed MV**. Ultimately, the best results were achieved by a heuristic combination of both OW and ISP approaches, which demonstrated **strictly higher accuracy than the single best individual LLM** on several real-world tasks.
+
+### Strengths
+- The design of the ISP aggregation scheme proposed by the study offers a practical, cost-effective solution to multi-agent LLM aggregation schemes because it relies only on correlations between LLM predictions. This capability makes ISP and related heuristics (OW-L, OW-I) directly applicable to unsupervised, real-world scenarios like automated data annotation.
+
+### Weaknesses
+- One of the assumptions of this study relies on the idea that we can model the difficulty of a prompt question given to an LLM using **α** (question difficulty) and **βᵢ** (model ability). However, these quantities are inherently **subjective** and cannot be reliably measured. This raises concerns about the realism and validity of the assumptions on which this work is based.
+
+- The study’s **comparative evaluation** is also limited, as it primarily benchmarks the proposed algorithms against **MV**, a relatively simple baseline. More advanced multi-agent reasoning strategies such as the ones described in the related work section (see Appendix A.1) were not included in the comparison, constraining the empirical scope of the findings.
+
+- Additionally, the paper **does not discuss statistical significance** or provide detailed information about its experimental setup. Key aspects such as, variance, and number of seeds used are not reported, making it difficult to assess the robustness and reproducibility of the results. Including these details would strengthen the credibility of the findings and provide clearer insight into the reliability of the proposed methods.
+
+### Questions
+- Why do OW‑L and OW‑I show nearly identical performance in the experiments (Tables 3–6)? Specifically, is there a correlation between the objective optimized in OW‑L and the metrics underpinning ISP, which OW‑I use to produce the weights?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4

human_reviews/HA8vzzT6Ax.md

@@ -0,0 +1,288 @@
+## Human Reviewer 1
+
+### Summary
+This paper revisits the trade-off between watermark strength and speculative sampling efficiency in large language models (LLMs). The authors argue that the incompatibility stems from the use of independent random sources: watermarking modifies token sampling using pseudorandomness, while speculative sampling introduces additional stochasticity through draft-token acceptance. This mismatch weakens watermark detectability and limits inference efficiency. To resolve this, the paper proposes a unified pseudorandom framework that decomposes random sources into three independent components: \(ζ_D\) (for the draft model), \(ζ_T\) (for the target model), and \(ζ_R\) (for acceptance decisions)
+
+### Strengths
+- The paper provides a clear theoretical explanation for why speculative sampling and watermarking interfere with each other—namely, their reliance on separate random sources.  
+- The proposed pseudorandom acceptance mechanism is simple yet elegant, turning the generation process into a deterministic function of pseudorandom variables.  
+- Theoretical analysis (Pareto frontier characterization) and empirical evidence consistently demonstrate that the proposed design breaks the previously assumed trade-off between watermark strength and efficiency.  
+- The work unifies two practically important but previously conflicting mechanisms, offering a potential path toward efficient and traceable LLM inference
+
+### Weaknesses
+1. The experimental evaluation only considers two unbiased watermarking schemes (Gumbel-max and SynthID). The analysis would be more convincing if extended to other representative watermarking frameworks, such as KGW or reweight-based methods [1,2].  
+
+2. The proposed method requires applying watermarking to both the draft and target models during speculative sampling. The paper does not analyze whether this dual watermarking might affect text quality or fluency.  
+
+[1] Hu, Zhengmian, Lichang Chen, Xidong Wu, Yihan Wu, Hongyang Zhang, and Heng Huang. *Unbiased watermark for large language models.* arXiv preprint arXiv:2310.10669 (2023).  
+[2] Wu, Y., Hu, Z., Guo, J., Zhang, H., & Huang, H. *A resilient and accessible distribution-preserving watermark for large language models.* arXiv preprint arXiv:2310.07710 (2023).
+
+### Questions
+1. Could the authors discuss whether the pseudorandom control framework can be generalized to biased or reweighted watermark schemes?  
+
+2. Considering that speculative sampling is primarily designed to accelerate inference, the paper should report additional runtime measurements comparing (a) standard speculative sampling, (b) speculative sampling with watermarking, and (c) the proposed method, to confirm that efficiency gains are retained.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper revisits the trade-off between watermark strength and speculative sampling efficiency in large language models. It formally quantifies watermark strength using KL divergence and derives a theoretical upper bound equal to the entropy of the base model. To overcome the claimed inevitable trade-off from prior work, the authors propose a pseudorandom acceptance mechanism that makes the entire sampling process a deterministic function of pseudorandom seeds. They further introduce new detection methods (Ars-$\tau$ and Bayes-MLP) leveraging this mechanism, and demonstrate through experiments that the approach improves detection performance without reducing decoding throughput.
+
+### Strengths
+The paper provides a theoretical framework connecting watermark strength and speculative-sampling efficiency. The proposed pseudorandom-acceptance mechanism is theoretically appealing, showing that the trade-off can be alleviated under certain conditions.
+
+### Weaknesses
+1. The proof that watermark strength can reach the upper bound $WS = H(P)$ (Theorem 3.2) assumes that each $P_\zeta$ is a completely deterministic distribution, where every generated token is fixed once the pseudorandom seed is given. However, even SynthID’s tournament sampling still has randomness after top-$k$ and $2^m$ rounds, so $P_\zeta$ isn’t truly deterministic. This means the bound is more of a theoretical ideal than something achievable in real LLM decoding. The paper should make this distinction clearer and discuss how far practical schemes are from that ideal. For example, how much does it improve over different m (the number of layers used in SynthID)?
+
+2. The claim of achieving both the best sampling efficiency ($SSE = 1 - \mathrm{TV}(Q, P)$) and the highest watermark strength ($WS = H(P)$) relies on several strong assumptions, such as perfectly unbiased decoders, fully deterministic generation, and independent pseudorandom seeds. If any of these assumptions don’t hold, the trade-off comes back. So should the result really be interpreted as showing what’s possible under idealized conditions, not as a general solution that always works?
+
+### Questions
+see weaknesses
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper revisits the trade-off between watermarking and speculative sampling. It (i) quantifies watermark strength as the expected KL divergence, showing this metric controls the asymptotic p‑value decay rate of the optimal detector (ii) formalizes the strength–efficiency Pareto frontier (iii) proposes pseudorandom acceptance by introducing an extra seed $\zeta_R$ so the accept/reject step is itself a deterministic function of pseudorandomness. Empirically, the method matches standard speculative sampling in AATPS while improving TPR@FPR=1% for detection compared with previous methods that preserve AATPS.
+
+### Strengths
+1. New perspective (not detection‑first, measure strength via KL).
+The paper propose to cleanly separates watermark strength from detector performance, grounding the former in an information‑theoretic quantity (expected KL) with a clear sample‑complexity interpretation (Theorem 3.1) and a tight entropy upper bound (Theorem 3.2). This presents a principled framework for further analysis.
+
+2. Evidence of better trade‑offs in practice.
+AATPS closely tracks the standard baseline, while detection improves compared with previous methods that preserve AATPS.
+
+3. Clear writing and organization.
+The paper is easy to follow, with crisp statements, clear algorithm, and figures that directly support claims.
+
+### Weaknesses
+### Whether the “trade‑off” truly disappears (terminology/interpretation).
+
+Theorem 4.1. builds on redefining strength as expected KL. This metrics makes some new proposed analysis possible. However, the paper itself clarifies that watermark strength ≠ detection efficiency (Remark 3.1): two schemes with comparable watermark strength may still differ in detection efficiency.
+
+Regarding specific methods, mixture of multiple pseudorandom sources will hurt detection efficiency, as more pseudorandom sources means more ambiguity at detection time. (Hu & Huang, 2024) uses one pseudorandom seed. (Dathathri et al., 2024) uses two seeds (draft + target), with multi-round tournaments there may be effectively more. This paper uses three seeds, further compounding interference among signals at detection time. Concretely, for any observed token the detector does not know whether it came from (i) an accepted draft token, (ii) a rejected token replaced from the residual, or (iii) the “extra” token when all draft tokens were accepted. This uncertainty necessarily makes practical detection weaker than the paper’s “WS” metric.
+
+In Experiment section of this paper, the claim is also “improved detectability,” not matching with non-accelerated watermarking.
+
+Therefore, the “trade‑off” still exists. The paper pushes the practical frontier but does not remove it, the headline “BREAK THE TRADE-OFF” is overstated. A more faithful summary is that the method steadily approaches a better trade-off.
+
+### Minors:
+
+The paper, for simplicity, treats positions as independent. A stricter analysis should avoid this assumption, and some results can hold without independence.
+
+Seed reuse is not discussed. Re-applying the same pseudorandom seed should be skipped (e.g., as in prior “context code history”) to maintain unbiasedness.
+
+### Questions
+I would encourage authors to also report ANLPPT, as with AATPS for efficiency, to quantify the detectability gap relative to a non-speculative watermarked baseline.
+How large is the remaining gap in detectable watermark signal strength?
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+## Summary
+This paper tackles a practical tension in LLM serving: **unbiased text watermarking** (for provenance) versus **speculative sampling** (for speed). Prior work framed this as a binary impossibility: you can either preserve the watermarked distribution or preserve the acceptance rate, not both. The authors (i) introduce a **quantitative notion of watermark strength** — the expected KL divergence $E_\zeta[D_{KL}(P_\zeta || P)]$ — and prove it is the **exact detection rate** (via Chernoff–Stein); (ii) cast watermark–efficiency trade-offs as a **Pareto optimization** and empirically map the curve for several watermark classes (Gumbel-max, SynthID, linear mixtures, prior classes); and (iii) propose a simple fix — **pseudorandomizing the accept/reject coin** in speculative sampling — so the whole decode becomes a deterministic function of seeds, achieving **maximum sampling efficiency** $1 - TV(Q,P)$ and **maximum watermark strength** $H(P)$ simultaneously. They also give practical detectors for speculative pipelines (Ars-$\tau$ / Ars-Prior and Bayes-MLP / Bayes-Prior).
+
+Production stacks already depend on speculative sampling; having a *measurable* frontier and a concrete recipe to sit on that frontier is valuable.
+
+## Main claims / contributions
+1. **Strength as information:** Define watermark strength as $E_\zeta[D_{KL}(P_\zeta || P)]$ and prove it controls the optimal detector’s p-value decay (through the per-token log-likelihood ratio).  
+2. **Tight frontier & optima:** For unbiased schemes, strength is upper-bounded by $H(P)$; speculative acceptance is upper-bounded by $1 - TV(Q,P)$. Gumbel-max and SynthID (as $m \to \infty$) are strength-optimal; the proposed **seeded acceptance** achieves *both* bounds simultaneously.  
+3. **Operational trade-offs & detectors:** A constrained program (Eq. 10) traces the **Pareto curve** $L(r)$. The paper designs detectors that account for draft/target paths; when the accept coin is known (Ars-$\tau$, Bayes-MLP), detection is stronger than prior averaging approaches (Ars-Prior, Bayes-Prior).
+
+### Strengths
+## Strengths
+- **Conceptual clarity:** Moving from a binary notion to a continuous, information-theoretic strength is the right abstraction. It cleanly explains earlier "impossibility" statements and yields a usable frontier.  
+- **Tight theory that lines up with practice:** The Chernoff–Stein interpretation makes sample-complexity predictions directly actionable; the entropy and TV bounds give interpretable ceilings.  
+- **Simple, impactful algorithmic tweak:** Seeding the accept coin is minimal and compatible with production speculative pipelines.  
+- **Detector design under speculative sampling:** The draft/target path issue is real; the two Ars and two Bayes variants are a neat, pragmatic treatment.  
+- **Figures are helpful:** The trade-off plots communicate the Pareto picture at a glance; markers (WS/SE maintained, optimum) guide interpretation.
+
+### Weaknesses
+## Weaknesses / limitations
+- **Oracle knowledge vs deployment reality:** Many results assume access to the base distribution $P_t$ (or accurate logits) and, for the strongest detectors, the **accept coin**. External forensics often lacks both; estimation error can materially affect strength and false positives.  
+- **Model/scale realism:** Simulated $(Q,P)$ pairs and small-vocab settings may not capture calibration quirks, long-tail tokens, or beam-blocking in large models.  
+- **Robustness / attack model:** The paper doesn’t stress-test adversaries who (i) alter spacing/punctuation, (ii) paraphrase with another model, or (iii) run ensemble-style removal — how quickly does the effective KL collapse?  
+- **Broader behavioral impacts not discussed:** Recent studies suggest watermarking can alter model behavior beyond detectability or quality trade-offs. *Downstream Trade-offs of a Family of Text Watermarks* (Ajith et al., EMNLP 2024) reports 10–20% drops in downstream task accuracy even for “unbiased” schemes like KGW, while *WaterJudge* (Molenda et al., NAACL 2024) quantifies a detectability–quality trade-off across watermark strengths. *Watermarking Degrades Alignment in Language Models* (Verma et al., ICLR 2025) further shows that watermarking can shift alignment properties such as truthfulness, safety, and helpfulness, mitigated partly by an external reward model. The paper would be stronger with a targeted discussion of how speculative sampling with watermarking affects alignment-relevant metrics (e.g., reward scores) and a short discussion citing these works to contextualize potential side effects.  
+- **Downstream behavior:** Stronger coupling can interact with alignment/quality. Recent work reports reward/quality/alignment degradations under watermarking; the paper doesn’t quantify the proposed method’s effect on reward score or human preference.  
+- **Reporting details:** Some derivations (e.g., Eq. 10) are explained tersely; detector training details and calibration of $\tau$ / $p$ could be easier to reproduce.
+
+## Actionable suggestions
+1. **Adversarial/transform robustness:** Evaluate detectability after common edits (paraphrase via another model, synonym swaps, punctuation/whitespace jitter, mild re-ordering). Report the induced drop in effective strength and how close it is to theoretical predictions.  
+2. **Latency & throughput accounting:** Since the goal is "fast **and** provable," include an end-to-end latency/throughput table under seeded acceptance (overhead of PRNG, acceptance calibration, detector extraction) to confirm the practical speedup survives.  
+3. **Downstream effects audit:** Report reward-model scores or human preference deltas for seeded-accept decoding vs baseline speculative and vs classic watermarks. Even a small ablation will reassure readers that “max strength” doesn’t quietly harm alignment/quality.  
+4. **Clarity & reproducibility:** Expand the derivation around Eq. 10 in the main text (one boxed paragraph). Right now it jumps out of the blue without providing much background.
+
+The narrative is coherent and the prior binary impossibility is treated respectfully. I appreciated the clean placement of this work relative to Hu & Huang and SynthID, as well as the detector connection to Bayesian scoring networks. I would like to see clearer statements about assumptions (knowledge of $P_t$, seed/coin access) up front.
+
+
+## Writing and Presentation
+Generally clear but somewhat difficult to follow; figures carry their weight. The "what to believe after reading" is crisp: *use KL as strength, plot the Pareto, seed the accept coin to sit on the frontier.* A bit more exposition around the constrained program and detector calibration would help readers who don’t already know speculative internals.
+
+## Nits / small issues
+- Define “accept coin” once in the intro or prelims to avoid hunting later.  
+- Around the entropy upper bound, one sentence explaining $WS = H(P) - E[H(P_\zeta)]$ would help a beginner.  
+- Minor typographical cleanups: consistent capitalization of "SynthID," spacing around $TV(\cdot,\cdot)$, and equation referencing.
+
+### Questions
+See above
+
+### Soundness
+3
+
+### Presentation
+1
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+This paper considers watermarking when speculative sampling is used. Their results seem modular, allowing for improved watermarking with speculative sampling by improving standard watermarking schemes and applying them to both the draft and target models. Making this optimal seems to require an interesting trick that (pseudo)randomizes acceptance. They appear to understand all the key points, like that this will not introduce new distortions if the watermarks for the two models are independent.
+
+### Strengths
+The paper has several strong results. The basic observation, that one can overcome the limitation of Hu et al. by allowing for pseudorandom draft-token acceptance, is very nice and important. Watermarking under speculative sampling is an extremely important technical issue for deploying watermarks in real-world LLMs.
+
+I haven't read the paper carefully, but it seems to make real progress on an important practical question.
+
+### Weaknesses
+I don't know of any.
+
+### Questions
+Is Google's watermark really a deterministic function of pseudorandom numbers? I thought that tournament sampling doesn't satisfy this property at all: Doesn't it work by first sampling a few tokens, and then choosing deterministically between them? In which case, there's still a bunch of entropy left on the table in the initial sampling process.
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 6
+
+### Summary
+This paper studies the trade-off between watermarking and speculative decoding. A relatively recent work had made a 'no-free lunch' observation, which shows that improving the watermark is at the cost of weakening the speculative decoding (acceptance rate), and vice versa. This paper challenges this observation by integrating the watermarking's pseduorandom mechanism into the speculative decoding mechanism. The paper studies the speculative-decoding and watermarking tradeoff and the resulting curves, proposes a method to 'break' these curves and presents some experiments, demonstrating the proposed method on the Gumbel and SynthID watermarks.
+
+### Strengths
+Overall, this is a solid and creative work. Here are some of its strengths:
+
+1. The paper is clearly written. Even though I have significant prior knowledge about both watermarking and speculative decoding I believe this paper will not be too hard to follow for a newcomer to these fields.
+
+2. The idea is novel and seemingly powerful. Furthermore, the authors address the combination of two important subjects in the AI community - trustworthiness and efficiency.
+
+3. The analysis is very interesting and creative! Reading this paper was really delightful.
+
+### Weaknesses
+### Major:
+
+- **Clarity on SynthID:** The paper uses SynthId as a case-study in this work, as it is an 'unbiased watermark'. However, SynthID is a relatively large class of watermarks that follow the tournament sampling mechanism. The general SynthID watermark is not even unbiased (when N>2). Unfortunately, the authors do not provide sufficient information to understand which specific case of SynthID is proposed in this work. I also believe that such information should be added for the sake of clarity.
+
+- **Missing related work:** The content of this paper misses several related works on 'unbiased' watermarks for better context, which can also be considered for the analysis and experiments. The first [1] considers a similar hypothesis test to the one that is considered int his work, however through the Bayesian setting. This work is highly relevant due to its analysis in the 'token-level', as considered in this work. Furthermore, the results of [1] are stated in terms of the total variation distance, which may lead to easier integration with the speculative-decoding analysis that is done in this work. Beyond that work, there are highly related work that is missing for context - [2] provides a general formulation in terms of score and watermark transform optimization, proposing methods that outperform the considered watermarks in this work. [3] studies optimality 
+
+- **Experiments section should be extended.** I would expect this section to consider more experiments, through more datasets (testing across specific tasks/domains) and testing more aspects of watermarking (robustness?). The current collection is quite narrow. Furthermore, I would expect a quantification of the watermark strength through p-values and quantification of the bias (through either textual quality or cross entropy/relative perplexity). 
+
+### Minor:
+- The title of section 3 "COMPLETE THE TRADE-OFF DIAGRAM" is not clear.
+
+[1] Tsur, D., Long, C. X., Verdun, C. M., Hsu, H., Permuter, H., & Calmon, F. P. (2025). Optimized Couplings for Watermarking Large Language Models. arXiv preprint arXiv:2505.08878.‏
+
+[2] Tsur, D., Long, C. X., Verdun, C. M., Vithana, S., Hsu, H., Chen, C. F., ... & Calmon, F. HeavyWater and SimplexWater: Distortion-free LLM Watermarks for Low-Entropy Distributions. In The Thirty-ninth Annual Conference on Neural Information Processing Systems.‏‏
+
+[3] He, Haiyun, et al. "Universally optimal watermarking schemes for llms: from theory to practice." (2024).‏
+
+### Questions
+Bellow are both questions and comments:
+
+- **Connection to (Hu & Huang 24) and the hypothesis test:** The difference from the results in (Hu & Huang 24). which is explained in Remark 3.1 is unclear to me. First, the comparison itself is a bit odd. The referred paper discusses asymptotic results of the hypothesis test, while this work (while not stating clearly) is a token-level mechanism. The notion of a sequence is not addressed at any stage of this work. Second, Both works quantify the watermarks strength (detection) through the KL between distributions. This is an immediate results from the error exponent of the hypothesis test. 
+
+- Following the previous comment, as the authors consider the 'token-level' hypothesis test, the use of the KL divergence to quantify the distributional divergence seems a little odd. In the hypothesis testing context, and as noted in this paper, KL divergence emerges in an asymptotic analysis (Through the error exponent of the type-2 error in the equivalent independence the between the 'i.i.d.' tokens and the psuedorandom process). The setting that manifests the KL divergence doesn't seem to fit strongly in the proposed operational setting of this work. I would appreciate the authors' thoughts on that. I believe that the mentioned related works [1,2] might be highly impactful to address this issue.
+
+- The expected KL divergence in eqn (7) is a popular quantity in information theory termed 'conditional KL divergence' - I would use that standard notation.
+
+- I believe that the watermark transformation can be regarded as a transition kernel similarly to the speculative decoding mechanism. Is there a reason to not address is as one in this work?
+
+- Do the results of this work extend to the unbiased case? I would believe that the tilting mechanism in [2, Section 2] might come in handy as a formulation of 'removing bias'.
+
+- Is the proposed method applicable to a general new watermark or does a new solution should be tailored to each new watermark?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/Hf7jMztvve.md

@@ -0,0 +1,135 @@
+## Human Reviewer 1
+
+### Summary
+This paper investigates whether large language models can engage in strategic deception and whether current interpretability tools (especially SAE features) can detect it. The authors design game-like tests to elicit deceptive behavior and analyze SAE activations during these scenarios. They claim that the deception features identified by SAE fail to activate when lying occurs, suggesting current interpretability methods are blind to such behavior. The work positions itself as an early attempt to probe “hidden deception” in LLMs.
+
+### Strengths
+1. **Creative experimental design**: The “Secret Agenda” and “Insider Trading” setups show some originality in trying to operationalize deception through game-like tasks.
+2. **Broad model coverage**: The study evaluates a large number of models (around 38), providing some breadth and comparative perspective across different architectures and providers.
+
+### Weaknesses
+I find the main weaknesses of this paper to be (1) its poor situating within existing literature and (2) major issues in clarity, rigor, and methodological grounding.
+
+First, the paper is poorly situated in related work. It makes strong claims about deception and interpretability without referencing key prior studies in truthfulness, mechanistic interpretability, or SAE analysis, leaving its novelty unclear.
+
+Second, the organization and clarity are weak. The introduction is too short, figures are oversized, and many technical details—such as how features are selected or how “feature fine-tuning” is done—are missing or unclear (e.g., “when tuned to -1 nor when tuned to +1” is never properly explained).
+
+Third, the authors misinterpret and underestimate current interpretability methods. Claiming that SAE “deception features” failed to activate does not justify dismissing the SAE framework; this reasoning overlooks the distributed and compositional nature of such representations.
+
+Fourth,  the paper criticizes existing evaluation methods without evidence. In line 100, the authors state that current evaluations are “unreliable, hard to replicate, and unrealistic,” yet they never specify which methods these are or provide comparisons to support the claim.
+
+Finally, the format and presentation are below standard. The text is brief relative to the figures, captions are vague, and the layout feels unbalanced and unfinished.
+
+### Questions
+Please refer to weaknesses
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper presents a reproducible testbed for eliciting deceptive behavior in LLMs based on a game that requires the player to lie to succeed. It demonstrates that this elicits deceptive behavior in 38 tested models. It explores the effectiveness of auto-labeled SAE features for detecting strategic deception in the new deception test bed. It claims that auto-labeled SAE features are limited in their ability to detect or suppress deception as demonstrated by small latents in "deception" related features and continued instances of deception despite suppressing those features. Lastly, it presents a method for clustering SAE latents using PCA+t-SNE to demonstrate domain-dependent effectiveness of SAE feature interpretability.
+
+Mandatory disclosure of LLM usage by the reviewer: This review was formatted using LLMs to structure the text into paragraphs and numbered lists.
+
+### Strengths
+1. The paper addresses an important topic of strategic deception using a new binary testbed.
+
+2. It demonstrates that deception can be elicited in across 38 different models.
+
+3. The results are reproducible and experimental setup is well documented.
+
+4. The effectiveness of SAEs at detecting strategic deception is put into question and backed up by experiment.
+
+### Weaknesses
+1. In section 4, the paper claims that Secret Agenda game aims to generate "high stakes deception", however given that the problem is framed as a game does not feel high stakes as in Agentic Misalignment (arXiv:2510.05179) where the agent reported that it was fully aware that its actions will cause harm to a human being.
+
+2. The game framing may limit generalizability in two ways: (1) models may be following game rules rather than exhibiting deception that would occur in deployment, and (2) the artificial stakes may not activate the same neural mechanisms as real-world deception. While the authors acknowledge this is a "controlled deception laboratory", the connect between game-based deception and deployment risks needs stronger justification.
+
+3. It would drive the point home if the Secret Agenda test bed could elicit deception in case deception is clearly harmful, though I admit that deception of any kind is evidence enough - it would be trivial for a third party to cause the model to deceive.
+
+4. Performing section 7 scatter plot for Secret Agent SAE latents seems like a natural step that could shed light on the proposed feature search methodology.
+
+5. Along the same lines I would be very curious to see how raising or lowering the states changes models behavior and the SAE activations. Will SAE features become more sensitive in a "real" scenario? Will the model avoid deception knowing that it will cause harm?
+
+### Questions
+1. In section 7.3, by "contrasting results" do you mean the difference between Secret Agenda vs Insider Trading or something else? If so, why was the "top discriminative features" not and PCA+t-SNE scatter plot analysis not applied to Secret Agenda SAE latents?
+
+2. The authors acknowledge under limitations that game scenarios might not generalize to all deception types. There are clearly tradeoffs here that make it difficult to pin down and also gets at the heart of the problem. If the scenario is a game entirely different features may activate for "in-game deception" compared to "real world deception". Could you understand whether what (if any) SAE features activate highly for "in-game deception"?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper outlines an investigation of strategic deception in LLMs using two developed testbeds – Secret Agenda and Insider Trading Compliance. The former testbed is based on its namesake, a social deduction game with the LLM set as the Dictator, wherein the games are set up such that lying is the easiest path for the model to win. Experiments in this “controlled deception laboratory” show that LLMs consistently resort to deception (Figure 1).  More troublingly, the models achieve this deception without activating features that are expected to be linked to deception. On the other hand, the Insider Trading Compliance testbed exhibited activation of deception-linked autolabelled features; the models continued to engage in insider trading almost half the time.
+
+Put together, these results illustrate the serious challenge and risk posed by LLMs strategically lying to maximize their stated objectives. Moreover, the disconnect between autolabelled deception-related features and the activations observed in the Secret Agenda tests limit the utility of tuning these features to limit strategic lying by the LLMs.
+
+### Strengths
+The paper sheds light on an important challenge, especially when generalizing LLMs to other domains. Honesty of models when providing responses is paramount to build trust for users and a necessity for their adoption in domains with high compliance standards. Consequently, testing methodologies to audit models for deception and testing the efficacy of corrective measures is an important contribution. 
+
+In particular, the result illustrated in Figure 1 is the most striking because it shows that all the tested models strategically resort to deception most of the time. Granted that Secret Agenda puts the models is adversarial situations where deception is the easiest path to victory, but it is desired and possibly a must, that LLMs do not lie even in such settings. While the result in Figure 3 is not as striking, the models electing to engage in insider trading nearly half the time is also a worrisome and significant result.
+
+Finally, the conclusion that the pertinence of autolabelled SAE deception-relevant features for model deception is likely setting-dependent is noteworthy as (a) it suggests that the utility of these features and deception mitigation strategies predicated on these features could be severely limited in some scenarios and (b) indicates that feature-based interventions and interpretations are more instructive with structured data.
+
+### Weaknesses
+While I appreciate the wider point and results conveyed by this paper, the writing of the paper leaves a lot to be desired. It is overfull with jargon and, missing references and explanations; this necessitates that the reader either be very familiar with this area or read multiple references so that he/she may understand the paper and the results contained therein.
+
+An example of this are the repeated references to SAEs and features labeled by SAE but its full form (Sparse Autoencoder) is only stated on Page 5. Moreover, there is no explicit discussion of SAE-based labelling of features and its uses for interpretability. Another example of the lack of details are the various game contexts in Section 5.3, with little explanation for the game contexts and explanations for the authors’ selection of these contexts in their tests.
+
+### Questions
+**Questions:**
+
+1.	What do the authors mean on line 59 when they aver that their “operationalization remains agnostic to assumed beliefs”?
+2.  On line 260, it is stated that refusal responses comprise the majority of outputs. This appears to contradict the results in Figure 3, and more broadly contrasts with the results of the Secret Agenda tests.
+
+**Suggestions:**
+
+1.	In the conclusion, the authors state that “identity-feature interventions often degraded outputs into repetitive, incoherent loops.” This is the first mention of any such result in any part of the main paper, and hence does not follow from any of the preceding discussion. The previous discussion of this result should be moved into the main paper or this should be removed from the conclusion to ensure coherence of the manuscript.
+2.	Please justify any design choices made in the proposed test beds.
+
+### Soundness
+4
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/IvG90aRAL0.md

@@ -0,0 +1,193 @@
+## Human Reviewer 1
+
+### Summary
+This paper provides a quantum version of the Frank-Wolfe algorithm for constrained convex optimization. Compared to prior work (Chen & de Wolf, 2023) which deals with regression with l1 and l2 norm constraints, the authors tackles the convex optimization in a more general setting, proposing methods for optimization on a smooth convex function with l1 norm constraints, a simplex constraint or a latent group norm constraint, as well as matrix nuclear norm constraints. Their algorithm can provide quadratic speedups in terms of the dimension of the domain.
+
+### Strengths
+1. The paper addresses an interesting problem in quantum optimization and generalizes results to a much larger set of convex optimization problems, extending the applicability of such algorithms from regression to other potential fields such as signal processing (sparsity constraints via l1 norm), game theory (zero sum games with simplex) and SDPs (nuclear norm optimization).
+2. Proofs and assumptions are clearly documented, with different scenarios of gradient obtainment for the vector case discussed in detail.
+
+### Weaknesses
+1. Cost of additional qubits and gates not provided for Theorem 6 in Table I.
+2. Applications of this paper can be more clearly articulated, i.e. what's been listed in strengths 1.
+3. Assumption for gradient obtainment in the matrix case seems a bit strong, as the gradient has to be low rank _and_ accessible in a KP tree.
+
+### Questions
+1. Given that some previous quantum algorithms that tackle the problem of convex optimization rely on the multiplicative weight update method eg. quantum LP and SDP solvers, I am curious as to how the quantum Frank-Wolfe method proposed in this methods compares to these methods and what potential advantages/disadvantages exist for the quantum Frank-Wolfe method.
+2. Why is a different assumption for gradients used for the matrix case different from the vector case? Is the gradient estimation algorithm not applicable to matrix cases? 
+3. I think the strength of gradient assumption in the matrix case may be stronger than the membership and separation oracles in ariXiv:1809.00643 as you could obtain such oracles when you have both function and gradient access. How would the algorithm perform this weaker set of assumptions is provided instead of the gradient assumption?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This work presents a framework for Quantum Frank–Wolfe (QFW) algorithms, extending classical projection-free convex optimization to the quantum setting. The authors design quantum analogues of the Frank–Wolfe method for both vector-domain (e.g., $\ell_1$-ball, simplex constraints) and matrix-domain problems (e.g., trace-norm constraints). By leveraging quantum primitives such as Dürr–Høyer maximum finding, Jordan’s quantum gradient estimation, and quantum singular value estimation (QTSVE), they show their algorithms achieve a $\sqrt{d}$ improvement in dimensional dependence while maintaining the classical convergence rate of $O(1/\varepsilon)$.
+
+### Strengths
+1. The manuscript presents proofs and complexity analysis, which seem reasonable.
+
+2. The proposed quantum framework attempts to address a range of convex optimization problems in both vector and matrix domains, including $\ell_1$ norm and trace norm (nuclear norm) constraints.  
+
+3. The manuscript presents a theoretical exploration of quantum algorithms in the context of convex optimization problems.
+
+### Weaknesses
+1. The lack of simulations and qubit estimates for the core matrix algorithms (QPM/QTSVE) makes it challenging to assess the practical feasibility and actual performance of the proposed quantum speedups. 
+
+2. The algorithm requires strong quantum oracle access (e.g., normalized row states and function value oracles), which are currently not feasible on NISQ devices.
+
+3. The algorithmic components (quantum maximum finding, SVD estimation, gradient estimation) are not new. 
+
+4. The use of $\tilde{O}$  notation hides constants that could have a significant impact on the practical runtime and performance of the algorithm.
+
+### Questions
+1. The QFW algorithm is designed for convex optimization, do you think its techniques could be adapted to solve non-convex problems like BQP, or provide useful insights for tackling such NP-hard problems?
+
+2. How does the quantum acceleration of the Frank-Wolfe algorithm for sparse convex optimization  compare with the optimization methods for bi-quadratic programming over unit spheres (as discussed in [1])  in terms of computational efficiency and the ability to handle high-dimensional optimization problems?
+  
+
+
+
+[1].  Li, S., et al., Tighter bound estimation for efficient biquadratic optimization over unit spheres, Journal of Global Optimization, 2024.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper studies how quantum computing can accelerate projection-free optimization methods, particularly the Frank–Wolfe (FW) algorithm, which avoids costly projection steps by solving linear subproblems. For example, they studied for $\ell_1$-norm and simplex constraints for vector optimization problems, and nuclear norm constraints for matrix variables.
+
+### Strengths
+- Originality: the authors study quantum acceleration for projection-free convex optimization for several classes of optimization problems. They propose some ways to achieve quantum acceleration under the FW framework, like quantum gradient estimation, singular value extraction, et.c.
+
+- Quality: They gave theoretical analysis of their methods, proving a worst case guarantee.
+
+- Clarity: The writing of the paper is good. Their problem formulations, algorithms and theoretical results are presented in a structured way.
+
+- Significance: Projection free methods for optimization is widely used and important in machine learning. Their results show the possibility of quantum speedups in this aspect.
+
+### Weaknesses
+The paper’s algorithmic advances rely heavily on existing quantum subroutines—such as quantum maximum finding, gradient estimation, and singular value extraction—raising questions about how much novelty lies beyond combining these tools within the Frank–Wolfe framework. The authors could better articulate what new technical challenges are overcome or what insights are unique to the projection-free setting. For example, what is the novelty of this work on vector variable problems compared with Chen & de Wolf (2023)?
+
+Moreover, the analysis would benefit from a clearer discussion of applicability and limitations. The claimed speedups depend on assumptions like low-rank gradients and favorable spectral gaps; in dense or ill-conditioned cases, the advantage may vanish. Explicitly characterizing these regimes and their practical implications would strengthen the paper’s significance.
+
+### Questions
+- Clarification on Novelty and Technical Contributions: Beyond integrating known quantum primitives, what are the key new technical insights or analyses specific to the projection-free (Frank–Wolfe) setting? For example, are there difficulties in ensuring convergence or oracle compatibility that required new ideas?
+
+- Regime of Quantum Advantage: The quantum speedups depend on parameters like the rank, singular value gap, and Lipschitz constants. Could you provide a more explicit characterization of the parameter regimes where your algorithms achieve practical advantage over classical FW? Also, are there examples (synthetic or theoretical) where these advantages disappear or become marginal?
+
+- Clarification of the Matrix-Case Analysis: In the matrix domain, your algorithms assume precomputed gradients. Would including gradient computation change the overall asymptotic complexity or speedup factor? The analysis relies on rank and spectral gap assumptions—can you discuss robustness when these assumptions are only approximately satisfied?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper investigates projection-free optimization for sparse convex problems. It proposes two quantum algorithms—handling vector and matrix domains respectively—that achieve improved complexities by reducing the dependence on the dimension d.
+
+### Strengths
+The paper is well-written, featuring a clear presentation of the methods, the obtained complexity results, and comparisons with prior works.
+
+### Weaknesses
+My primary concern pertains to the paper's motivation. I find the core motivation insufficiently justified. The manuscript does not adequately establish a compelling need for specialized quantum algorithms for this particular class of problems. From my perspective, the proposed method appears somewhat ad-hoc, lacking both a clear demonstration of practical application and the provision of new, general insights for the field of quantum algorithm design.
+
+### Questions
+1.	What is the motiveitaion of this paper? Why do we need to solve this type of problem? 
+2.	Why do we need to accelerate Frank-Wolfe algorithms? Is it possible to design special algorithm for these sparse learning problems?
+3.	Recent advanves in non-convex optimization often solves sparse (low-rank) problem by reformulating the problem by a quadatratic parametrization problem? Why do you work on this way?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 5
+
+### Summary
+The paper is focused on convex optimization with convex constraints in high dimensional settings. The authors propose a quantum approach for achieving quadratic speedup, in terms of the problem dimensionality, by using quantum computing to solve the subproblems in Frank-Wolfe optimization algorithm.
+
+### Strengths
+The paper provides an original approach towards solving constraint optimization problems. The significance of the result is diminished, as noted in weaknesses, but deficiencies in complexity analysis of the method and its assumptions. The paper is clearly written, although the assumptions should be more prominently analysed.
+
+### Weaknesses
+The main weakness involves unrealistic assumptions that affect the complexity of the full algorithm. For example, step 7 in Alg. 2 involves preparing, in each Frank-Wolfe iteration, a state
+$\sum_{i=0}^{d-1}|i>|x^{(t)}>|0>$, with $|x>$ defined in Assumption 3 to be $|x> = |x_1>|x_2>…|x_d>$. The state is over $d+2$ qubits. Similar qubit $O(d)$ qubit requirement is present in the rest of scenarios considered in the paper. This linear qubit requirement severely limits the applicability of the method for the stated goal of addressing high dimensional (high $d$) optimization problems. 
+
+Preparing a $d$-dimensional arbitrary qubit state is hard in general, yet the impact of this step on the complexity of the whole method is not discussed.
+
+Some terminology could be revised, e.g. authors mention “sparsity-constrained problems” in the introduction, but actually address L1 norm and not L0 norm constraints.
+
+### Questions
+Are there any specific properties of $|x^{(t)}>$ over the iterations that would make state preparation efficient?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3

human_reviews/JEYWpFGzvn.md

@@ -0,0 +1,110 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes an adaptive visual tokenizer that allocates a more optimal compressed sequence length than existing methods based on information theoretic principles. This paper demonstrates that prior works’ training methods are often suboptimal, imputing bias that makes the adaptive compressions lower-quality despite their adaptability. It further claims that the optimal design for the compressed token length should be proportional to the negative log likelihood of the input, and thus uses the learned ELBO to modulate the input compression factor. The proposed router uses this principle to more optimally allocate tokens for compression, and the top K tokens are selected based on their log likelihood. The results demonstrate that InfoTok achieves better performance at the same compression rates as comparable methods.
+
+### Strengths
+1) This paper provides some much-needed theoretical grounding to the adaptive tokenizer space, and attempts to pin down what `optimal’ adaptive compression should look like.
+
+2) Experiments are very comprehensive and clearly demonstrate that InfoTok achieves better reconstruction at the same compression rate compared to other adaptive tokenizers.
+
+3) While the results are incremental on their own, the entire paper is well structured and motivated, and properly places all the contributions in context, making the result valuable for the community.
+
+4) While the claims in the theorems are initially hard to parse, they are empirically supported by experiments that make the intuition clearer.
+
+### Weaknesses
+1) Key definitions are not well highlighted in the text, and the mathematical sections are poorly written and hard to follow. Since definitions are missing from the theorems, the explanation is rather hard to follow. For example, the entropy H is not clearly defined, \mathbb{D} not defined. Similarly r(N | x) not defined in Alg 1 input. While I may have missed these definitions, I would suggest not burying them in the text and making key components clear and easier for readers to refer to.
+
+2) It would be helpful if there were some visual results demonstrating the differences in InfoTok's reconstruction at different token lengths for the same image, so we could see what inputs are determined to need lower ones and which are determined to need higher ones. The reconstructions shown in Figure 2 are not particularly compelling - the compression results are considerably worse than Cosmos-DV.
+
+3) There are no video samples provided in the supplementary material or project page which makes qualitative analysis of the results challenging.
+
+### Questions
+I don't have any strong suggestions, but would appreciate responses to my above critiques and would like to see them addressed in the next version of the paper.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes InfoTok, an adaptive discrete tokenizer that leverages an information-theoretic formulation to allocate token budgets based on the data's compressibility. The method introduces a principled way to determine token lengths using a normalized ELBO-based router and achieves strong empirical performance with a conceptually simple approach.
+
+### Strengths
+* The paper is well structured, and the mathematical formulation is precise.
+* It achieves competitive or superior performance using a lightweight and interpretable mechanism.
+* The authors provide rigorous justifications connecting ELBO with optimal token length, and the derivation is insightful.
+* The ELBO-based routing and token selection reuse existing encoder-decoder structures and introduce minimal inference cost, which is appealing in practical deployments.
+
+### Weaknesses
+* Is the per-token ELBO computed purely based on the encoder-decoder’s end-to-end reconstruction path? Must this be explicitly introduced during training, or can the method be plugged into any VAE-style model without changes? Specifically, can non-VAE tokenizers be adapted to this framework?
+* If N_max is small or the compression ratio is extremely low (e.g., β< 0.1), what are the observed effects on stability and convergence? Would the KL term dominate or explode in such settings, and does it impact model converging or generalization?
+* The method is designed for videos. Would it generalize to other structured data types such as audio and 3D point clouds? Do different modalities affect the ELBO-based router's assumptions or effectiveness?
+
+### Questions
+Please refer to the weaknesses.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+In this paper, an algorithm called InfoTok is proposed for adaptive video tokenization. The proposed algorithm determines the token length based on ELBO for near optimal compression. For compression, InfoTok utilizes the ELBO based token selection. Thus, it can be integrated on the top of existing tokenizer architectures. Also, the proofs for the mathematical theorems are provided. The experimental results show that the proposed algorithm outperforms the existing methods.
+
+### Strengths
+- This paper clearly describes the proposed algorithm and is easy to follow.
+- This paper provides the proofs for the theorems.
+- The proposed algorithm achieves good performance on various benchmark tests.
+
+### Weaknesses
+- It would be helpful if the paper included a discussion of the limitations of the proposed approach. 
+- Since the method seems general enough to be applied to images, it would be helpful to explain the rationale for limiting the experiments to videos.
+- typo
+  - L184: an more accurate -> a more accurate
+
+### Questions
+Though I have carefully reviewed the paper including the appendix, I did not observe any major drawbacks and I believe the proposed algorithm makes a meaningful contribution to adaptive video tokenization. Please see my minor concerns in the weakness section.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+8
+
+### Confidence
+4

human_reviews/LD6B7AvkZq.md

@@ -0,0 +1,194 @@
+## Human Reviewer 1
+
+### Summary
+This paper investigates how Transformer-based autoregressive models learn languages generated by Probabilistic Context-Free Grammars (PCFGs). The authors provide a theoretical framework showing that the Kullback-Leibler (KL) divergence loss decomposes across the PCFG's subgrammar structure. Empirically, using small Transformer models, they demonstrate that models reduce loss across all subgrammars in parallel, rather than mastering simple structures first. Their experiments also reveal that these models perform poorly on deeply recursive structures and that pretraining on subgrammars can improve the final loss for these smaller models.
+
+### Strengths
+1. The paper provides an elegant theoretical derivation (Theorem 4.6) that helps explain why Transformers perform poorly on deeply nested grammatical structures . The recursive formula for KL divergence based on expected recurrence is a strong conceptual contribution.
+
+2. The experiments show that Transformer models reduce loss across all subgrammars simultaneously, which contrasts with the developmental stages observed in human language acquisition.
+
+3. The work includes a clear set of experiments on small models demonstrating that subgrammar pretraining can improve final performance , offering a practical insight for training on structured data.
+
+### Weaknesses
+1. The key "understands composition" assumption, which forms the basis for Corollary 4.5 and Theorem 4.6, is very strong. The paper does not provide sufficient empirical validation to determine to what extent this assumption holds for practical, trained models, which limits the applicability of these specific theorems.
+
+2. The experiments on Large Language Models (LLMs) are limited to anecdotal tests on "ChatGPT-5-Instant". This experiment may conflates two different tasks: calculating the result of an arithmetic expression and modeling the grammar of that expression. The task of calculating a result is not equivalent to the task of generating an expression from a grammar, especially from a model training perspective. Therefore, the difficulty an already-trained model shows in calculating a deeply recursive arithmetic expression is not direct evidence that a Transformer-based model would struggle to be trained on a deeply recursive PCFG (the paper's primary claim).
+
+3. The comparison in Section 5 between pretrained models and those trained from scratch seems unfair. Pretraining on a subgrammar essentially gives the model a strong hint, or an "inductive bias," about the PCFG's correct structure. The baseline model, however, gets no such hints and must discover this hidden structure all on its own. Therefore, the finding that pretraining improves the final loss for smaller models might primarily demonstrate the benefit of curriculum learning and injecting this strong structural prior.
+
+### Questions
+1. There appears to be two minor typo in the derivation in Section 4.2 (lines 211-213, page 4). In the second line of the equation.
+
+The correct expansion of - log Q_\theta(\alpha a\beta) should be [ - log Q_\theta(\alpha ...) - log Q_\theta(a|\alpha) - log Q_\theta(\beta|\alpha a)]
+
+The current text incorrectly uses positive signs for the second and third components. Additionally, the conditioning context for the final term appears as \beta|a\alpha, which seems to be a notational typo for the correct \beta|\alpha a.
+
+In the third term of the equation on line 214, the probability $P(a)$ should likely be written as $P_A(a)$.
+
+2. ChatGPT-5-Instant and ChatGPT-5-Thinking seems not formal name of these two models.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The work analyses the language models’ learnability of PCFGs. They introduce a subrammar-based framework for probing an architectures learnability. A subgrammar decomposed KL is derived. The authors find that the loss decreases in parallel over the subgrammars. The work also considers the impact of pre-training and depth of the grammars.
+
+### Strengths
+The idea of studying the learnability of sub-grammars is sound and interesting. And using formal languages is a, by now, common approach to study learnability. Using hierarchical subgrammars is a natural, and the decomposition of the KL is as expected.
+
+The pedagogical approach of the paper, laying out standard definitions and derivations is welcome. And giving "worked examples” of e.g. the KL decomposition at the bottom of page 4.
+
+The three kinds of analysis are nice (albeit a bit shallow individually)
+
+### Weaknesses
+The work overstates the novelty of using CFGs to probe language models' learnability, e.g. in the introduction “this work initiates the theoretical and empirical study of training dynamics over CFGs” but this is repeated in many places. See for instance Sennhauser and Berwick 2018 (Evaluating the Ability of LSTMs to Learn Context-Free Grammars), or Allen-Zhu et al 2023 (How Language Models Learn Context-Free Grammars).
+
+Another drawback is that the grammars are tiny, and the model sizes not ablated with grooving grammars. I understand a framework is being suggested, but to justify the framework it would be good to do some kind of scaling. It doesn't have to be for very large models, or CFGs with thousands of rules. But sampling grammars (and averaging over them) and doing so for grammar sizes of e.g. 10,100,200,400,800 rules would indicate whether the findings are due to the simplicity of the languages or not. Especially in a section titled “How Language Models Learn”
+
+In line 301 you say you train on several synthetic CFGs on transformers, but the settings are missing or not referenced. I.e. architecture and training parameters, grammars sizes.
+
+In line 379 you make a claim about an effect that diminishes as models size and complexity increase but it seems to be backed by a single example.
+
+The ChatGPT-5-Instant experiment is unclear, what is the setup? It is also missing a citation or description. 
+
+I recommend the authors fix the overly bold claims and smoothen the writing a bit, consider the related literature a bit better and strengthening the empirical study when revising their work. Ensuring full reproducibility is also important, make sure hyperparameters and configurations are reported and referenced.
+
+Minor:
+In the introduction you state that training an LLM for studying its training dynamics is unfeasible. Even so, there is plenty of room to study their learnability without reaching SOTA and no reason to believe the findings don't apply to the top models.
+
+You could condense the standard definitions (kl, entropy, mle etc) a bit as prose, but im ok with the definition environment.
+
+Please use numbers by equations (the equation or align environments), it’s hard to refer to them otherwise.
+
+### Questions
+In the activation space analysis you show that pretraining on in-domain data helps. How would you take this further, i.e. do you see some options for using more complex/gradual curricula?
+
+The deeper recursion experiment is interesting, do you see a way to connect it to the curricula approach?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper aims to study how transformers learn context-free grammars as a step toward a theory of language modelling. The authors claim that transformer training on context-free grammars can be understood by decomposing the cross-entropy loss into contributions from subgrammars. Under the assumption that a language model "understands composition", the authors derive a decomposition of the loss that elucidates the role of subgrammars and their recursive occurrence. This is an interesting and timely direction, but the contribution is unclear to me. Due to this lack of clarity, it's impossible for me to properly judge this contribution unless some of my doubts are addressed first (see below).
+
+### Strengths
+This is an ambitious and relevant topic, and I do believe that studying how language models acquire simplified formal languages is the path towards a predictive theory of large language models.
+
+### Weaknesses
+### Relation to previous literature
+
+First, the claim that the authors' work "initiates the study of the learning dynamics of transformers trained on PCFGs" is wrong. There are at least two groups contributing in this direction: https://arxiv.org/abs/2406.00048 (see also https://journals.aps.org/prx/abstract/10.1103/PhysRevX.14.031001?__cf_chl_rt_tk=UOc8KN3VOXQLYip1DgN1hgH_sgm85mQO4KFwHUmz38A-1761924271-1.0.1.1-05csO._aVplfQKTVpNHOiSYCZ.oTz8qXQePSbQqMnLE, https://arxiv.org/abs/2505.07070, ) and https://arxiv.org/abs/2305.13673v4 (one iteration https://openreview.net/forum?id=qnbLGV9oFL also shares a portion of the title with the present submission). These contributions should be acknowledged and their relation to the present paper discussed.
+
+---
+
+### Clarification of the exact contribution
+
+If I understand correctly, the main technical contribution is the proposed loss decomposition. What is the intended consequence of this result? For instance, the authors claim that the decomposition implies that the model learns all subgrammars in parallel. However, a decomposition of the population loss can only constrain the location of the final optimum, not the dynamics of optimisation. In particular, it has no direct implications for convergence rates, sample complexities, or learning order. What additional assumptions would be required to justify the claim of “parallel learning”?
+
+---
+
+### Technical claims
+
+1. **Inconsistent notation (pp. 4–5).** Symbols change mid-derivation (e.g., (P(\alpha\dots)) vs (P(\alpha\mid\varepsilon))), and in *Definition 4.2* the dependence of the summand on (a) is unclear. Please fix the notation globally and clarify how (a) enters the RHS of Def. 4.2/
+
+2. **Application of Definition 4.2.** When (A) is a fixed string (as in (D_{\mathrm{KL}}(P\Vert Q)_{\alpha})), I cannot reproduce the final equation on p. 4. An explicit worked example would help verify the derivation.
+
+3. **KL decomposition  of Theorem 4.3. from Definition 4.2.** Please show the full derivation of the KL decomposition.
+
+4. **“Understanding compositionality.”** This assumption is too vague. Does it hold before, during, or after training, and what measurable property defines it?
+
+6. **Implications of Theorem 4.6.**  Specify whether the result concerns the population optimum or predicts learning dynamics (rates, ordering, or sample complexity). If the latter, state the additional assumptions required for “parallel learning.”
+
+---
+
+### Empirical validation
+
+The empirical section does not clearly test the theoretical claims. If the main statement is that the overall loss equals the sum of the losses over subgrammars, this should be explicitly validated in a figure, rather than requiring the reader to infer it by visually summing multiple curves. Moreover, the role of Figure 2 is unclear relative to Figure 1---what new information does it convey? Finally, the dependence on the recursion number (R, or rather its expected value), is never systematically explored or quantified.
+
+### Questions
+See weaknesses section.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper investigates how transformers learn probabilistic context-free grammars (PCFGs). It first derives several theoretical properties describing how the KL divergence of PCFGs decomposes over subgrammars. Using this framework, the authors show that transformers learn all grammatical components in parallel, unlike humans who acquire syntax hierarchically. Experiments with small transformers confirm these findings and demonstrate that pretraining on subgrammars improves performance and yields more structured internal representations, though the benefit decreases for larger models. Finally, the study shows that models can manage long but shallow dependencies yet struggle with deeply recursive syntax, a limitation also observed in large language models.
+
+### Strengths
+It is important to understand how transformers learn CFGs. Although previous work (Allen-Zhu et. al. 2023) has studied in detail of the mechanism on how transformers implement CFGs, the dynamic analysis remains blank. The paper studies this important problem. They reveal an interesting phenomenon that transformers lean all subgrammars simultaneously.
+
+### Weaknesses
+Although the topic is important, the results feel incomplete and lack depth.  
+
+1. The theoretical part at the beginning is quite trivial and unnecessary for understanding the dynamics. The first main result only appears at the end of page 6, making the presentation hard to follow.  
+
+2. Regarding how transformers implement CFGs, Allen-Zhu et al. (2023) have already provided detailed and insightful analysis that largely covers the results here. The paper does not clearly explain its own contribution beyond that work.  
+
+3. The only genuinely new finding seems to be that transformers learn all subgrammars simultaneously, but the analysis is shallow. I would like to see more detailed investigations and ablations to uncover the mechanisms behind this behavior.
+
+### Questions
+None.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4

human_reviews/NIvqiJ8R4J.md

@@ -0,0 +1,213 @@
+## Human Reviewer 1
+
+### Summary
+This work presents an adaptive tutoring framework consisting of two stages, integrating collaborative cognitive diagnosis with dynamic instructional adaptation. The first stage aims to model the student’s cognitive state through collaborative cognitive diagnosis. The teacher utilizes a successor-first method to efficiently generate diagnostic questions, ensuring their accuracy through an expert-assistant-verifier pipeline. In the second stage, based on the estimated cognitive state, the teacher uses slow-thinking-based methods to select teaching strategies from a strategy pool to guide the student in solving problems.
+
+### Strengths
+1. The entire framework is meaningfully designed and the authors present the prompt design clearly. 
+ 
+2. The authors present a systematic evaluation by comparing with baselines in cognitive diagnosis and adaptive tutoring, with ablation studies in different modules and backbones. 
+
+3. Several case studies and deeper analysis regarding the effectiveness of this framework in specific education scenarios.
+
+4. The related work presents necessary background for readers to understand the context.
+
+### Weaknesses
+1. Unclear model design rationale. It is unclear what the unique strength of this proposed model is and why it should be, compared with prior models. The introduction and related work sections briefly stated that prior work can not do something, but the detailed research gap and how this work addressed such gap are unclear.
+
+2. The current framework is meaningful, but there is no fundamental breakthrough or unique insights in algorithms or model side. It is more like a manually crafted application system in education with prompt-engineering.
+
+3. Dataset Scale. There is only one dataset with only 184 questions. 
+
+4. The main experiment used simulated students for evaluation, which does not seem to be convincing. It is also unclear how many simulated students were generated, the differences among the simulated students, and rationales for such design. Without such details about student settings in the experiment, it is hard to evaluate the rigor and effectiveness of this work.
+
+5. Lack of statistical tests (such as t tests) for most tables to show the significance.
+
+6. Main Experiment: The experimental measurements may not really support the claimed contribution in line 107: "stimulating critical thinking". Which metrics measure how the tutoring system improves the critical thinking of students in either simulation or real student study?
+
+7. Real Student Experiment: 
+
+The real student experiment is unclear and does not look like a formal study. Is it a between-subject design or within-subject design or a mixed study design? 
+
+For this statement: "one of six tutoring methods is randomly selected (with equal probability) to tutor the student", does it mean that each student only receives one tutoring method (between-subject design) or each student receives different tutoring method for different questions? If so, how do you control so many human factors like different students in different questions with different tutoring methods? If one question is delivered to different students in different tutoring methods, how do you know the result difference is due to student subject difference or due to the tutoring method difference? 
+
+Participant scale and limited samples: So many human factors usually need larger-scale user study rather than N = 67. The authors can report the power via power analysis to show the statistical significance with the current participant number. Moreover, for this statement "The student who selected the most questions chose 12, while the student who selected the
+fewest chose 2", does it mean that each student only tested 12 questions at most? This limited sample size is probably not convincing to show its effectiveness.
+
+
+Study design and metrics: Another concern is that the current study design and metrics may not really support the claim of enhancing tutoring performance. Most metrics are subjective (e.g., Appropriateness, Sentiment, Inspiration) and do not provide a very objective measurement. The success rate may be better quantified and more convincing, but Pelican's success rate (87%) is lower than simple step-wise methods (87.3%).
+
+
+Minor Issues:
+
+There are some typos.
+
+### Questions
+Please check my concerns and questions in Weaknesses section.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper proposes an adaptive tutoring framework called PELICAN, which is designed to address the limitations of existing LLMs in personalized education. It achieves student-centered teaching through a two-stage process: collaborative cognitive diagnosis and dynamic instructional adaptation. Specifically, it first organizes knowledge points into a hierarchical dependency structure and then conduct CD to get student's cognitive state. Then, the teaching strategies are dynamically selected from "fast thinking" and "slow thinking" modes based on the diagnostic results. Experiments conducted on GAOKAO bench demonstrate the effectiveness of the proposed framework.
+
+### Strengths
+1. The proposed 2-stage personalize education framework is well-defined and reasonable, and the introduction of dual system in stage-2 is convincing and fitting the problem well.
+2. The experiments on GAOKAO Bench is comprehensive, effectively demonstrates the superiority of the proposed PELICAN.
+3. The writing and structure of this paper are clear and easy to understand.
+
+### Weaknesses
+1. In my understanding, the construction of knowledge tree may heavily relies on manual work, which may hurts the scalability of the proposed framework. For example, scaling to full K12 education or vocational education would lead to a sharp increase in manual costs.
+2. The evaluation limits to only 1 benchmark (GAOKAO Bench), it will be more convincing if authors validate PELICAN framework on more benchmarks.
+
+### Questions
+See weaknesses.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The authors introduce PELICAN: a way to personalize education utilizing both LLMs and cognitive diagnosing. The LLM utilizes cognitive diagnosis to identify appropriate responses tailored to the student's cognitive level and understanding. Each problem has different knowledge components that the student must master and these are organized into a hierarchical structure. Based on this knowledge state, the adaptive tutoring will select an appropriate response to solve problem p with the student via dialogue. If the number of dialogue rounds with a student exceeds a threshold M, it is assumed that they are facing persistent cognitive obstacles, and slow-thinking is enabled to focus on smaller subproblems. The authors tested their method on the public Gaokao dataset, as well as a real-world study involving high schoolers.
+
+### Strengths
+The paper is well-motivated, with a clear introduction, and is overall easy to follow. I also think the figures are well-made and help aid the explanation of the ideas and methodologies. The appendix is thorough.
+
+### Weaknesses
+Overall, the paper lacks deep substance and novelty. It is more of an applied research paper, but it lacks statistical rigor (e.g., ANOVA tests, effect sizes, p-values). It could become more interesting and impactful with greater care and attention to detail for deeper results, analysis, and presentation, etc. However, I also don't find it to be a great fit for ICLR and don't think the audience of ICLR would be interested in this topic. This paper appears to be more closely aligned with AI applications in education venues, such as AIED, EDM, and EAAI, among others, once it has been improved. I have little to specifically point to in a critique of it, other than that it is a limited-scope paper that utilizes GPT models for teaching. This also raises concerns about the future reproducibility of the reported results. It is very time-specific and niche.
+
+Abstract:
+* Have LLMs really generated attention in education because of their extensive knowledge base and reasoning capabilities?
+* Extra space between "at" and "here" for code reference.
+
+Introduction:
+* Change "I Don't understand" to "I don't understand" in Figure 1
+* Add spacing: "OK,I get it!" between "OK," and "I" in Figure 1
+* First two paragraphs need citations to support their various claims
+* Parentheses around citations are needed. In-text citation references are likely not the style for ICLR.
+* Line 80: random word "Planning." appears
+* Figure 3: Why does "Explanation" lead to "Explanation" in the slow-thinking? Why does "Decomposition" have a crown?
+* Line 148: random word "Planning." appears
+
+3.2: "introduce a successor-first strategy, in which the teacher prioritizes assessing leaf nodes or nodes whose successors have already been evaluated," I thought line 175 said "a student can only master a child node after mastering its parent"?
+
+3.3: "dual system theory" citation? lines 215, 246
+
+Line 264: Is this what you are proposing? Simulated teaching tree? Or is that something that already exists in the literature?
+
+Line 270: Does m = |S| (i.e., the length of the strategy pool)? If not, why?
+
+Section 3: The formulas seem overly high-level to the point of being complete black boxes with no reproducibility
+
+Section 4: Reporting overhead in cost of dollars? Maybe should use something more concrete, like token usage.
+
+Section 4: Why not cite baseline methods?
+
+How are suitability, logicality, informativeness, reliability, and overall quality calculated in the results?
+
+The real-world experiment sounds limited. What is the average and standard deviation for the problems that the students voluntarily selected to be tutored? And what was the distribution like amongst the 6 conditions? Why no ANOVA test results?
+
+Line 1314 in Appendix L.4.: "FIANL" should be "FINAL"
+
+### Questions
+* Have LLMs really generated attention in education because of their extensive knowledge base and reasoning capabilities?
+
+* Figure 3: Why does "Explanation" lead to "Explanation" in the slow-thinking? Why does "Decomposition" have a crown?
+
+3.2: "introduce a successor-first strategy, in which the teacher prioritizes assessing leaf nodes or nodes whose successors have already been evaluated," I thought line 175 said "a student can only master a child node after mastering its parent"?
+
+3.3: "dual system theory" citation? lines 215, 246
+
+Line 264: Is this what you are proposing? Simulated teaching tree? Or is that something that already exists in the literature?
+
+Line 270: Does m = |S| (i.e., the length of the strategy pool)? If not, why?
+
+Section 4: Why not cite baseline methods?
+
+How are suitability, logicality, informativeness, reliability, and overall quality calculated in the results?
+
+The real-world experiment sounds limited. What is the average and standard deviation for the problems that the students voluntarily selected to be tutored? And what was the distribution like amongst the 6 conditions? Why no ANOVA test results?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The authors introduce PELICAN, a large language model–based tutoring framework to deliver personalized education. It is designed to give adaptive feedback to students. It has two primary components: cognitive diagnosis and adaptive tutoring. The cognitive diagnosis uses a hierarchy of concept knowledge graph and uses an expert–assistant–verifier to generate diagnostic questions. Adaptive tutoring creates explanations and strategies based on the diagnosed cognitive state. It uses a dual-system strategy: fast and slow thinking to tailor based on the necessity of student. By extensive experiments across synthetic and real students, the authors show that PELICAN can be highly effective for LLM-driven personalized tutoring that mirrors human pedagogical behavior.
+
+### Strengths
+The primary strength of the paper lies in the clarity of its presentation. The authors take an intuitive idea to adapt LLMs to humans and show such system can work in practice. The two major components in PELICAN have been clearly written and motivated. The authors explain the importance of each and every component in the training pipeline. For example, the cognitive modeling component uses a hierarchical knowledge tree, with a well-defined curriculum that traverses from the leaf to the root. In order to ensure robust question quality during diagnosis, Expert–Assistant–Verifier Pipeline. Finally, the authors explain the importance of both short and fast thinking in adaptive tutoring and how they are decided based on the diagnosed cognitive state. By clearly showing improvements on both synthetic and real world students, the authors show the efficacy of their proposed tutoring sysem.
+
+### Weaknesses
+As such, I don't have many concerns with the paper. Please find some of my questions regarding the experiment setup.
+
+a) How fine-grained are the knowledge concepts in the hierarchy? Furthermore, does the system allow more granularity in the concepts, depending on the student?
+
+b) The expert-assistant-verifier pipeline depends on the accuracy of the two LLMs involved. Do the authors conduct ablations on what fraction of the diagnostic questions are noisy? And how much is the performance of PELICAN affected by the noise?
+
+
+Furthermore, the authors should discuss on the diversity of the students involved in the human study. More than just performing well, did the students also report favorable experiences when interacting with LLMs? For example, a hyperparameter that can be controlled is M, which controls the number of times the LLM switches from fast to slow thinking. Is M decided based on whether the student progresses in the problem?
+
+### Questions
+Please see above for my questions.
+
+### Soundness
+4
+
+### Presentation
+4
+
+### Contribution
+4
+
+### Rating
+8
+
+### Confidence
+4

human_reviews/NlHHlqP1zk.md

@@ -0,0 +1,139 @@
+## Human Reviewer 1
+
+### Summary
+This paper evaluates the capabilities of (multimodal) LLMs to produce fine-grained concept annotations. Prior work either relied on human evaluations or used downstream performance as a proxy. The key idea of this paper is that if annotations are sufficient, a LLM should be able to infer the correct class from the annotation alone (Definition 3.1).
+
+The authors investigated concept annotations of (multimodal) LLMs through several stages of annotation refinements (l. 234-240). A so-called fast mode, where the model directly predicts the class from the input image (like in standard object recognition). In the so-called slow mode, the model progressively predicts concept annotations from coarse (e.g., background or superclass) to fine-grained levels (e.g., visual attributes such as the shape or color of bird’s wing).
+
+The experiments show that (multimodal) LLMs generally fail to produce concept annotations that are sufficiently accurate or informative to enable correct class inference from them.
+
+### Strengths
+* S1: The finding that using more fine-grained annotations (slow mode) hurts the correct class inference for fine-grained data like CUB is surprising and interesting.
+
+* S2: The concept annotation sufficiency definition is quite interesting. It captures well what we want concept annotations to look like.
+
+### Weaknesses
+* W1: A core issue with the evaluation framework is that a LLM is used to infer the class. There’s no guarantee that this LLM is good in inferring the correct class from the annotations.
+
+* W2: The claim that “strong performance in downstream tasks may not correlate with adequate conceptual supervision” (l. 462/463) is too strong. Also, I’d like to note that these models often weigh the concepts or only use a subset of them on a per-instance basis. Thus, the conclusion seems not warranted. It’d be more meaningful to take, say, the top-5 concepts per image and try to see if models can infer the correct class.
+
+* W3: The finding that (multimodal) LLMs work not that well to capture fine-grained details is well-known. The paper doesn’t provide a novel insight regarding this point.
+
+* W4: The paper’s writing could overall be improved for clarity.
+
+### Comments
+
+* C1: The explanation of “utility-as-proxy assumption” is very unclear and would be good to refine for better clarity.
+
+* C2: Comparing the 0-shot classification performance to the concept annotation is likely not that fair. Because, for example, these models might have seen the dataset, so the fast-mode numbers might be therefore a too optimistic estimate. However, I understand that this might be hard to control for the authors and, thus, only listed it here as a comment.
+
+### Questions
+* Q1: Limiting the number of classes to four plus the correct class seems very constraining. Why did the authors choose this setting?
+
+* Q2: How would concept annotators that also include distinguishing similar classes compare? The prompts (Appendix B) don’t focus on this aspect and therefore the models might yield annotations that can be confused with other classes.
+
+* Q3: Regarding Fig. 3: What makes post-hoc textual annotations so much stronger than visual-grounded annotations?
+
+* Q4: How do models compare on ImageNet in Tab. 3?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper investigates whether large language models (LLMs) and vision-language models (VLMs) can serve as reliable automated annotators for concept-based explainable AI (XAI) systems. 
+The work introduces the Fast and Slow Effect (FSE) framework, which evaluates the sufficiency of concept-class annotations without human supervision. 
+FSE models the annotation process as a continuum from a fast (opaque, intuitive) mode to a slow (explicit, conceptual reasoning) mode, and introduces a new metric, the Class Representation Index (CRI), to quantify whether generated concepts adequately represent the target class.
+Through experiments on fine-grained and general vision datasets (e.g., CUB-200, Cars-196, CIFAR-100), the authors find that current LLM-generated annotations are often insufficient, particularly in fine-grained classification. 
+Surprisingly, “slow” conceptual reasoning frequently underperforms the “fast” intuitive mode, suggesting that models possess implicit visual knowledge they struggle to express explicitly. 
+The paper also critiques the “utility-as-proxy” assumption, that higher downstream accuracy implies better annotation quality, showing this correlation breaks down under FSE analysis.
+
+### Strengths
+- The FSE framework is an original and well-motivated contribution that systematically examines annotation sufficiency, filling a notable gap in XAI evaluation methods.
+
+- The paper offers a convincing empirical analysis revealing a consistent gap between intuitive inference and explicit conceptual reasoning, raising important questions about the interpretability of LLM-driven explanations.
+
+- The paper is well-organized, clearly written, and transparent about procedures, datasets, and metrics, which enhances reproducibility and conceptual clarity.
+
+### Weaknesses
+- The definition of sufficiency is somewhat circular and lacks grounding in formal interpretability theory. It might mistake a good result for true understanding.
+
+- The five-stage annotation process (Background, Superclass, Salient, Detailed, Auxiliary) may heavily influence results. The robustness of findings to prompt variations or alternative hierarchies is unexplored.
+
+- Although the paper critiques “utility-as-proxy,” its proposed CRI metric is still accuracy-based, measuring recoverability of class labels rather than interpretive richness or human-understandable sufficiency.
+
+- The current research strictly limits its focus to image data. Therefore, its effectiveness is uncertain for other forms, such as text, tabular data, time series, or mixed-media reasoning.
+
+### Questions
+- How can CRI or FSE be extended to capture human interpretability, not just internal consistency?
+
+- Would the same phenomena occur in text-only domains (e.g., sentiment analysis explanations) or multimodal reasoning tasks beyond classification?
+
+- Can fine-tuning LLMs explicitly for conceptual disentanglement improve slow-mode sufficiency, or is this limitation architectural?
+
+- Is there empirical evidence that higher CRI correlates with human trust or usability of explanations? Without this, the framework’s practical XAI relevance remains uncertain.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper investigates the problem of explainable AI in deep learning, whether large language models (LLMs) and vision-language models (VLMs) can generate reliable concept-based annotations. While previous studies have shown that LLMs can produce plausible annotations, the semantic adequacy of these annotations is unclear. To address this, the paper introduces the Fast and Slow Effect (FSE) framework, an autonomous self-evaluation methodology. FSE transitions annotation from a fast mode to a slow mode and evaluates each stage using a Class Representation Index (CRI) metric, which measures how well accumulated concepts represent their target classes. Experiments conducted on various visual classification datasets demonstrate that current annotations generated by LLMs lack sufficiently conceptual depth and need more effective and transparent strategies.
+
+### Strengths
+1. The paper is well written, and its motivation is clear.
+2. This paper proposes a new explanatory framework that can automatically evaluate annotation sufficiency without human intervention. 
+3. The paper conducts experiments on fine-grained and common datasets with various model families and further re-examines the utility-as-proxy evaluation.
+
+### Weaknesses
+1. The paper initially discusses annotations for both LLMs and VLMs; however, the proposed methodology and experiments focus only on the explanation of vision tasks with VLMs, without evaluating textual explanation tasks or reporting results for LLMs.
+2. The reliability of the concept-chain gathering process and the CRI metric as accurate measures of “semantic sufficiency” remains unclear, as they have not been rigorously validated through human evaluation.
+
+### Questions
+1. How sensitive is the prompt design to different vision tasks? The current five-stage refinement process is evaluated only on general-domain datasets. How well does this template perform in generating visual concepts for more specialized domains, such as medical imaging?
+2. The concept generation process follows a hierarchical structure, for example, from background to detailed features. How does the order or number of steps in this concept chain influence the final results?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/OAXECnLxuk.md

@@ -0,0 +1,153 @@
+## Human Reviewer 1
+
+### Summary
+This work introduces a new MLLM for synthesizing scientific figures as TikZ graphics programs conditioned on raster images. To facilitate this, the authors introduce a highly curated dataset of 30k TikZ images. Compared to related work, this dataset might seem small, but the automated curation process yields higher-quality training examples that lead to better downstream performance. The authors also introduce a new set of task-specific reward functions that can further improve the model via GRPO. Their final model outperforms both proprietary and open-weight baselines on a wide range of automated metrics.
+
+### Strengths
+* The central contribution of the work is the dataset and its curation pipeline. Through manual analysis, the authors identify shortcomings of existing datasets and improve their training examples through automated methods. These findings will be useful for future work (such as injecting comments as an "inline" reasoning trace).
+* Likewise, the defined reward functions seem effective and scalable and will be useful for future work.
+* In general, the paper is well-written and easy to follow.
+
+### Weaknesses
+* The comparison between DaVinci and DeTikZify is not entirely fair, as DaVinci is based on a stronger base model. This means that conclusions about the effectiveness of the data curation strategy (l.375f) could also be due to this difference.
+* Even though the dataset is a central contribution, its creation is not entirely clear. For example, in l.154, the authors mention that they originally collect 360k TikZ snippets. It is not clear at which step these examples are filtered down to only 30k examples.
+* While the findings are interesting, there are few technical novelties presented.
+
+### Questions
+* In l.61, Belouadi et al. use the MCTS algorithm primarily to find outputs with higher perceptual similarity, not to improve the compile success rate.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper addresses the timely topic of diagram-to-code generation by proposing a two-stage learning method: a supervised fine-tuning (SFT) "cold start" followed by GRPO post-training. It also explores two overlooked features, code ordering and descriptive comments. However, the paper's key findings lack support from rigorous experimental data, leading to ungrounded claims about the technical method's efficacy. Moreover, the dataset construction and evaluation methodology present a potential risk of bias, indicating that the work requires further improvement before its conclusions can be considered reliable.
+
+### Strengths
+S1. The proposed two-stage learning approach with SFT + GRPO is a timely fit for the field.
+
+S2. This work draws attention to the often-overlooked features of code ordering and descriptive comments in diagram-to-code generation.
+
+S3. The paper is well written, and the methodology is solid.
+
+### Weaknesses
+W1. The main concern lies in the data construction process, which contains only 30k samples, whereas previous work (DATIkZ v1: 117k, v2: 360k, v3: 456k) used much larger datasets. Though the authors provide more details in Appendix C, key information is still missing. In line 795, they refer to 225k samples, but after stratified sampling by token length, only 58k remain (line 978), and it is unclear how the final 30k samples (line 149) were selected. Moreover, stratified sampling by token length may introduce bias, making the proposed dataset focus more on the long-tail data of the original dataset.
+
+W2. We encourage the authors to provide a concise and clear description of the data construction process in the main content, showing how 366,075 (line 764) valid pairs were reduced to the final TikZ-30k dataset. It is also necessary to provide comprehensive statistical analyses of the dataset at each step to ensure that no bias is introduced.
+
+W3. Bias in the evaluation method is another critical weakness. In line 338, the authors mention sampling 542 instances from 456,000 instances in DATIkZ v3. This small-scale test set cannot provide statistically convincing results, and the unclear sampling method raises concerns about potential distortion. Moreover, the authors should include evaluation metrics consistent with previous studies for better performance comparison.
+
+W4. The reviewer is curious about the “error-free precise feedback” (line 076); however, the authors only introduce it as “a method” (line 263) without experimental support or technical details.
+
+W5. Format/typos: 
+- In Figure 1, the Tikz code is changing from “yellow” to “red” while the text is describing “from green to light red” in the top right part.
+- In Table 1, did not bold the best score for SigLIP
+
+### Questions
+See weaknesses
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This work addresses the problem of parsing raster-based scientific figures into TikZ code. The authors introduce TikZ30K, a high-quality dataset constructed through extensive preprocessing using large language models (LLMs) and vision-language models (VLMs). The data curation pipeline includes low-quality data removal via VLMs, code reordering via LLMs, and comment injection to enhance planning capabilities. A Qwen2.5-VL multimodal LLM is fine-tuned using a two-step training pipeline consisting of supervised fine-tuning (SFT) followed by reinforcement learning (RL). The authors design novel reward functions evaluating textual information, geometric elements, image fidelity, and rendering success rate. The trained model is evaluated on a range of proprietary and open-source MLLMs, demonstrating that the SFT+RL approach outperforms baselines such as GPT-5 on automatic metrics.
+
+### Strengths
+- The data preprocessing pipeline is novel and intuitive. The use of VLMs for labeling and filtering, LLMs for code reordering, and comment injection for planning is well-motivated. Ablation studies confirm that reordering improves performance.
+- The reward design is comprehensive, capturing textual, geometric, and rendering aspects. The inclusion of ablations on different reward settings is appreciated.
+- The analysis provides interesting insights, particularly regarding reasoning vs. non-reasoning behavior and the observation that code similarity decreases while performance improves after RL. Evaluation includes both proprietary and open-source MLLMs.
+
+### Weaknesses
+- It would be helpful to include a code efficiency metric, such as average token or character length, to better understand the complexity and verbosity of the generated TikZ code.
+- There is no information about the data timeframe or sourcing, which raises potential concerns about data contamination or leakage.
+- The lack of human evaluation is a major limitation. Relying solely on automatic metrics makes it difficult to assess the real-world quality and fidelity of generated figures. Even a small-scale human evaluation would substantially increase the credibility of the results and help correlate human judgments with automatic metrics.
+
+### Questions
+- What is the exact difference between Reorder30K and TikZ30K? Is it simply the inclusion of comments introduced during data curation?
+- For DaTikZ-V3, is the preprocessing pipeline the same as TikZ30K? Also, is TikZ30K a subset of DaTikZ-V3, or was it sourced independently? Clarifying the dataset relationships and data collection process in the main paper would improve transparency.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper presents a two-stage learning framework for diagram generation using TikZ, consisting of supervised fine-tuning (SFT) followed by reinforcement learning (RL).
+In the first stage, the authors investigate the effects of drawing order and comment placement, aspects that have not been thoroughly examined in previous studies.
+In the second stage, they introduce a composite reward function in the RL phase that incorporates multiple factors, including image quality, code correctness, and structural consistency, and demonstrate its effectiveness through experimental results.
+
+### Strengths
+- The authors analyze previously overlooked properties of the data itself—specifically, the order in which data are generated and the use of comments within the code—and demonstrate their positive effects.
+- They further propose a method for defining the RL reward that incorporates text alignment, bounding-box consistency, and element matching directly extracted from the vector images.
+
+### Weaknesses
+- Although a new dataset has been constructed, there is no description of its license information, making the details of the dataset’s licensing unclear.
+- The evaluation relies solely on automatic metrics, and no human evaluation is conducted. Since the proposed automatic metrics partially overlap with the reward function used in RL, verifying through human evaluation whether the improvement is genuinely meaningful—or merely a result of metric hacking—would greatly enhance the credibility and reliability of the results.
+
+### Questions
+- In Section E.3 of the paper, which describes the training details, it is mentioned that the ViT and MLP projector are frozen during the SFT stage but unfrozen during the GRPO stage. Did you experiment with both configurations—freezing and unfreezing—in the GRPO stage, and find that unfreezing these components led to better performance?
+- Will the model, code, and dataset be released to the public?
+- The dataset is referred to as TikZ30K, but according to Appendix C.4, there are 58,000 training samples. Were the data further filtered during the optimization or post-verification steps? Does this mean that the final dataset actually consists of 30,000 samples?
+- The code augmentation process does not seem trivial for an LLM. The paper mentions that the procedure was repeated when rendering failed or when the rendered image was dissimilar to the original. On average, how many attempts were required per sample to obtain a valid code conversion? Moreover, were all samples successfully converted into valid code, or were some data filtered out due to unsuccessful conversions?
+- Could you please clarify which specific tools or libraries were used to extract the elements from the PDF files when computing the geometric similarity reward?
+- In the computation of the geometric similarity reward, I assume that matching the corresponding elements is not a trivial task. Could you please elaborate on how reliable the matching results are when inspected by humans? In other words, to what extent does the proposed algorithm produce visually or semantically reasonable correspondences between elements?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/OPFE1zPYbU.md

@@ -0,0 +1,195 @@
+## Human Reviewer 1
+
+### Summary
+The paper analyzes the diffusion and flow matching models and argues that the learning objective suffers from sparsity in higher dimensions.
+
+### Strengths
+I think that the problem of analyzing diffusion/flow models' objective is an important one.
+
+### Weaknesses
+The papers has the following strong weaknesses:
+1. **Lack of proper positioning/context in the literature**. The authors make a very strong and false claim in the introduction that they "present the first rigorous analysis of the diffusion model objective in high-dimensional sparse scenario". This problem is a known one and an active area of research. See for example [1] and references therein. The authors do not acknowledge, cite or compare with the existing literature on the topic. Furthermore, a large portion of the proofs (e.g. appendix A) are known results and the authors do not explicitly say that, potentially misleading readers that these are new results. Similarly, the entire discussion in section 2 about the similarities between diffusion and flow matching is well known, and nicely explained e.g. in the blog post: https://diffusionflow.github.io/. The authors should be more explicit that all of this is already known and properly cite related works.
+2. **The paper lacks mathematical rigour**. A lot of statements are vague and sometimes I genuinely did not know what the authors were trying to convey. For example
+    1. what does 'variance is relatively fixed' mean?
+    2. The notation is also confusing. What does E
+    3. What does "Furthermore, due to limited sampling during training, each $p(x0|xt = Xt)$ cannot be sufficiently sampled, so the actual degradation ratio should be higher than the statistics show" mean?
+    4. What does "In high dimensions, each $p(x0|xt = Xt)$ should be complex" mean?
+    5. "It easily predicts non-submerged frequencies (likely copying them)." What are non-submerged frequencies?
+    6. "Since the model compensates for the submerged frequencies, it can also be regarded as an information enhancement operator." - I don't understand what this sentence means.
+    7. "Based on the principle of train-test matching". What is the principle of train-test matching?
+    8. The entire section 4.1 lacks mathematical rigour. I_good and I_bad should be explicitly defined.
+    9. "The linear combination of independent noise is still noise"
+3. **There is no contribution**. Finally, the paper presents no contribution. They argue that the model objective is suboptimal. What they propose instead is "Self guidance", which in principle sounds similar to "Autoguidance" [2] (which is not cited) and "natural inference", but the authors provide no details on how it actually works. The entire section 4.2 is tough to understand. Figure 5 makes it more confusing rather than clarifying. Instead of rigorously defining what the authors propose with equations, they provide a list of bullet points that summarize the "core ideas". Finally, the authors do not provide any empirical results with their new framework, and how they compare with known sampling frameworks.
+
+Some more details:
+
+1. The proofs in appendix A:
+    1. Are imprecise - f_\theta is not defined. I assume that it's a function from R^D to R^D. But there are typos with missing norms, i.e. f_\theta^2 is undefined when f is not one-dimensional
+    2. All these results are well known are derived multiple times in various works. The authors don't mention this anywhere, and might mislead someone into thinking that these are novel results
+2. The discussion about flow matching in lines 138-152 is not correct in general. A flow matching model can be defined with various probability paths (and in particular, the prior distribution need not be Gaussian). It should be mentioned here explicitly that the authors only mean this simplified case (which is equivalent to a diffusion model with a certain noise schedule)
+3. Line 231. The reason for flow matching showing different statistics is that it uses a different noise schedule than VP. In other words, the "time" $t$ means different levels of noise for both models.
+4. "When weighted sum degradation occurs, it is equivalent to using a single sample as an estimator of the mean, which typically have large error". I have two issues with this statement
+    1. it is not true. The authors define degradation as the probability distribution over training data having an entry > 0.9. This is not equivalent to a dirac delta.
+    2. Secondly, we do the same thing with VAEs - we learn the model by only using a single sample, and it works fine in practice.
+5. "If we cannot provide an accurate fitting target, we argue that the model is unlikely to learn the ideal target accurately"
+    1. The question is: do we want to lean the ideal target? The ideal target will never generate any unseen data, because of the form of Eq14. So it is not really desirable to try to learn the ideal target.
+    2. Therefore, in practive, diffusion models certainly do not learn the ideal target, and it's good, because they are able to generate unseen examples. An interesting question is why that happens.
+6. Figure 5 is very difficult to parse. Some undefined symbols appear: a_t, b_t, c_t, y_t, \hat{a}_t.
+
+---
+References:
+
+[1] Lukoianov et al. "Locality in Image Diffusion Models Emerges from Data Statistics" (NeurIPS 2025)
+
+[2] Karras et al. "Guiding a Diffusion Model with a Bad Version of Itself" (NeurIPS 2024)
+
+### Questions
+1. In tables 1 and 2, for time=600, it seems like the "degradation" is much more pronounced than "degradation to x0". This is not intuitive to me. Are you saying that after adding some amount of noise I am much more likely to be closer to a different training example than the one I started with, and the weighted sum is still degenerate?
+2. "When training a model to predict X0 from noise-mixed samples, the model prioritizes frequencies based on their SNR"
+    1. Is this claim proven anywhere in the paper?
+3. "This frequency-dependent process is confirmed during inference: early steps (large t) generate contours, while later steps (small t) add details."
+    1. I think this is indeed considered general knowledge in the field, but I don't see the authors providing evidence of this claim, nor citing any works that demonstrate this
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This article posits that diffusion models are fundamentally limited in their ability to learn high-dimensional distributions with sparsity. To address this, the authors introduce a natural inference process that serves to unify existing methodologies.
+
+### Strengths
+- **Significance**: This article concerns the problem of diffusion models where the data distribution is sparse in high dimension, which could be a critical problem in the real-world applications.
+
+### Weaknesses
+- **Soundness**:
+  - Unknown facts: The authors demonstrate that in high-dimensional sparse scenarios, machine learning models can not effectively learn complex hidden probability distributions and their essential statistical quantities, but they do not show the references or provide other evidence. 
+  - Equation (15) is unclear, since the expectation is taken over $x_{t}$ rather than $x_{0}$. 
+  - Figure 1 seems to be incorrect, since $x_{t}$ can be at positions near other data points other than $x_{0}$. 
+- **Completeness**:
+  - Lack of experiments in the major text. 
+  - Lack of a related work section that contains the relevent references of this study. 
+- **Reresentation**:
+  - Redundancy: (1) In Sec. 1, the descriptions of assumptions made for diffusion models, flow-matching models; (2) the proofs for the equivalence of score-matching objective and denoising score matching objective in Sec. A.1 is not necessary, since it is well-known and not made by the authors. 
+  - Formatting Distractions: The overuse of emphasized text (e.g., excessive bolding or italics) biases the reader's attention and disrupts the flow of the narrative.
+  - Inproper figures: (1) Figure 5 is overly cluttered with mathematical symbols and equations that are not essential to its point and are not adequately explained in the caption or main text.; (2) all figure captions are missing a period at the end. 
+  - In Sec. 3.1, the text reads "Eq. equation (1)" instead of the standard "Eq. (1)". This formatting error should be corrected throughout the manuscript.
+
+### Questions
+- How can this study be applied in the real-world applications?
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper presents an analysis of diffusion models training objectives and its degradation to learning from a single sample in a low signa-to-noise regime. The paper also proposes a framework to unifies existing inference methods.
+
+### Strengths
+The paper proposed an interesting perspective to understand the training dynamics of diffusion models; in particular, how different frequences are picked up during different time and how the model's learning objective gradually shifts to focusing on a single sample as the signal to noise ratio increases. There are some empirical evidence in Table 1 that supports the authors' claim.
+
+### Weaknesses
+- The paper's language is very vague and the analysis not rigorous. A major claim in the paper is that diffusion models "cannot effectively learn the essential statistical quantities of the underlying data distribution, including the posterior, score, and velocity field", there are not rigorous mathematical arguments that support this claim. In fact prior work (e.g. [1]) have characterized the convergence properties of diffusion models. How would the proposed statement fit in this existing literatures on diffusion model theory?
+
+- Many "supporting" arguments are not original and not surprising. For example, the "weighted sum degradation" behavior is expected: as the likelihood signal gets stronger, the posterior concentrates on the most consistent sample. 
+
+- The purpose of Section 4 on A UNIFIED INFERENCE FRAMEWORK -NATURAL INFERENCE is unclear. It appears more like an interpretation of existing inference methods. It is not clear what new  insights are drawn, or whether there are new inference methods that can be proposed and improve upon the existing ones. 
+
+
+References
+[1] Chen, Sitan, et al. "Sampling is as easy as learning the score: theory for diffusion models with minimal data assumptions." arXiv preprint arXiv:2209.11215 (2022).
+
+### Questions
+See Weaknesses section.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper challenges the conventional probabilistic interpretation of diffusion models. The authors argue that, in high-dimensional and sparse data regimes, the objective of diffusion models “degrades” — instead of learning statistical quantities (posterior, score, velocity field), the model merely learns to predict the original clean sample $x_0$ from its noisy counterpart $x_t$.
+
+The paper presents both analytical reasoning and empirical statistics to support this claim. Specifically:
+
+* The authors analyze the posterior $p(x_0|x_t)$ and show that, in high-dimensional spaces, its mean (normally a weighted sum over all samples) collapses to a single dominant sample — the so-called **Weighted Sum Degradation** phenomenon.
+* Empirical measurements on ImageNet-256 and ImageNet-512 show nearly 100% degradation for small $t$ ($t < 600$).
+* Building on this, they propose a “Natural Inference Framework” that unifies multiple sampling algorithms (DDPM, DDIM, DPM-Solver, DEIS, etc.) as simple autoregressive signal combinations without invoking probabilistic concepts.
+
+The authors claim this provides a new, intuitive, and non-statistical understanding of diffusion model training and inference.
+
+### Strengths
+* The new perspective is interesting
+
+* The proposed “Natural Inference” formulation offers an intuitive, signal-processing-style view of sampling.
+
+### Weaknesses
+* The **main claim is overstated and unsupported**: posterior concentration does not invalidate the probabilistic formulation of diffusion models.
+* The “Natural Inference Framework” is **only a re-expression** of known samplers in linear form, offering no new theory, algorithm, or empirical advantage.
+* No experimental evidence demonstrates that this perspective improves understanding or performance.
+* The argument conflates numerical concentration with conceptual failure of probabilistic mode
+
+### Questions
+* In Section 3.1, $p(x_0)$ is represented using an empirical distribution over training samples. This is not a common practice in diffusion literature. Could you clarify whether there are existing works that also adopt such an empirical formulation as the basis for theoretical derivation?
+
+* While many deterministic formulations of diffusion models exist, they generally do **not** reject the probabilistic foundation. What is the motivation or added value of explicitly denying the probabilistic perspective?
+
+* I agree that in modern machine learning there are cases where probabilistic reasoning can be dropped—for example, linear regression can be derived either from a Gaussian noise assumption or purely geometrically from minimizing MSE. However, methodological proposals should ideally be grounded in **simple and principled intuition**. The probabilistic formulation of diffusion models has such a foundation through *score matching*. Does your proposed “Natural Diffusion” framework have an equally simple and principled starting point?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+3

human_reviews/OUQ8kLRK3m.md

@@ -0,0 +1,171 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes DRE-Bench, a cognitive-level benchmark for abstract reasoning, showing that while reasoning-focused LLMs outperform general ones, all models still fail on higher-level sequential and conceptual tasks, revealing a significant gap from human fluid intelligence.
+
+### Strengths
+**Cognitive alignment:** Tasks are designed based on the psychological cognitive hierarchy (attribute, spatial, sequential, and conceptual) to structuredly evaluate the abstract reasoning ability of LLMs. This structured assessment is reasonable and meaningful.
+
+**Dynamic robustness:** Generating diverse tasks by varying parameters to avoid data leakage while revealing the stability and generalization ability of the model under different levels of complexity, which is novel.
+
+### Weaknesses
+This paper only uses the all-right/all-wrong grid matching criterion, ignoring cases where some reasoning is correct or intermediate steps are reasonable, limiting the fine-grained analysis of model capabilities.
+
+### Questions
+1. see weaknesses
+
+2. In your experiments with supplementary visual information, you found that text-only input outperformed both single-image and multi-image formats. I'm curious: what happens if you combine visual input with Chain of Thought (CoT) cues? Would there be a performance boost?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces DRE-Bench, a dynamic reasoning evaluation benchmark designed to assess the fluid intelligence of LLMs. The benchmark consists of 36 abstract reasoning tasks organized across four cognitive levels, with each task featuring multiple dynamic variants. The authors evaluate a range of LLMs, including general-purpose models and reasoning-specialized models. Experimental results reveal that while most LLMs perform well in low-level cognitive tasks, they struggle with high-level cognitive tasks, exhibiting limited generalization capabilities.
+
+### Strengths
+Proposes an abstract reasoning framework based on cognitive levels.
+Designs verifiable code generators and solvers to ensure data quality and scalability.
+
+### Weaknesses
+Inadequate Human Experiment Design:
+Alignment with human cognition is the core advantage of this Benchmark. However, the handling of this critical aspect is very weak in this paper. Although a human comparison experiment is provided, there is a lack of detailed age distribution and significance testing. Additionally, there is no explanation of how participants were motivated to complete the questionnaire seriously or how invalid responses were excluded. For the complex task of designing cognitive test questionnaires, the paper seems to lack rigor; neither the main text nor the supplementary materials provide detailed explanations of how the questionnaire design excludes potential influencing factors. This leads to the conclusion "validates the justification of our 4-level framework" appearing unreliable. This point is particularly important because if the foundation of the human experiment is unreliable, the entire Benchmark's cognitive alignment will be questioned, thereby affecting its validity as a Benchmark.
+
+Lack of Detailed Analysis of Model Failure Modes: Specifically, error types in high-level cognitive tasks are not thoroughly explored, failing to reveal the specific reasons for model failures in complex tasks.
+
+Key Findings Lack Depth: The paper emphasizes five key findings, but these findings do not demonstrate particularly outstanding innovation or in-depth analysis, appearing rather superficial.
+
+Insufficient Analysis of Visual Information Impact: This is one of the key challenges in human and LLM testing. However, the analysis of how visual information affects model performance is overly simplistic, providing results for only two visualization formats and lacking deeper exploration.
+
+### Questions
+Why were only 20 annotators selected as the sample for the human comparison experiment? This number appears particularly insufficient given the multitude of potential influencing factors in this experiment. Since this is a questionnaire-based test, recruiting more participants would not have been an expensive or labor-intensive task, yet it could have significantly improved the reliability and statistical robustness of the results.
+
+Is there sufficient statistical significance to support the conclusions?
+How was the seriousness of participants and the validity of the questionnaire ensured? Were any incentives or quality control measures implemented?
+Were potential cognitive biases or other influencing factors considered during the questionnaire design process? How were these interferences excluded to ensure the reliability of the experimental results?
+What are the specific failure modes of the models in high-level cognitive tasks? Are there certain types of errors that repeatedly occur?
+
+### Soundness
+1
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper introduces a large-scale abstract reasoning benchmark structured along a four-level cognitive hierarchy (Attribute, Spatial, Sequential, and Conceptual). The framework is designed to evaluate Large Language Models (LLMs) on tasks of increasing cognitive complexity and to assess whether current models exhibit good fluid intelligence. Through systematically generated tasks and extensive experiments, the authors find that model accuracy declines with higher cognitive levels, especially for physics-related conceptual reasoning. Reasoning-optimized models (e.g., OpenAI-o1, DeepSeek-R1) outperform general-purpose LLMs, and the performance gap widens as task complexity increases.
+
+### Strengths
+- It is an interesting finding that the models achieve higher and more consistent accuracy in vertical (up/down) directions than in horizontal (left/right) ones in Move. Similarly, in symmetry tasks, performance is better for horizontal symmetry than for vertical symmetry
+-  The authors implement a verifiable, scalable data engine capable of generating diverse reasoning tasks with controllable complexity—an important methodological contribution that ensures reproducibility and adaptability.
+- The results highlight key cognitive distinctions: (a) model accuracy degrades sharply at higher cognitive levels, (b) inference-time scaling contributes mainly to lower-level reasoning, and (c) visual input may not aid abstract reasoning. These insights deepen our understanding of the limits of current LLM cognition.
+
+### Weaknesses
+- Several aspects of the experimental design and presentation require clarification and stronger consistency.
+ the selection of models across figures lacks transparency and methodological coherence. For example, Figure 5 does not specify why those particular four models were chosen, Figure 6 focuses solely on DeepSeek-R1 without justification, and Figure 7 switches to yet another subset of models while testing only two tasks. 
+
+Such inconsistent model selection makes it difficult to assess whether observed relationships are generalizable. A more systematic comparison across models, task types, and difficulty levels would provide stronger evidence for the claimed trends. Table 2 also introduces new models without explanation
+
+- question design and ground-truth validity raise concerns. In Figure 8, certain problems might have multiple correct solutions (e.g., the Level-3 task where the path could start from one red square to the upper blue square not only the bottom one), yet only one is labeled as correct. Additionally, the prompt mislabels the target as a “red dot” instead of a “red square.”
+
+ In the Level-4 example, two obstacle types exist, but the prompt fails to specify that only the blue obstacle is reflective, potentially confusing both human and model respondents. These ambiguities undermine the reliability of the evaluation.
+
+- discrepancies between text and data should be addressed. 
+- Line 340 claims that human accuracy decreases with task level, but Table 1 shows relatively stable human performance across Levels 1–3, except for a few specific tasks (e.g., sorting). Interestingly, these same tasks are also where models perform poorly. This suggests that further validation is needed to ensure the generated items reflect increasing cognitive difficulty rather than artifacts of task design.
+
+### Questions
+- What criteria guided the selection of models in Figures 5–7 and Table 2? Were these choices based on availability, performance tiers, or specific architectural features?
+- Why was DeepSeek-R1 uniquely highlighted in Figure 6?
+- How many tasks and difficulty levels were used to derive the inference-time–accuracy relationship, and do these results hold across other models or scales?
+- In Figure 8, how were ground truths defined for multi-path problems, and were alternative valid solutions considered during evaluation?
+- How were ambiguous or mislabeled prompts (e.g., “red dot” vs. “red square,” unspecified reflective obstacles) handled in scoring and analysis?
+- Given that human performance appears stable across the first three levels, how to justify the statement that accuracy “generally decreases” with increasing cognitive level?
+
+Suggestions:
+
+- Unify model selection and reporting. clearly specify which models are used in each experiment and maintain consistency across figures. If subsets differ, explain the rationale.
+
+- Broaden the inference-time analysis. test multiple models across several tasks and difficulty levels to confirm that observed correlations are not model-specific.
+
+- Validate and disambiguate tasks. review prompts and ground truths to ensure each question has a unique, well-defined solution and that instructions are precise.
+
+- Re-evaluate human baselines. conduct additional validation to verify that increasing task complexity indeed reflects higher cognitive difficulty, rather than inconsistencies in dataset generation.
+
+- Clarify inconsistencies between text and data.
+
+### Soundness
+1
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+2
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This work introduces DRE-Bench, a benchmark designed to evaluate dynamic reasoning through a hierarchical cognitive framework including 36 reasoning tasks. It frames fluid intelligence as the ability to generalize beyond memorized knowledge and reason, and designs rule-based code generators and solvers for each task. Empirical evaluations across different LLMs show that while models perform well on lower-level reasoning, they struggle with higher cognitive levels, indicating a gap from true fluid intelligence.
+
+### Strengths
+1. The task design is well aligned with cognitive psychology, clearly reflecting different types of reasoning and the varying levels of intelligence required for each task.
+
+2. The proposed benchmark is a valuable contribution to the evaluation community, offering diverse and controllable challenging tasks.
+
+3. The paper provides a thorough evaluation and analysis across multiple LLMs, shwoing how accuracy and stability vary with task complexity and cognitive level.
+
+### Weaknesses
+While the paper proposes a cognitively inspired hierarchy of reasoning tasks, it does not measure the time humans take to solve them. Given the small-scale human evaluation and potential variance, reporting the average solving time per task would better substantiate the claimed cognitive alignment and reveal the true difficulty gradient.
+
+### Questions
+In Table 1, why does o3-mini perform significantly better than other models on the Level-4-Mechanics tasks, achieving 31.75% accuracy while most other models show near-zero performance?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/Q5SSA6IonA.md

@@ -0,0 +1,151 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes Vision Filter (ViF), a novel vision backbone that replaces Transformer self-attention with a Fourier Neural Filter (FNF). Built upon the Fourier Neural Operator (FNO), FNF introduces adaptive modulation and selective activation to dynamically couple spatial and frequency-domain representations, addressing FNO’s limitations in modeling high-frequency local details. Experiments on classification, detection, and segmentation show that ViF achieves competitive or superior performance to Swin Transformer, ConvNeXt, and VMamba while maintaining quasi-linear complexity.
+
+### Strengths
+1. The paper provides a clear theoretical motivation.
+2. The introduced adaptive modulation and selective activation mechanisms effectively mitigate the over-smoothing and bandwidth bottleneck issues of conventional FNOs.
+3. Experiments on classification, detection, and segmentation tasks demonstrate consistent improvements over strong baselines such as ConvNeXt, Swin, and VMamba, confirming the model’s effectiveness and generality.
+
+### Weaknesses
+1. The ablation study is relatively limited, providing insufficient analysis of the independent contributions of the two key FNF submodules. A more detailed exploration of how each component individually influences performance would strengthen the empirical validation.
+2. The paper lacks interpretability analysis. Although the authors claim that adaptive modulation adaptively scales different frequency components in the spectral domain, the explanation remains purely mathematical without concrete visualization or empirical evidence. It is unclear how the modulation behaves across different images or tasks and whether any consistent adaptation patterns exist.
+3. The comparison with recent state-of-the-art methods are missing, such as MLLA and OverLoCK.
+4. In Table 3, the performance improvement under the Mask R-CNN 3× MS schedule is marginal. The paper does not discuss possible reasons for this weak gain.
+
+### Questions
+See weaknesses.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This manuscript proposes a novel visual backbone network called Vision Filter (ViF). Its innovation lies in the introduction of a core component, the Fourier Neural Filter (FNF). Building on the Fourier Neural Operator (FNO), FNF introduces an input-dependent adaptive integral kernel. Through selective activation and adaptive modulation, it establishes a dynamic information flow between the frequency and time domains, addressing the bandwidth bottleneck and over-smoothing issues inherent in FNO. ViF outperforms current mainstream Transformer, Mamba, and Fourier-based models on ImageNet-1K image classification, COCO object detection, and ADE20K semantic segmentation tasks, while maintaining high computational efficiency.
+
+### Strengths
+1.The manuscript 's motivation is sound, and it clearly identifies the limitations of FNO, such as bandwidth bottlenecks and over-smoothing.
+2.This manuscript introduces an input-dependent integral kernel into the Fourier operator, constructing a unified time-frequency representation space, which is quite innovative.
+3.An obvious advantage of this method is computational efficiency. Compared with Transformer-type models, it has lower computational complexity and is more advantageous than Mamba-type models in spatial structure modeling.
+4.This manuscript achieves state-of-the-art or competitive results on multiple mainstream vision tasks, particularly on ImageNet-1K, significantly outperforming similar Fourier methods (such as GFNet).
+
+### Weaknesses
+1.Although FNF solves some of the shortcomings of FNO, its core idea is still based on the traditional framework of "frequency domain processing + time domain supplementation". It does not propose a new visual representation learning paradigm, which weakens the innovation of this article.
+2.Although ViF performs well on ImageNet-1K, its improvement on COCO and ADE20K compared to models such as ViM is relatively small (e.g., mAP increases by 0.2-0.4), failing to significantly widen the gap.
+3.ViF only tested three small and medium-sized variants, ViF-T/S/B, and did not evaluate the performance of larger models, such as those of ViT-L/16, on larger datasets.
+4.Definition 7 mentions that adaptive modulation is achieved through the amplitude-sensitive weighting function \(\|z\|\) and learnable parameters \(\alpha\) and \(\beta\), but does not give the range of \(\alpha\) and \(\beta\) and whether they are adjusted with training iterations.
+
+### Questions
+As mentioned above.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes a theoretically sound and empirically validated vision backbone that bridges local and global modeling through the frequency domain, offering an alternative to attention-based and state-space models. The FNF formulation is well-motivated and results across benchmarks substantiate its benefits.
+
+### Strengths
+1.The paper presents a theoretically grounded extension of the Fourier Neural Operator (FNO), introducing an input-dependent integral kernel that enables adaptive and dynamic coupling between the spatial (time) and frequency domains.
+2. The proposed Vision Filter (ViF) integrates both local convolutional和global frequency operations within a single framework.
+3. Results in Figure 1 demonstrate that ViF achieves a more favorable accuracy–throughput trade-off compared to both Transformer and Mamba-based counterparts, showing practical efficiency for high-resolution image processing.
+
+### Weaknesses
+1. The paper does not evaluate ViF on larger datasets (e.g., ImageNet-22K, LAION) or with pretraining strategies such as self-supervised learning (MAE, BEiT). As such, the scalability and transferability of ViF remain uncertain compared to modern foundation models.
+2. The ablation study (Table 5) confirms that each module contributes to performance, but lacks deeper interpretability or visualization (e.g., frequency response maps, energy spectrum analysis). More qualitative analysis could help clarify how ViF balances local and global representations in practice.
+3. Although the authors cite GFNet and AFNO, the distinction between ViF’s adaptive gating mechanism and prior dynamic spectral filters could be elaborated further.
+
+### Questions
+1.The manuscript seems only to swap the entire Multi-Head Attention block in ViT for the proposed FNF module, but never compares the two under an otherwise identical architecture. An ablation is recommended that exchanges only the attention mechanism.
+2.Although the authors claim “quasi-linear complexity”, only FLOPs are reported. No end-to-end latency, throughput, or peak-memory measurements versus Swin, VMamba or GFNet on the same GPU/CPU are provided.
+3.FNF introduces several new hyper-parameters (α, β in adaptive modulation, number of Fourier modes, kernel size of local convolutions, etc.). The manuscript gives a single setting without any hyper-parameter sensitivity analysis. It is indiscernible whether the performance gains are statistically robust or an artefact of ad-hoc tuning.
+4.The claim that “directional scanning inevitably leads to spatial disruption” in Mamba-based models is not substantiated. The manuscript neither quantifies the degree of spatial-information loss nor compares patch-order robustness between Mamba and the proposed FNF. It is necessary to compare Acc of Mamba and FNF after random patch-shuffling to quantify patch-order robustness.
+5.The implicit periodic extension inherent in FFT introduces spectral leakage and degrades the accuracy of the frequency-domain filtering operation when the input dimension is not a power of two or when the image contains strong edges. This work has not examined whether the proposed FNF module suffers from such boundary artefacts, nor has it evaluated remedies such as windowing or reflection padding. A systematic comparison of cyclic, symmetric and zero-padding schemes is required to determine the susceptibility of FNF to spectral leakage and its consequent impact on dense prediction tasks.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes a new generic vision backbone called Vision Filter (ViF) , built upon a novel Fourier Neural Filter (FNF) module. The authors argue that while the Fourier Neural Operator (FNO) is promising for its global modeling capabilities and quasi-linear complexity , it suffers from an "over-smoothing" effect and a "bandwidth bottleneck," making it difficult to capture local, high-frequency patterns. To address this, the FNF module introduces two key components: 1) Adaptive Modulation (AM) in the frequency domain to enhance sensitivity to high-frequency components , and 2) Selective Activation (SA) to balance local time-domain and global frequency-domain information flow. This design creates an input-dependent integral kernel. Extensive experiments on image classification (ImageNet-1K) , object detection (COCO) , and semantic segmentation (ADE20K) demonstrate that ViF outperforms existing SOTA Transformer- and Mamba-based backbones.
+
+### Strengths
+1. The authors introduce an adaptive kernel that effectively couples time- and frequency-domain information, providing theoretical novelty.
+2. The paper demonstrates strong empirical performance across multiple visual benchmarks and model scales.
+3. The authors conduct comprehensive ablation studies that clearly quantify the contribution of each component (AM, SA, LC).
+4. The proposed ViF achieves a favorable balance between model complexity and performance, offering quasi-linear computational efficiency.
+
+### Weaknesses
+1. The paper lacks experiments exploring the scalability of ViF on larger datasets such as ImageNet-22K or LAION.
+2. The authors do not evaluate ViF under self-supervised or foundation model pre-training settings, which limits its generalization claims. I believe this is crucial for vision backbone models.
+3. The paper lacks detailed frequency-domain visualizations (e.g., spectral activation maps) to empirically support its frequency modeling claims.
+4. The improvements over strong Mamba variants on dense prediction tasks are relatively marginal compared to the architectural complexity.
+5. The authors have not released any implementation code at submission time, reducing reproducibility.
+
+### Questions
+In addition to the above weaknesses, I also have several questions and suggestions for the authors:
+1. Could ViF be integrated into existing architectures (e.g., ConvNeXt or BEiT) as a modular replacement, and what challenges might arise?
+2. Does the proposed FNF introduce any implicit bias toward certain frequency bands, and how is this mitigated?
+3. Are there plans to extend FNF to video or multimodal vision tasks?
+4. Could the authors provide additional visualization results illustrating spectral activations or learned frequency responses?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4

human_reviews/Q6TymStAvS.md

@@ -0,0 +1,176 @@
+## Human Reviewer 1
+
+### Summary
+This paper introduces ShadowFM, a generative framework for learning quantum ground states from their classical shadows. The core idea is to apply geometric flow matching on non-Euclidean manifolds inspired by the Bloch sphere, rather than using standard Euclidean approaches. By respecting the intrinsic geometry of quantum measurements, the proposed methods, Spherical Flow and Anisotropic Dirichlet Flow, compared to vanilla flow matching, achieve more accurate predictions of physical observables like correlation functions.
+
+### Strengths
+The paper introduces a novel framework by applying geometric flow matching to learn classical shadows of quantum states. It thoughtfully motivates this approach by connecting the non-Euclidean structure of quantum measurements to the Bloch sphere, a departure from prior works that assumed Euclidean geometry.
+
+The proposed "Spherical Flow" and "Anisotropic Dirichlet Flow" methods are empirically validated on the Transverse-Field Ising and Heisenberg models.
+
+### Weaknesses
+1. The major weakness is the paper's structure and writing. A significant portion of the early sections is dedicated to explaining well-established concepts, which may be inefficient for an expert audience at ICLR conference.
+- Sections 2.1, 2.2, 2.3 and Section 3 provide lengthy introductions to Classical Shadows and flow Matching. While context is necessary, these concepts are foundational and could likely be summarized more concisely, perhaps by focusing only on the specific aspects directly built upon by the authors. This would free up valuable space to elaborate on the novel contributions. If the author desires to present more specific or self-contained content, detailed information (including Algorithm 1 and 2) can be placed in the appendix.
+- Furthermore, the motivation in Section 4.1 is good, but the preceding three pages of background and related works dilute the paper's focus and delay the reader's engagement with the key ideas. In a highly competitive venue, it's crucial to present the core innovation as early and clearly as possible. 
+
+A more effective structure might have been to briefly introduce the necessary concepts from FM and classical shadows within the introduction or a much shorter, combined background section, and then move directly to the motivation and detailed methodology of ShadowFM.
+
+2. The paper's core methodological sections (4.2 and 4.3) blend established theory with the authors' modifications. It would be better if the authors delineate their novel technical contributions from the baseline frameworks of RFM and DFM. As presented, it is difficult to isolate the exact innovations beyond the application of existing tools to a new domain.
+
+3. For the Anisotropic Dirichlet Flow, the increased methodological complexity does not consistently yield superior performance over the simpler Spherical Flow (e.g., in the Heisenberg model results of Table 3). What is the justification for this more complex model if its empirical advantage is not universally demonstrated across the tested problems?
+
+4. A critical observation from Table 1 is that the learning-free Classical Shadow baseline consistently outperforms all trained generative models, especially on the correlation metric and in the high-sample regime. Could the authors explain this performance gap and justify the practical utility of the generative approach if it fails to surpass a direct, learning-free estimation method?
+
+4. The primary motivation for introducing a non-Euclidean geometry is an experiment (Figure 2) suggesting that "spin errors" (e.g., $|X^{+}\rangle \rightarrow |X^{-}\rangle$) are significantly more detrimental than "basis errors" (e.g., $|X^{+}\rangle \rightarrow |Y^{\pm}\rangle \text{or} |Z^{\pm}\rangle$). However, the provided plot does not strongly support this claim. The performance gap between the two error types appears marginal across the tested error rates. The paper's assertion that spin errors are "significantly higher"  seems overstated. This weakens the central premise that a geometry specifically designed to maximize the distance between spin-flipped states is necessary and justifies the added model complexity.
+
+5. The paper's experiments compare the proposed methods primarily against other flow-matching variants. This comparison is too narrow and fails to benchmark against the true state-of-the-art in generative modeling for quantum states. Crucially, there is no comparison against machine learning models such as [1,2,3] and a benchmark [4], which have shown strong performance on similar tasks.
+
+[1] Huang H Y, Kueng R, Torlai G, et al. Provably efficient machine learning for quantum many-body problems[J]. Science, 2022, 377(6613).
+
+[2] Wang H, Weber M, Izaac J, et al. Predicting properties of quantum systems with conditional generative models[J]. arXiv preprint arXiv:2211.16943, 2022.
+
+[3] Yao J, You Y Z. ShadowGPT: Learning to Solve Quantum Many-Body Problems from Randomized Measurements[J]. arXiv preprint arXiv:2411.03285, 2024.
+
+[4] Zhao Y, Zhang C, Du Y. Rethink the Role of Deep Learning towards Large-scale Quantum Systems[C]. Forty-second International Conference on Machine Learning, 2025.
+
+### Questions
+The selection of "Generative Models" as the primary area suggests the paper's main contribution lies in advancing the fundamental methodology of generative modeling. While the paper does introduce novel geometric adaptations to flow matching, these innovations are exclusively motivated, designed for, and validated on the specific problem of learning quantum states. This raises a crucial question about the work's intended contribution and audience. The work might be more appropriately positioned within an area like "Applications to Physics".
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper introduces Shadow Flow Matching, a geometric and generative framework for modeling quantum measurement data. Rather than treating measurement outcomes as raw values, the authors embed them into a cross-polytope manifold that captures the symmetry and combinatorial structure of Pauli measurements, essentially a high-dimensional, geometry-aware analogue of the Bloch sphere. Within this space, they apply flow matching to learn distributions of outcomes conditioned on Hamiltonian parameters, allowing the model to follow the curved geometry of quantum shadows instead of a flat Euclidean one. This leads to smoother sampling and better reconstruction of observables like correlation matrices and entanglement measures, even under noise or limited data. By blending geometric insight with generative modeling, the work moves beyond treating quantum data as structureless, aligning machine learning more closely with the intrinsic geometry of quantum mechanics.
+
+### Strengths
+1. Mapping quantum measurements into a cross-polytope space is an original and elegant idea—it captures the symmetries of Pauli measurements while moving beyond standard Euclidean embeddings.
+
+2. The paper nicely links discrete measurement data with continuous geometric flows, bringing flow-based generative modeling into quantum ML in a fresh and promising way.
+
+3. The framework shows solid potential for tasks like shadow tomography, expectation reconstruction, and entropy estimation, where structure-aware modeling really helps.
+
+4. The theory and context are well presented, clearly explaining why a geometric approach makes sense and how it connects to broader machine learning and quantum ideas.
+
+5. The visualization of the cross-polytope flow is particularly interesting, helping to convey the intuition behind how flow trajectories respect geometric constraints and how they differ from flat-space generative flows.
+
+### Weaknesses
+1. Despite its novelty, the scope of practical applications appears somewhat narrow at this stage. The framework’s advantages are demonstrated primarily on specific shadow-based tasks, and it remains uncertain how easily the approach can scale to broader or more complex domains of quantum learning.
+
+2. A current limitation is the method’s dependence on the structure of Pauli shadows. Since the model architecture and training procedure hinge on these symmetries, adapting it to non-Pauli or arbitrary measurement settings could require substantial reworking of the geometric foundation.
+
+3. The empirical evaluation is somewhat underdeveloped. Comparisons are made to only one baseline, and while qualitative results are encouraging, they leave open questions about quantitative robustness and competitiveness against state-of-the-art generative quantum models.
+
+### Questions
+1. Are the current experiments confined to the 1D anti-ferromagnetic Heisenberg model, or has the framework been tested on other many-body systems?
+
+2. How does the proposed approach compare to diffusion-based quantum models, such as Generative Quantum Machine Learning via Denoising Diffusion Probabilistic Models? 
+
+3. Given the method’s dependence on Pauli-based shadows, could the framework extend to alternative measurement protocols like? Is the model’s success contingent on Pauli-specific symmetry?
+
+4. Could this geometric-flow framework be repurposed for broader quantum modeling tasks—for instance, simulating quantum dynamics?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper presents Shadow FM, which is a flow based method to generate Pauli POVM measurements of a quantum ground state. The Hamiltonian of the system can be used as a conditioning input to generate these samples. There are two approaches in the paper: (i) a spherical / Riemannian flow aligned with the Bloch-sphere geometry and (ii) an anisotropic Dirichlet probability path. The authors test their methods on Heisenberg and TFIM chains.
+
+### Strengths
+The paper is well writing and the flow matching framework that the paper uses is well motivated by the geometry of the quantum states. Being a non-autoregressive method is also a positive,  as the shadow distribution is unlikely to have 1D nature for interesting quantum states. The experiments are thorough and interesting.
+
+### Weaknesses
+1. All the experiments presented are for 1D models. This is a concerning as the ground states of 1D models have efficient classical representation in terms of MPS representations. 2D experiments would have substantially improved the quality of the results
+
+2. Motivation of restricting to ground states of Hamiltonians is unclear to me? Why not thermal states or states produced by real time evolution? I don't see the learning task itself using any information about the fact this is a ground state of a Hamiltonian.
+
+### Questions
+1. For learning ground states, could this method be enhanced by adding a variational component to the loss that also tries to minimize the estimated energy of the state?
+
+2. What is the main bottleneck faced by the authors to go to 2D experiments?
+
+3. In the phase transition studies, do the authors observe any changes in the behavior of the learning algorithm across the phase transition?
+
+4. How do the authors ensure that the distribution over shadow states that the learned model samples from for a new Hamiltonian actually correspond to a physically allowed quantum state? Can a projection step be built into the inference pipeline to project to physically allowed states?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper introduces ShadowFM, a geometric flow matching framework for learning ground-state quantum many-body wavefunctions via the distribution of classical shadows.
+Instead of modeling full quantum states, ShadowFM learns to generate shadow measurements—compact randomized representations of quantum states—and uses them to estimate physical observables such as correlation functions and entanglement entropies.
+
+### Strengths
+- Originality: the paper presented a novel framework called ShadowFM that combines low matching generative modeling with classical shadow tomography for learning quantum many-body ground states.
+- Quality: The technical development looks sound and link both geometric and quantum information theory. The paper demonstrates a good understanding of both and integrates them coherently. 
+- Clarity: The paper is clearly written, with a well-organized structure. Figures and tables of experiments look good. 
+- Significance: The work study a problem that is significant in bridging geometric deep generative modeling and quantum many-body learning. It proposes a scalable, data-driven alternative to classical shadow reconstruction and autoregressive models.
+
+### Weaknesses
+- Computational overhead: The anisotropic Dirichlet flow requires precomputing and integrating Beta-function–based terms, which could limit practicality; runtime and memory costs are not quantified in very detail.
+- Restricted empirical scope: Experiments are limited to 1D spin chains (TFIM and Heisenberg). The scalability and performance of ShadowFM on larger or higher-dimensional systems are not demonstrated.
+
+### Questions
+- Could the authors explain more about the scalability of their methods? For example, for Heisenberg model, could the experiments for $L=30$ be conducted? Also, what about other models?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+2

human_reviews/QIIrjgUnL1.md

@@ -0,0 +1,187 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes a position-aware attention mechanism designed to make the standard self-attention more sensitive to relative positions.
+It introduces a position-effect and an enhanced position-effect functions which modulates the attention weights according to pairwise positional distance.
+The goal is to combine the flexibility of self-attention with an explicit, parametric control over positional bias.
+The authors present detailed analyses, including theoretical derivations, simulated experiments on synthetic data distributions (random, structured, clustered, etc.), and consistency metrics comparing theoretical versus actual attention responses.
+They claim that this mechanism improves positional interpretability and stability relative to standard Transformers.
+
+### Strengths
+1.	Interesting idea: 
+The notion of explicitly controlling positional effects via a decaying kernels (Eqs, 1 and 6) is conceptually clean and connects attention to classical signal-processing intuitions about locality.
+2.	Analytical exploration:
+The paper includes extensive mathematical and empirical analyses of how the proposed positional kernel influences attention behavior.
+The “theoretical vs. actual” consistency metric is a thoughtful way to test whether the implementation aligns with the analytical definition.
+3.	Comprehensive appendix:
+The appendix contains a wealth of supporting materials — visualizations, parameter sensitivity analyses, and ablations — which demonstrate significant effort and curiosity about the model’s internal mechanics.
+
+### Weaknesses
+1.	The main text lacks clarity and structure.
+Many of the explanations and empirical results appear only in the appendix or not at all.
+The main sections barely reference the relevant figures, forcing the reader to guess which figure corresponds to a given discussion.
+This significantly reduces readability.
+2.	Use of undefined or double-defined formulations.
+For example the abstract relates to the parameters $\alpha$, $\beta$ and $\gamma$ without any ability of the abstract reader to understand what they represent…
+$P_{effect}$, $A_{ij}$, $V(i)$, $pos^*$ are all defined twice, how do I know which one you mean in the text. Give them a slight differentiator for example $P_{effect}^+$
+
+3.	Limited experimental validation.
+The study is conducted almost entirely on synthetic data. There are no evaluations on real tasks (e.g., language modeling or vision benchmarks), so the practical usefulness of the method remains untested.
+4.	Ambiguous methodology.
+The paper introduces metrics such as $pos_{actual}$ and $pos_{theoretical}$ but does not specify precisely how they are computed,  requiring the reader to reconstruct the procedure themselves. Specifically for these two, there is some reference in the appendix, but I could not find a clear definition. 
+5.	Literature positioning is weak.
+The paper references related work sparsely. Many strong prior efforts on relative positional encodings (Shaw et al. 2018; Dai et al. 2019; Press et al. 2021; Su et al. 2021) are either missing or only briefly mentioned. On the other hand, all the references in the Abstract makes it cumbersome to follow. It should be precise and short.
+
+### Questions
+1. How does this mechanism compare empirically with standard relative position encodings (e.g., RoPE) on real benchmarks?
+2. Can the method be integrated into standard Transformer architectures without major computational overhead?
+3. Why are key visual results not cited or summarized in the main body? Would the paper benefit from moving Figures A.1–A.8 into the main section?
+4. Line 105 – which matrix values do you refer to? I guess it is not $A_{ij}$ because these are all normalized…
+5. Line 109 - add “In Fig. 3”…  do so for all references to Figures
+6. Line 144 + Figure 6. “Position influence magnitude” was never defined (not also in the appendix AFAIK) 
+7. The same goes for other naming in Section 2.3
+8. Plenty of typos in citing. lines 035, 037, 039, etc.,
+9. Line 207 and Line 863 – what is the difference between $S(pos)$ and $V(i)$
+10. I’m not sure what does $I_j$ means. In line 215 I see it is torch.norm(data, dim=-1) is it a the norm over all $i$ values? 
+11. Eq. 7. $A_{ij} already include $P_{eff}$. Do we multiply it again by $P_{eff}$?
+12. Table 3. I’m not sure I understand. I don’t see any significant difference between architectures. Highest is better or lowest is better?
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+3
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces a position-aware attention mechanism designed to address the inherent limitations of conventional attention models, such as those introduced by Ashish Vaswani et al. (2017). The key technical contribution is a positional effect function parameterized by three coefficients: \alpha which controls positional influence intensity, \beta which governs spatial decay rate, and \gamma which compensates for over-attenuation at longer distances. The authors establish the mathematical properties of this function (continuity, differentiability, monotonicity) and propose an adaptive triple-attention architecture integrating position-, task-, and content-aware modules for dynamic weighting. Experiments reportedly demonstrate performance improvements in structured and clustered data scenarios.
+
+### Strengths
+1. The paper introduces a positional effect function that offers a formalized mechanism for incorporating positional influence into attention computation.
+
+2. It develops a triple-attention architecture that jointly models position, task, and content information, which is conceptually interesting.
+
+3. It defines quantitative evaluation metrics aimed at measuring attention distribution quality, contributing to more systematic evaluation of attention mechanisms.
+
+### Weaknesses
+1. The proposed position-aware attention formulation appears largely as a mathematical restatement of existing positional modulation techniques without a clear, theoretically grounded justification for why it should outperform established methods (e.g., Rotary Position Embedding (RoPE), ALiBi, or relative positional encoding by Peter Shaw et al. (2018)). A stronger theoretical motivation or comparative analysis is needed.
+
+2. The experimental evaluation is insufficiently comprehensive. There are no direct comparisons with strong baselines such as RoPE, ALiBi, or Shaw et al. on realistic benchmarks (e.g., long-document modeling, QA).
+
+3. The implementation details are incomplete or unclear. Important information about model architecture, parameter sizes, training setups, and hyperparameters are missing, which prevents reproducibility.
+
+4. Some core concepts lack precise definition or justification. For example: I_j in Eq. (3) and (4) is referred to as “information importance,” but its definition and computation are not explained. The proposed “consistency” and “ranking correlation” metrics are not well motivated or compared against established alternatives.
+
+5. The claim that the triple-attention architecture achieves superior performance is not strongly supported by the results in Table 3, where the improvements over other configurations are marginal.
+
+6. The description of experimental results is ambiguous: it is unclear whether larger or smaller metric values indicate better performance, and some tables lack sufficient explanation.
+
+### Questions
+1. Please include experimental comparisons with strong and widely recognized baselines (e.g., RoPE, ALiBi, Shaw et al. 2018) to substantiate the claimed advantages of the proposed method.
+
+2. Please provide a clearer and more rigorous definition of I_j and its role in the method. Also, please justify the choice of evaluation metrics or align them with established practices in the field.
+
+3. Please clarify how to interpret each evaluation metric (e.g., whether higher or lower is better) and provide more structured and detailed explanations of results.
+
+4. Please include comprehensive implementation details, such as model configurations, training setup, hyperparameters—to facilitate reproducibility and fair comparison.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper introduces a position-aware attention mechanism that explicitly models positional relationships in attention computation through a position effect function, parameterized by $\alpha$ (intensity) and $\beta$ (decay). It mathematically analyzes their properties (continuity, differentiability, monotonicity) and proposes an enhanced variant with an additional $\gamma$ coefficient to mitigate over-attenuation in long-distance dependencies. Built on this, the authors develop a triple-attention architecture incorporating position-aware, task-aware, and content-aware attention to achieve task-specific flexibility.
+
+### Strengths
+The quality of the Appendix seems better than the main text
+
+### Weaknesses
+1. It is not clear why the proposed position effect function addresses the limitations of traditional attention mechanisms by providing fine-grained control over positional relationships, as stated on line 94-96. Or what’s the advantage to existing methods like relative positional encoding and ROPE? If it’s only about modeling the dependency between $i, j$, ROPE also mathematically supports this, no?
+2. I am confused of section 2.2. What is the analysis subject to? On lines 105-106, the author states, “Specifically, α = 0.5 yields maximum values of approximately 0.5, α = 1.0 produces values of 1.0, α = 2.0 generates values of 2.0”. But there is no explanation of what the values impacted by the $\alpha$ values refer to.
+3. Experiment section is not clear enough in introducing the settings such as data, model, baselines.
+4. Eq.1 seems redundant as it is introduced again from line 92
+5. Typo, Line 357: we previously -> our previously
+
+### Questions
+Is it better to rearrange so that the analysis in section 2 appears later, say the experiment section?
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes a position-aware attention mechanism that extends traditional Transformer attention (Vaswani et al., 2017) by incorporating a novel positional effect function. The method builds upon relative position representations (Shaw et al., 2018) and rotary position embeddings (Su et al., 2021), introducing parameters α and β to control positional influence and spatial decay rate. To alleviate long-distance over-attenuation, the authors further introduce an enhancement coefficient γ and design an adaptive triple-attention architecture that integrates task-aware and content-aware modules for dynamic weight adjustment. The paper also provides theoretical analysis on the proposed function’s properties (continuity, differentiability, monotonicity) and introduces new evaluation metrics for consistency and positional benefit. Experiments demonstrate promising results on structured and clustered datasets, particularly for information retrieval and document understanding tasks.
+
+### Strengths
+Theoretical novelty: The paper provides a mathematically grounded extension to existing position encoding methods, offering interpretable control of positional influence and decay behavior.
+
+Innovative architecture design: The proposed adaptive triple-attention framework with task- and content-aware weighting is conceptually interesting and well-motivated by recent multi-scale attention work.
+
+Clear motivation and formulation: The paper is well-written and logically structured, clearly explaining the motivation behind each modification to the standard attention mechanism.
+
+Potential practical impact: The approach could improve transformer interpretability and adaptability in tasks where fine-grained positional relationships are important.
+
+### Weaknesses
+Limited empirical validation: While the theoretical contributions are strong, the experimental section is relatively weak. The evaluation mainly covers structured and clustered datasets, without sufficient diversity to support claims of general effectiveness.
+
+Computational analysis: No discussion on computational cost, convergence, or scalability compared to baseline attention mechanisms
+
+### Questions
+see weakness
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+3

human_reviews/QSPKIO3XV8.md

@@ -0,0 +1,206 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes a method to improve neural operator learning by incorporating a dimension- and domain-aware co-decoder architecture. The key idea is to decouple solution representations into dimension-specific components and domain-conditioned features, then fuse them to approximate the target PDE operator. The authors argue that this decomposition improves generalization and sample efficiency for solving PDEs in different geometric domains and dimensional settings. Empirical evaluation is performed on benchmark PDE problems, comparing against several neural operator baselines, showing lower relative L2 error in some test cases.
+
+### Strengths
+- Clarity of high-level motivation: The paper identifies a meaningful and real challenge — neural operator models typically struggle to generalize across domains or dimensional configurations. This is a relevant problem in scientific machine learning.
+
+- Clean architectural formulation: The proposed “co-decoder” idea is structurally simple and easy to integrate with existing neural operator backbones.
+
+- Readable exposition: The manuscript is well structured and technically consistent at the high level, with decent visualizations and clean mathematical notation.
+
+- Experimental setup uses multiple PDE benchmarks: This is a step beyond trivial toy problems, indicating an attempt at broader evaluation.
+
+### Weaknesses
+While the problem is important, the contribution is weak both conceptually and empirically:
+
+- Limited novelty:
+The central idea—decomposing and fusing dimensional and domain representations—is not fundamentally new. Variants of similar factorization and conditional embedding approaches have already been explored in operator learning, meta-PDE frameworks, and equivariant neural operator designs. The paper lacks a crisp theoretical justification or truly novel algorithmic insight. It feels like another architectural tweak rather than a new paradigm.
+
+- Insufficient baseline coverage:
+The paper only compares against basic operator learning baselines (e.g., FNO, DeepONet, or UNet-like variants). To substantiate the claims of improved generalization, it is essential to compare against stronger and more recent models, including: (1) Koopman neural operator as a mesh-free solver of non-linear partial differential equations, (2) Solving High-Dimensional PDEs with Latent Spectral Models. These methods are explicitly designed for high-dimensional PDE problems and constitute state-of-the-art baselines.
+
+- Superficial ablation analysis:
+The ablations only show minor performance differences when components are removed. This raises concerns about whether the proposed “co-decoder” mechanism truly drives the reported improvements or if the gains are marginal/random noise.
+
+- Lack of theoretical insight:
+The authors repeatedly claim that dimension/domain factorization improves generalization, but no formal analysis or empirical probing (e.g., transfer learning between dimensional settings, robustness to domain shifts) is provided to support these claims.
+
+- Limited scope of experiments:
+The experiments are narrowly focused, and it’s unclear whether the method scales or generalizes to more challenging PDEs (e.g., higher-dimensional or irregular geometries). There is no comparison on computational overhead, parameter efficiency, or training stability.
+
+- Unconvincing performance gains:
+Even within the provided benchmarks, the improvements are often modest and lack statistical rigor (e.g., no error bars, limited repetitions, unclear significance testing). This weakens the claim of superiority over baselines.
+
+- Overclaiming in narrative:
+The abstract and conclusion use strong language (e.g., “universal”, “robust generalization”, “significant improvement”) that is not backed by the presented evidence. This is a recurring issue in weak submissions.
+
+### Questions
+- Clarify the real contribution:
+What is fundamentally new about the co-decoder mechanism compared to existing conditional or factorized representations used in neural operator models? A clearer theoretical framing or architectural justification is needed.
+
+- Stronger experimental validation:
+Please evaluate on more challenging PDE families, including higher-dimensional settings and irregular domains. Include robustness and scalability analyses. Provide multiple runs with error bars to assess statistical significance.
+
+- Ablation rigor:
+Conduct deeper ablations: Test the effect of removing either the domain or dimension factor separately. Vary the embedding capacity to show necessity. Test sensitivity to the number of training domains.
+
+- Efficiency and complexity:
+What is the computational overhead of the co-decoder compared to baseline neural operators? If the architecture is more complex, does it provide any real benefit in efficiency or generalization to justify it?
+
+- Transfer/generalization tests:
+If the paper claims better generalization across domains or dimensions, please provide explicit cross-domain or cross-dimensional generalization experiments, not just single-domain evaluation.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This work combines several popular techniques to solve PDEs with PINNs, including tensor decomposition, domain decomposition, and Mixture-of-Experts (MoE). It is like a technique report not a research paper. The motivation of combining such techniques is not clear. I recommend rejection.
+
+### Strengths
+Using a neural network inspired by tensor decomposition is a good idea. In fact, several recent studies have explored solving PDEs with tensor-based neural network architectures.
+
+### Weaknesses
+1. The proposed framework simply adds a Mixture-of-Experts (MoE) layer to a dimension-decomposition backbone. This combination seems arbitrary and lacks a clear reason or theoretical justification. The MoE layer makes the model more complex but does not show any real advantage over a well-tuned single network or simpler decomposition methods. The claimed “automatic” domain decomposition is not properly compared with existing, simpler domain decomposition approaches used in PINNs, so its benefit remains unclear.  
+
+
+2. The reported improvements in accuracy and efficiency are small compared with standard PINNs or other PINN-based methods. The added model complexity and training cost are not justified by the limited gains shown on low-dimensional test problems.
+
+### Questions
+see the Weaknesses part.
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes Dimension Domain Co-Decomposition (3D), a unified framework for solving partial differential equations (PDEs) based on Physics-Informed Neural Networks (PINNs). The framework jointly performs dimension decomposition, which factorizes the solution into dimension-wise components using a shared MLP for parameter efficiency, and domain decomposition, implemented through a Mixture-of-Experts (MoE) mechanism that automatically partitions the solution space into subregions without pre-defined boundaries. In addition, the paper introduces a novel quantitative interpretability metric, Variable Interpretability (VI), which measures how well the learned dimension-wise latent components align with ground-truth factors. Experimental results on several PDE benchmarks (Poisson, Wave, Burgers, and Linear Transport equations) demonstrate that the proposed method achieves improved accuracy, scalability, and interpretability compared to standard PINNs and prior decomposition methods.
+
+### Strengths
+1. The paper elegantly combines dimension decomposition and adaptive domain decomposition into a single framework, addressing both scalability and local adaptivity issues in PDE solving.
+
+2. The introduction of the VI metric provides a quantitative way to assess per-dimension interpretability, an aspect rarely considered in PINN literature.
+
+3. The introduction of the shared MLP is well motivated and effectively reduces the parameter count, especially in high-dimensional settings.
+
+4. The evaluation includes multiple types of PDEs (elliptic, hyperbolic, nonlinear), with detailed parameter analysis on rank $r$ and expert number $K$, providing empirical evidence.
+
+### Weaknesses
+1. The paper presents a clean combination of dimension decomposition, MoE-based domain decomposition, and a variable interpretability metric. While the empirical results are solid, the theoretical innovation is limited：Both parameter sharing and MoE are well-established techniques. The paper would benefit from a clearer articulation of the core technical challenge or insight specific to this setting, and from theoretical analysis connecting the VI metric to the proposed model’s design to make the contribution more solid.
+
+2. The experiments do not include comparisons with state-of-the-art PDE solvers such as Fourier Neural Operator (FNO) [1], which are highly relevant baselines.
+
+3. Ablations that remove the dimension factorization or the MoE router are necessary to substantiate the co-decomposition claim.
+
+4. Although the paper tests a high-dimensional (10D) Poisson equation, the PDEs considered remain relatively simple. The robustness and generality of the proposed 3D framework would be more convincing if tested on more complex or non-analytically-solvable systems.
+
+[1] Fourier Neural Operator for Parametric Partial Differential Equations.
+
+### Questions
+1. Have the authors evaluated the consistency and robustness of the learned domain decompositions？
+
+2. The paper highlights parameter and memory efficiency, but does not report actual training time or convergence comparisons. Could the authors provide quantitative evidence to support the claimed computational advantages of 3D?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper proposes a unified Dimension Domain Co-Decomposition (3D) framework that integrates dimension decomposition and Mixture-of-Experts (MoE)–driven domain decomposition for solving high-dimensional partial differential equations (PDEs). The key contributions include:
+1. A shared-MLP dimension decomposition architecture that processes coordinate–index pairs, reducing parameters and improving scalability.
+2. A Variable Interpretability (VI) metric, which measures the alignment between learned latent dimension representations and ground-truth components.
+3. A MoE-based automatic domain decomposition, which adaptively partitions the computational domain without predefined regions or interface constraints.
+
+The framework demonstrates improved accuracy, efficiency and interpretability across several PDE benchmarks. However, certain aspects of the theory and experiments require further clarification. If the authors address these points in their rebuttal, I would be willing to raise my score.
+
+### Strengths
+1. **Originality**.
+The introduction of the VI metric is a creative and measurable approach to interpretability in PDE learning. The unified framework that combines shared-MLP design for multi-dimensional inputs with MoE based domain decomposition is also an innovation that enhances scalability.
+
+2. **Clarity**.
+The paper is generally well-structured, with clear explanations of the architecture, mathematical formulation, and experimental setup. Figures effectively illustrate both the decomposition mechanisms and domain partitions.
+
+3. **Reproducibility**.
+The paper provides detailed implementation settings, including network architectures, training procedures, and hyperparameters, ensuring high reproducibility.
+
+### Weaknesses
+1. **Lack of theoretical justification for the VI–convergence relationship**.
+The paper implicitly proposes an important but unproven proposition: if for all dimensions $j$, $VI_j \to 1$, then the model prediction $\hat{u}$ converges to the true separable solution $u$.
+   Currently, this claim is only supported empirically, where high VI values are observed to correlate with low prediction errors. However, no formal mathematical proof is provided to substantiate this relationship.
+
+2. **Fixed number of experts $K$ limits adaptivity and may cause compromise responses**.
+The number of experts $K$ is manually chosen rather than learned adaptively. When the PDE solution contains multiple fine-grained or highly localized structures (e.g., shocks, vortices, or high-frequency regions), a fixed small $K$ may fail to provide adequate domain resolution. Consequently, the router’s softmax weighting may average across distinct subregions, diminishing the local specialization property of the Mixture-of-Experts and leading to compromise responses. This design constraint restricts flexibility for more complex or multi-scale problems.
+
+3. **Router–expert coupling risks instability and expert starvation**.
+The strong interdependence between the router and experts introduces potential training instability. If an expert achieves slightly lower error early in training, the router may continuously reinforce its weight allocation, causing other experts to receive negligible gradients—a phenomenon known as expert starvation. In problems with rapidly evolving or multi-shock dynamics, delayed router adaptation can further trigger oscillatory expert switching and degrade convergence stability.
+
+### Questions
+1. Is it possible to provide a rigorous mathematical justification for VI?
+
+2. During training, was a multi-stage training strategy used, where the routing is fixed first and then the experts are updated, in order to improve overall convergence stability?
+
+3. According to the discussion in Section 3.2, VI is closely related to the separability of the ground truth and the properties of the matrix itself. Why, then, in Table 2, is VI correlated with frequency?
+
+4. In Table 2, the 10-dimensional Poisson requires a smaller $r$ than the 5-dimensional Poisson, which seems somewhat counterintuitive. Conventionally, higher-dimensional problems typically require a larger $r$ to adequately represent multi-dimensional separable structures.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3

human_reviews/QnuJR7qA3z.md

@@ -0,0 +1,192 @@
+## Human Reviewer 1
+
+### Summary
+This paper highlights deployment issues with layers like GELU, Softmax and LayerNorm present in transformer models due to the use of expensive operations such as div, exponential. Present solution either focuses on simplified approximation (GELU→ReLU) and/or hardware design specifically to efficiently run one or more of these blocks on hardware. Both of these solutions are limited in terms of their effectiveness and/or generalization. The Author proposed a unified framework (HARA-Hybrid Arithmetic-ReLU Networks Approximation)  to decompose these functions into simple arithmetic operators and an optimized parameter initialization pipeline for these approximations to work. The results are demonstrated by showing validation accuracy and hardware estimation for both FP32 and INT8 models (BERT, LLaMA, Swin, Stable Diffusion) 
+
+Overall, this paper is well written and proposed solution to a very important problem related to deployment of transformers on different devices. The paper has good contribution and has potential be a good starting for others to work on top of this work. Results are conclusive (except the hardware benchmark) and shows that this method can solve both the issue of hardware complexity and quantization issues related to transformers.
+
+### Strengths
+- Authors have targeted a much needed topic in a timely manner as transformers are gaining a lot of popularity but their deployment is still limited to GPU devices and many layers need to be reworked to make it suitable for other devices.
+- Approximating power/compute hungry operators has been proposed in the literature and the Authors have compared their results with few of the those (especially LUT based methods)
+- The dynamic programming based initialization is a novel concept. Any model optimized with a better initialization strategy has a potential to converge fast.
+- Comprehensive results on four models and comparison with other SOTA methods makes the claim stronger.
+- Detailed mathematical explanation of approximation is a strong point.
+- Most important contribution is to include hardware synthesis results that shows the power and area reduction achieved using HARA
+
+### Weaknesses
+- One of the main limitations (also mentioned by the Authors) is that all the benefits are based on the hardware synthesis and the impact of these changes on RAM and latency can be significantly higher (or lower).
+- Authors are also encouraged to see the impact of these changes on other domains such as computer vision using ViTs.
+- Minor : Citation for RMSNorm and LayerNorm for the first time is missing L037
+
+### Questions
+- How does hardware synthesis translate to real RAM/latency improvement comparing to existing architectures (a qualitative discussion would suffice) 
+- Can you include the ablation study to prove that the training initialization strategy results in better accuracy of the model (end to end model evaluated on a dataset). Table 4 includes ablation study but it is based on MSE calculated from a sample ? Please clarify how the MSE is exactly calculated and if it is based on a sample, it will be good to do ablation study of validation accuracy of end to end model.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper presents HARA (Hybrid Arithmetic-ReLU Approximation), a unified framework designed to approximate diverse non-linear activation and normalization functions using a single ReLU-based computational unit. The core motivation is that AI accelerators often require dedicated hardware units for different non-linear operations, leading to inefficient area and power utilization. HARA addresses this by proposing a unified approximation architecture that leverages a ReLU-centric design combined with a dynamic programming–based parameter initialization pipeline to enable accurate functional approximation across various nonlinearities. The authors claim that HARA achieves ≤0.1% accuracy degradation on representative workloads (BERT, Swin, LLaMA, Stable Diffusion) while providing substantial projected area and power savings compared to baselines with specialized functional units for each operation.
+
+### Strengths
+- The paper introduces a clean, generalizable framework (HARA) that unifies diverse nonlinear operators (e.g., GELU, Softmax, LayerNorm) under a single reconfigurable ReLU-based architecture.
+
+- The unified operator maintains accuracy within <0.1% of the baseline, demonstrating excellent approximation fidelity.
+
+- Hardware synthesis projections show significant area and power savings (≈62% area, 52% power) compared to separate specialized LUT units
+
+### Weaknesses
+- Results rely on synthesis-based estimations only, with no end-to-end performance, latency, or energy measurements on real hardware (FPGA, or ASIC prototype) (noted as limitation) or simulation. Ideally, the results for area and power reported should be post place-and-route.
+
+- Evaluations cover only standard precision settings and moderate workloads; there is no stress testing under long-sequence LLMs, mixed precision, or extreme numerical conditions.
+
+- The paper does not compare against reconfigurable functional units (RFUs) or FPGA based designs that already provide similar arithmetic flexibility.
+
+- Models and datasets are reasonable but small. Ideally, to convincingly portray the benefits of the method a large range of model sizes must be employed with reports of accuracy and inference performance metrics.
+
+- Evaluation omits key runtime baselines such as GPU fused kernels or optimized accelerator designs.
+
+### Questions
+Technical Concerns/Questions and Points to Address in Rebuttal:
+
+- Missing citations for related work in lines 119-128.
+
+- I suggest having an extended related work section in the appendix having a round-up of quantization techniques etc. particularly the ones focused on similar co-design such as [1], [2] and [3].
+
+- Area, power results must be post place-and-route and not post synthesis.
+
+- Provide range-reduction proofs and error bounds, plus catastrophic-case tests (very peaky logits; near-zero variance in LayerNorm; half-precision under/overflows).
+
+- Larger model evaluations needed. BERT and Swin are not really exciting.
+
+- The paper does not analyze how quantization affects the approximation quality of the learned nonlinear mapping. 
+
+- HARA is positioned as inference-time replacement. There’s no evidence on fine-tuning stability with HARA in the loop, or on co-training to reduce approximation error.
+
+- Quantization setup (scales, accumulators, rounding) is not detailed, limiting reproducibility.
+
+- A key limitation is, this work does not compare inference throughput with GPUs or other existing accelerators. I'd suggest the authors to do a comparison along the lines of [1], [2] using either GP-GPUSim or (if possible) real GPU results to truly evaluate how well HARA performs compared to GPUs. While the authors note the difficulty in physical implementation in the limitation section, I feel it is not unreasonable to do simulation as prior work have done to get preliminary performance evaluation. Not showing any performance results is a red flag and if performance results cannot be shown the authors must wait and re-submit a more complete work in the next iteration.
+
+- More recent baselines must be identified for comparison. 
+
+- How is HARA different from a scheme where you have a general-purpose computational unit with all primitive operations such as add, sub, mult etc and we can use that same unit by changing the routing to create a reconfigurable unit to perform any non-linear operation ? Please do area/power comparison with such a unit.
+
+- When operating at low precision, why not simply employ small LUTs for each non-linear function at sub–8-bit resolution? Wouldn’t that be a more straightforward approach? Because, at lower-precision such as 4-bit  the LUT size is fairly small.
+
+- The URN structure likely involves intermediate activations with addition/multiplication accumulation. The bit-width of these accumulators (e.g., 8-bit vs. 16-bit partial sums) is never stated, making it impossible to evaluate numerical error growth or overflow risk.
+
+- Is the 6nm by TSMC ? 
+
+
+References:
+
+[1] Ramachandran, A., Kundu, S., & Krishna, T. (2025, June). Microscopiq: Accelerating foundational models through outlier-aware microscaling quantization. In Proceedings of the 52nd Annual International Symposium on Computer Architecture (pp. 1193-1209).
+
+[2] Guo, C., Tang, J., Hu, W., Leng, J., Zhang, C., Yang, F., ... & Zhu, Y. (2023, June). Olive: Accelerating large language models via hardware-friendly outlier-victim pair quantization. In Proceedings of the 50th Annual International Symposium on Computer Architecture (pp. 1-15).
+
+[3] Sharma, H., Park, J., Suda, N., Lai, L., Chau, B., Kim, J. K., ... & Esmaeilzadeh, H. (2018, June). Bit fusion: Bit-level dynamically composable architecture for accelerating deep neural network. In 2018 ACM/IEEE 45th Annual International Symposium on Computer Architecture (ISCA) (pp. 764-775). IEEE.
+
+### Soundness
+3
+
+### Presentation
+4
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes HARA (Hybrid Arithmetic–ReLU Approximation Networks), which replaces diverse and power-hungry nonlinear operations with simple arithmetic primitives combined with a shallow ReLU network. In this way, a single hardware platform for HARA can flexibly support a wide range of nonlinear operations while maintaining efficiency.
+
+### Strengths
+- A single hardware framework for HARA can support various nonlinear operations, offering architectural flexibility.
+
+### Weaknesses
+- Several critical details necessary for fully understanding HARA are missing.
+  1. The size of the ReLU network (e.g., hidden dimension) is not provided, making it difficult to estimate the processing cost of HARA.
+  2. The conversion mechanism from GeLU to HARA is not described.
+  3. The storage and memory-access overhead for the ReLU network weights is not adequately discussed. Conventional nonlinear operator implementations (including LUT-based approximation methods) do not require external memory access, so the latency and energy implications of fetching ReLU network weights should be analyzed for a fair comparison.
+  4. The parameter tuning cost for HARA (via dynamic programming and fine-tuning) is not fully discussed.
+- This paper does not clearly describe the hardware resource requirements of conventional LUT-based approximation methods or their impact on inference accuracy. For example, NN-LUT [1] employs a simple LUT with only 16 entries to approximate nonlinear functions, achieving low hardware cost while maintaining accuracy. Although increasing the number of LUT entries could further reduce the mean squared error (MSE), such improvement may not necessarily lead to meaningful gains in inference accuracy. Hence, improving MSE at the expense of hardware efficiency may not be justified. Consequently, directly comparing MSE values between HARA and conventional LUT-based approximation methods may not represent a fair evaluation. A more comprehensive assessment, including hardware area, power overhead, processing latency, and inference accuracy, would be necessary to establish the practical advantages of HARA.
+
+[1] Yu, Joonsang, et al. "NN-LUT: Neural approximation of non-linear operations for efficient transformer inference." Proceedings of the 59th ACM/IEEE Design Automation Conference. 2022.
+
+### Questions
+Please check the weaknesses.
+
+### Soundness
+2
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper presents HARA (Hybrid Arithmetic–ReLU Approximation), a unified framework for replacing all non-linear operations in Transformers (e.g., GELU, Softmax, LayerNorm) with a single hardware-efficient ReLU–arithmetic module. The core idea is to use a dynamic programming–based parameter initialization pipeline that systematically finds near-optimal breakpoints for piecewise linear approximation, converts them analytically into ReLU parameters, and fine-tunes them. Experiments show that HARA achieves orders-of-magnitude lower MSE compared to NN-LUT and RI-LUT, while maintaining almost identical end-to-end performance with various transformer-based models.
+
+### Strengths
+- The paper is well written, detailed, and clearly organized. The methodology and experiments are described systematically, and results are easy to reproduce.
+- Compared to NN-LUT and RI-LUT, HARA significantly reduces approximation error (often by several orders of magnitude), leading to more robust end-to-end accuracy across multiple Transformer models.
+
+### Weaknesses
+- **Methodological contribution seems incremental**: While the paper frames its contribution as a unified framework with an optimized DP-based initialization pipeline, the core technical novelty beyond prior work such as NN-LUT or RI-LUT appears limited. The main differentiator seems to be the use of dynamic programming for systematic breakpoint selection and analytical ReLU parameter conversion, which is a well-motivated but relatively moderate algorithmic improvement. If I have misunderstood the extent of this difference, clarification on how HARA’s optimization pipeline fundamentally departs from previous LUT-based or neural approximation methods would be helpful.
+
+- **Lack of strong empirical motivation**: The paper does not convincingly establish that non-linear operators are the primary computational bottleneck in modern Transformer inference, especially compared to attention or matrix-multiplication components. While the proposed framework demonstrates excellent numerical approximation accuracy, it remains unclear how much practical speedup or memory benefit these approximations bring at the end-to-end system level. In addition, the experimental scope is somewhat narrow—evaluations focus on a small set of tasks and omit runtime or latency measurements that could justify the broader motivation.
+
+ - The work’s focus and contributions are primarily hardware- and implementation-oriented, with less emphasis on learning dynamics or algorithmic understanding. Therefore, the paper may align more naturally with a hardware or systems venue (e.g., DAC, MLSys) rather than ICLR, which typically emphasizes conceptual or methodological advances in machine learning.
+
+### Questions
+See weakness
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3

human_reviews/RNe17WTR38.md

@@ -0,0 +1,180 @@
+## Human Reviewer 1
+
+### Summary
+This paper proposes a self evolution framework in which a single language model acts as both generator and verifier to construct preference data for self improvement. The method introduces two variants, SimpleGV (single turn verification) and RevisionGV (multi turn verification), and applies iterative DPO. The authors evaluate the approach mainly on mathematical reasoning benchmarks.
+
+### Strengths
+The paper is clearly written and conceptually easy to follow.
+
+### Weaknesses
+**Limited novelty.**\
+The proposed framework of self-generation and self-verification is not new. Similar iterative preference learning or self-rewarding approaches have already been explored in prior works such as [1,2,3,4,5]. The paper differs only in minor details (e.g., thresholded verification, curriculum scheduling), which do not constitute meaningful methodological novelty. It also overlaps heavily with [6] in both structure and training procedure.
+
+**Lack of proper comparison and discussion.**\
+The paper does not compare against the most relevant prior works on self-improvement or self-rewarding LMs. There is also little discussion on how the proposed method differs conceptually or empirically from existing iterative preference optimization frameworks.
+
+**Weak experimental validation.**\
+The experiments are narrow in scope (e.g., training on a single dataset, using only two models, and evaluating mostly on math relaative easy reasoning tasks). There are no experiments on other domains (e.g., coding) or stronger reasoning benchmarks (e.g., AIME, AMC). As a result, the claims of general self-evolution remain unsubstantiated.
+
+**Cost inefficiency.**\
+The generator–verifier framework requires multiple generations and verification passes, increasing computational cost significantly. However, the observed performance gains are modest, suggesting poor cost–performance trade-off.
+
+**Overall.** The method appears to be a minor extension of existing self-reward/self-evolution frameworks, without demonstrating clear advantages or generality.
+
+Reference\
+[1] Self-Rewarding Language Models\
+[2] Self-consistency preference optimization\
+[3] Meta-Rewarding Language Models: Self-Improving Alignment with LLM-as-a-Meta-Judge\
+[4] CREAM: Consistency Regularized Self-Rewarding Language Models\
+[5] ARIES: Stimulating Self-Refinement of Large Language Models by Iterative Preference Optimization\
+[6] ReVISE: Learning to Refine at Test-Time via Intrinsic Self-Verification
+
+### Questions
+See the weakness above.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+1
+
+### Rating
+2
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper studies generator-verifier games where the same instruction‑tuned model acts as a generator that proposes multiple answers and a verifier that judges whether each answer is correct. Preference pairs are then formed and used to fine‑tune the model with DPO. The paper studies two variants: SimpleGV, which involves single‑turn, verifier as a judge with thresholded majority voting, and RevisionGV, which uses multiple turns, generator revises responses using verifier feedback. The method is evaluated on synthetic KK puzzles and math benchmarks, with strong KK improvements.
+
+### Strengths
+* Timely problem: Overall the problem the paper aims to solve is good. Reducing dependence on human labels or domain-specific verifiers is very important, and a single model generator/verifier is very a simple formulation.
+* Clear Writing: The paper does well at explaining how they form preference pairs and provides concrete settings. I found this easy to follow.
+* Ablations: The ablations on KK were really useful and important giving a sense of how the synthetic task behaves.
+
+### Weaknesses
+I think this paper needs additional evaluations to be stronger and seems unfinished. I list my concerns below:
+
+* Improvements on Real Datasets: The paper finds really small improvements on datasets like GSM-8K or MATH-500 in comparison to the synthetic task. I think this needs more analysis, see one concern below on this thread.
+* Distribution Leakage: I was not familiar with OpenThoughts3, but I looked into it and found that it was really close in distribution to GSM-8K and MATH-500. It aggregated reasoning problems. I wonder if this leads to inflated gains and the modest gains seen are from using OpenThoughts3 rather than the set up seen here. 
+* Hyperparameters: How are hyperparameters selected? My reading right now is that they are selected using the testing set... 
+* Verifier Signal: The core claim of the paper is that we can extract reliable signals from noisy self-verification via thresholded voting. But, when measured against a strong model with access to ground truth, the unsupervised verifier is only ~60% accurate on KK with Gemma‑4B. At scale, this is far from real preference labels. Even when making the model larger, you're still topping out at around 91%. 
+* Small Models: In the abstract, the authors claim that the approach "enhances reasoning in small models" yet Gemma‑1B shows little to no benefit. In the limitations section, this is acknowledged as well. I would reduce claims on generality in the abstract implied upfront. 
+* Novelty: Conceptually, I am having trouble differentiating between this method and self-refinement. To me, the paper is in this way, purely empirical. The major problem with this is that the main improvement the paper demonstrates is with KK and very limited improvement on the math benchmarks. Without stronger gains or deeper analysis, the contribution doesn't feel very strong to me.,
+
+### Questions
+* Could the authors discuss OpenThoughts3 and its training distribution a bit more?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper proposes Simple Generator–Verifier (GV) Games, a self-evolution framework that enables a single language model to generate and verify its own outputs without external supervision. By constructing high-confidence preference pairs through thresholded majority voting and optimizing via DPO, the model self-improves its reasoning ability. The authors demonstrate consistent gains in reasoning accuracy and easy-to-hard generalization across multiple benchmarks, showing that simple self-verification can approach supervised performance efficiently.
+
+### Strengths
+1. The paper conducts extensive experiments across multiple benchmarks, model scales, and ablation settings, demonstrating robustness and reproducibility.
+
+2. The problem setting—learning from unlabeled prompts and unverifiable tasks—is timely and important.
+
+3. The idea of turning a model’s own verification capability into a generator–verifier game, rather than relying solely on majority voting, is novel.
+
+### Weaknesses
+1. The paper lacks sufficient explanation of baseline methods such as INTUITOR, Absolute Zero (AZR), and AZR-Coder—it is unclear how these baselines are implemented or differ from the proposed approach.
+
+2. Although the related work section mentions Absolute Zero and R-Zero, there is no direct empirical or conceptual comparison, making it difficult to assess the advantage of the proposed method.
+
+3. Without a comparison to an external or supervised verifier, it is hard to evaluate how much of the improvement comes from self-verification itself versus incidental effects of DPO fine-tuning.
+
+4. Different tables use different model backbones (e.g., Table 1 uses Qwen2.5-7B-Instruct, while Table 4 uses Gemma-3-1B-it), raising concerns about consistency and fairness in comparison.
+
+5. Figure 2 shows that the SimpleGV verifier accuracy improves over the base model, but since the verifier is not explicitly trained, it is unclear why or how this improvement emerges.
+
+6. In Table 1, several baselines perform worse than their original reports (e.g., GRPO with Qwen2.5-7B drops from 90.2 → 82.9), yet no interpretation or justification is provided.
+ ---
+Presentation.\
+a. The gray-highlighted rows in the tables are difficult to interpret .\
+b. The numbers in Figure 5 are too small and hard to distinguish.\
+c. The texts in Table 3 are overly long, making it hard to understand each configuration clearly
+
+### Questions
+1. How would performance change if the verifier were trained or fine-tuned separately rather than sharing parameters with the generator?
+
+2. Is the model update applied only to the generator role, or does it indirectly improve verification ability as well?
+
+3. Could the authors clarify whether iterative GV training leads to verifier drift (i.e., changes in its reliability over iterations)?
+
+4. Could the authors provide more explanation and evidence for lines 264–266 — specifically, what causes “redundancy and verifier noise to begin to dominate” beyond a moderate model size?
+
+5. Could the authors include an error case analysis to better illustrate the failure modes of self-evolution?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper proposes as test time inference method relying on majority voting with an LLM as a judge utilized for a single-turn and multi-turn setup.
+
+### Strengths
+- The paper evaluates on several empirical domains such as knights and knaves and also performs ablations over thresholds and data and model sizes
+
+### Weaknesses
+- LLM as a judge and Majority Voting/Ensembling has been well studied in prior work (e.g [1]) thus is unclear the main contributions of this work with respect to these approaches
+- Several baselines (e.g MCTS) from the test time inference literature hasn't been compared to in this work, making it hard to judge the contribution of the proposed method
+
+[1] Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters
+
+### Questions
+See weaknesses above.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+5

human_reviews/Rip3EJc4nx.md

@@ -0,0 +1,175 @@
+## Human Reviewer 1
+
+### Summary
+The paper proposes a pruning criterion for large language models based on the information entropy of the model’s output distribution. Instead of using one-hot cross-entropy (which focuses on the single next token) or relying on a separate teacher for self-distillation, the method computes per-layer entropy-based importance scores within a Taylor-style framework and prunes parameters to better preserve the model’s global output distribution. The approach is label-free, aims to maintain fidelity after pruning, and is evaluated on zero-shot benchmarks for LLaMA and Qwen-family models.
+
+### Strengths
+The authors propose a relatively simple yet reasonable approach for pruning large language models (LLMs), which uses the entropy of output distributions as an indicator of neuron importance, instead of relying solely on next-token cross-entropy. The method does not depend on additional teacher models or complex distillation procedures, making it practically appealing for real-world pruning applications. The experiments cover multiple model families (e.g., LLaMA, Qwen) and several zero-shot benchmarks, showing that the proposed approach consistently outperforms existing pruning baselines. This provides some evidence of general applicability and reliability.
+
+### Weaknesses
+1. The dataset coverage is narrow and the evaluation tasks are of low difficulty, relying on a small set of relatively simple zero-shot benchmarks. This makes it difficult to assess robustness in areas such as instruction following, multi-step reasoning, long-context understanding, or multilingual and multi-domain settings.
+2. There may be a mismatch between the benchmarks and the pruned model components: if the tasks do not sufficiently engage the pruned submodules, the reported “distribution fidelity” may be overstated and external validity remains uncertain.
+3. Baseline and tuning transparency is limited, with missing comparisons to stronger or more recent pruning baselines, as well as a lack of systematic hyperparameter exploration under fair alignment conditions (e.g., matched training steps and learning-rate sweeps).
+4. The evaluation metrics are limited, focusing heavily on perplexity or simple zero-shot accuracy, without including direct generation results.
+
+### Questions
+1. The evaluation datasets are insufficient — they’re all too simple. The pruned parameters might not cover the activation patterns of these models, so more diverse datasets should be added.
+2. In addition, the model sizes don’t seem sufficient either, for example models like Qwen3-8B.
+3. For reasoning models, does the entropy criterion remain reliable?
+4. The main contribution is the use of entropy for pruning, which feels rather limited.
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+This paper introduces HFPrune (High-Fidelity Pruning), a structured pruning method for LLMs that aims to preserve model fidelity while reducing computational and memory costs, replacing the conventional loss-based Taylor pruning criterion with an information entropy–based criterion that measures the global prediction distribution of the model, instead of focusing only on the ground-truth token.
+
+### Strengths
+- Provides a label-free, holistic signal for neuron importance estimation. 
+- HFPrune consistently outperforms strong baselines: LLM-Pruner, LoRAPrune, SDMPrune, on LLaMA and Qwen families.
+- Comprehensive ablation studies validate the entropy criterion’s role in preserving output distributions. 
+- The algorithmic description is clear and reproducible. Implementation details are systematically reported.
+
+### Weaknesses
+- The pruning ratio $\rho{mlp}$  is fixed across all MLP layers, despite entropy potentially varying per layer, this could limit the functionality of HFPrune.
+- Lack a comparative discussion or empirical correlation analysis between entropy-based and Fisher-based importance scores.
+- Training-time FLOPs for fine-tuning (post-pruning recovery) are omitted.
+
+### Questions
+- Can a per-layer entropy-based adaptive cutting ratio improve the trade-off between fidelity and compression?
+- How does the entropy gradient behave in the low entropy versus high entropy regions of the output distribution?
+- What are the effects of using entropy calculated from logits vs. softmax probabilities?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+3
+
+### Rating
+4
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes HFPrune, a structured pruning method for LLMs that replaces traditional one-hot cross-entropy with information entropy as the criterion for Taylor-based neuron importance evaluation. The authors argue that entropy-based evaluation considers the full output distribution rather than just the ground-truth token, leading to better preservation of model capabilities. The method focuses on pruning MLP modules and demonstrates consistent improvements over existing methods across LLaMA and Qwen models on zero-shot benchmarks.
+
+### Strengths
+The method avoids computational overhead of teacher models and resolves gradient initialization issues in self-distillation approaches, showing 3x speedup over SDMPrune with 31% less memory usage.
+
+Demonstrates improvements across multiple model families (LLaMA, Qwen) and sparsity levels, with some configurations even exceeding dense model performance after fine-tuning.
+
+The approach is straightforward to implement, requiring only standard forward-backward passes without custom kernels or auxiliary models.
+
+### Weaknesses
+The paper fundamentally lacks theoretical justification for why entropy-based importance should preserve model performance. This is not a minor omission, in my view it's a central issue that undermines the contribution's scientific rigor.
+
+**Limited Evaluation Scope**: 
+- Exclusively focuses on zero-shot QA/classification tasks
+- No evaluation on reasoning, long-form generation, or conversational capabilities  
+- Largest model tested is only 7B parameters
+- Limited architectural diversity beyond LLaMA/Qwen families
+
+**Methodological Limitations**:
+- Uses uniform pruning ratios across layers, ignoring heterogeneous sensitivity
+- Limited sparsity range testing (only 20-30%)
+- The performance gains over dense models likely result from fine-tuning rather than pruning itself
+
+**Insufficient Analysis**: 
+- No explanation of which types of neurons are pruned and why (recent works have revealed super neurons, super weights, super experts etc.)
+- No investigation of the relationship between entropy reduction and specific capabilities
+- Missing analysis of method sensitivity to calibration data selection
+
+**Questionable Claims**: The assertion that the method "minimizes the change of global prediction distribution" is not rigorously nor theoretically established, and the connection between this and performance preservation remains speculative.
+
+### Questions
+1. Can you provide rigorous mathematical analysis of why minimizing entropy change should preserve model capabilities better than minimizing cross-entropy change?
+2. What is the information-theoretic justification for treating high-entropy neurons as more important?
+3. How does your approach relate to existing theories of neural network capacity and information flow?
+4. How does the method perform on reasoning tasks (GSM8K, BBH), long-context generation, and conversational AI beyond zero-shot classification?
+5. Can you evaluate on larger models (70B+) and more diverse architectures to support generalizability claims?
+6. What explains the performance improvements over dense models - is this due to fine-tuning effects rather than pruning benefits?
+7. Why use uniform pruning ratios instead of layer-sensitive approaches?
+8. How sensitive is the method to calibration data selection and size?
+9. How does the method perform under more aggressive pruning (40-70% sparsity)?
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes HFPrune, a novel method for compressing LLMs through pruning of the MLP modules within transformer architectures. HFPrune introduces an information entropy-based criterion for evaluating neuron importance, offering a more holistic approach than traditional methods, which rely on one-hot cross-entropy loss. This entropy-based method minimizes the global prediction distribution change, effectively preserving model performance.  Extensive experiments on LLaMA and Qwen models demonstrate the effectiveness and efficiency.
+
+### Strengths
+The proposed idea is simple, straightforward, and aligns well with intuition. From the engineering perspective, it is also very easy to implement.
+
+Extensive experiments were conducted on multiple LLM models, across diverse benchmarks, comparing the proposed method with several previous methods. The results demonstrate significant improvements in both performance and efficiency of the proposed method.
+
+### Weaknesses
+**Explanation on the Choice of Baselines**: As discussed in the Related Work section, LLM pruning is a highly focused research area with a large body of work, which can be categorized into different approaches. However, in the experimental section, only a few methods such as LLM-pruner, LoRAPrune, and LoRAP are compared, and the rationale behind choosing these baselines is not explained. It remains unclear whether the state-of-the-art methods from each category are all covered by the baselines used in this paper. If they are, an explanation should be provided; if not, more baselines should be included (or a justification should be given for why they cannot be included for a fair comparison).
+
+**More Comparison on Efficiency**:  For Efficiency, the paper only compares the proposed method with SDMPrune, demonstrating that the proposed approach is more efficient than methods like SDMPrune, which require a teacher model. The efficiency of other methods (LLM-pruner, LoRAPrune, LoRAP) should also be compared.
+
+**Writing Needs Improvement**:
+- For example, in lines 013-014 of the Abstract: The statement "A common approach uses Taylor Expansion" does not clearly explain the field of research. It should at least mention that "it is a common approach for LLM pruning."
+- In line 015 of the Abstract, the sentence "However, its reliance on one-hot cross entropy loss, ..." contains a grammatical error. It should be "it relies on" instead.
+- Many more problems on writing exists
+
+### Questions
+**Regarding whether "focus on pruning MLP" is the contribution of this paper**: In lines 039-043, when explaining why the paper focuses on pruning MLP, the authors use phrases like "We find that" and "we observe that" (and then prove this in Section 5.3.3 with experiments), but these statements are also referencing previous works. Is this an innovative contribution of the paper, or is it the findings from previous works?
+
+**considering a subset of tokens**：Between the extremes of "only consider the ground-truth label token" and "consider all possible tokens", an intermediate approach could be: only considering a subset of tokens (for example, the top K tokens with the highest probability) when evaluating neuron importance. Did the authors experiment with this approach? Experiments or discussions on this would make the paper more insightful.
+
+### Soundness
+3
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+5

human_reviews/SkmkGKEZ1U.md

@@ -0,0 +1,188 @@
+## Human Reviewer 1
+
+### Summary
+The paper introduces **O-Forge**, a workflow that couples a frontier LLM with a computer algebra system (CAS), specifically Mathematica’s Resolve, to prove asymptotic inequalities by (i) asking the LLM to propose a domain (or index) decomposition and (ii) using Resolve to verify each piece via first-order quantifier elimination. A toy inequality and a series bound from Tao are used as case studies; the paper claims the approach “moves beyond contest math” by offloading the creative split to the LLM and the verification to CAS. The system exposes a CLI and a website front-end; evaluation is described as ~40–50 “easier problems,” with qualitative takeaways about typical numbers of subdomains and the usefulness of leading-term simplifications.
+
+### Strengths
+N.A.
+
+### Weaknesses
+There are too many drawbacks in this paper. In general, this submission is more like a blog post instead of a rigorous paper as it lacks of enough and solid experiments and evaluations to demonstrate its claims. Some of the weaknesses are as follows.
+- **Lack of Novelty:** LLM + CAS could be viewed as part of Tool-use LLM research. The proposed idea is not novel.
+- **Insufficient rigor and experimental substance:** The “evaluation” consists of two main case studies plus ~“40–50 easier problems,” but there are no well-defined benchmarks, success metrics, or failure analyses (rates of correct/incorrect splits, wall-clock, query counts, sensitivity to prompts, comparisons to baselines like hand-crafted heuristics or SMT-based pipelines, etc.). As is, this reads more like a prototype report/blogpost than an ICLR-level empirical study.
+- **Underspecified method details:** Key pieces are missing or skeletal. For example, the “structured prompt” is left with placeholders (“describe the structure of the prompt”), and the code snippet for Resolve is fragmentary; the search over constants C is a coarse grid without justification or sensitivity analysis. This makes the approach hard to reproduce or evaluate scientifically.
+- **Heavy reliance on closed-source Mathematica without proof objects:** While Resolve is powerful, the paper acknowledges there is no proof term and asks the reader to trust a closed system; this undermines the claim of “rigorous verification,” especially for research-level math. No attempt is made to cross-check with open tools (e.g. SageMath) on a subset, or to export certificates.
+- **Reproducibility & accessibility concerns:** Running the system requires Mathematica and a frontier LLM API, making reproduction costly; even the authors’ ethics section notes access costs. There is a website, but that can’t substitute for open artifacts or independent verification.
+- **Scope creep vs. precise problem definition:** The paper oscillates between “asymptotic inequalities” and the specific case study. I could feel the motivation of this paper: making a true AI4Math tool or project to better help professional mathematicians instead of some fuzzy LLMs that could only do math questions. But there is no crisp task definition, no sanity check for the motivation, and no quantitative experiments to show the helpfulness and effectiveness. This vagueness makes it hard to judge whether this is something scientifically helpful or just some course project.
+
+### Questions
+- Please use \citep for non-subject/object citations.
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+5
+
+---
+
+## Human Reviewer 2
+
+### Summary
+Following Terry Tao’s proposed algorithm, a system for asymptotic analysis is implemented. It combines an LLM for the creative part of sub-domain choice, and uses Mathematica for the rest of the steps - notably “Resolve” for proofs in specific sub-domains.
+
+2 non-trivial asymptotic bounds were proven via this system.
+
+### Strengths
+If works well, applicable to a broad range of scientific endeavors.
+
+Final results are grounded via a reliable CAS system.
+
+Shown proof of concept for useful mathematical results.
+
+### Weaknesses
+Line 86: “(** describe the structure of the prompt**)”...
+This is unprofessional at best.
+
+Line 91: All the mentioned solvers need citations, as well as the Mathematica Resolve function - due to its central role in your system.
+
+Line 100: Fig.1 is mentioned initially on page 2 but appears on page 4. Why?
+
+Line 101 (and multiple others): “o-forge.com”.
+At the top right corner of the front page of this website it clearly states:
+
+Created by
+
+Vijay Ganesh
+
+Ayush Khaitan
+
+Violating the ICLR guidelines regarding anonymity. 
+
+
+Regarding the website itself, while I liked the UI and readily available examples, it sometimes returns weird results.
+For example, Example Series 1 returns a summary that contains a single word: “The”. I ran it a few times to make sure it’s not some random LLM aberration.
+Other times the output shows a Python error stack trace (“Execution Failed”).
+Seems underbaked and the code is unstable.
+
+Line 109,113,119 (and others): Citations. You’re mentioning Prof. Tao many, many times - but he is a prolific mathematician. Point to the specific works you use.
+
+Line 123: While I personally agree with this claim on (most) interesting series bounds, it needs to be quantified (benchmark vs. other methods), or at least supported by multiple examples from the literature.
+
+Line 127: The novelty claim is unclear to me. It seems like an engineering project - implementing a (good) idea by Prof. Tao. While I agree that such an online tool can be useful for mathematicians around the world - and would like to encourage you to improve it and fix the bugs - the system does not constitute *scientific* innovation done by you.
+
+The generate (via LLM) -> verify (via CAS) approach was done in multiple projects (though not specifically in your use case) - so the core approach is also not novel on its own.
+
+If the strongest results are the proof of the specific use-case (which to my understanding is indeed novel, but I don’t know enough to say how impactful this singular result is), then perhaps consider submitting to a mathematical / experimental math journal?
+
+Line 295: I’m a supporter of readable papers and not-too-formal language, but this is too much. 
+The paper is not your blogpost.
+
+Line 348: What do you mean “around 40-50”??
+
+Section 5: You need concrete statistics for all the claims here. Currently it’s anecdotal.
+
+Section 6: There are multiple other works in AI for Math that apply combinations of CAS/code generation + LLMs. The current overview is quite limited.
+
+### Questions
+Line 256: This “elaborate Mathematica code’ seems like an important part of the system - perhaps ~50% of its power?
+
+Line 270: Once in the entire process? If I’m reading the outputs on your website correctly, there are recursive attempts to re-define the sub-domain partition, until you get “True” on all of them?
+
+Line 328: How is the output passed to Mathematica? Is there a specific format that you force it to use in its output? Do you parse the output text and translate it to Mathematica code? If so, how?
+
+Line 359: How do you count decompositions? Number of separate domains or number of boundaries?
+
+Line 367: Any ideas why this happens? 
+How do you do this replacement?
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+2
+
+### Rating
+0
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+he paper presents O-Forge, a system that integrates a large language model (LLM) with a computer algebra system (CAS) to aid in the verification of asymptotic inequalities. This is an instance of classical synthesis paradigms such as oracle guided inductive synthesis combining an inductive LLM with deductive reasoning system to generate formal artifacts. The paper describes this as an “In-Context Symbolic Feedback Loop”.  The LLM proposes domain decompositions (i.e., how to split a problem into manageable subdomains). The CAS (via Mathematica’s Resolve function) then verifies whether each subdomain satisfies the proposed inequality using first-order logic and quantifier elimination.
+
+### Strengths
+The authors claim their tool can handle research-level asymptotic inequalities, going beyond standard competition-style problem solving by combining LLM creativity with CAS rigor. This is left to some subjective interpretation. 
+Two case studies: an asymptotic AM-GM inequality and a series decomposition, are used to demonstrate the concept.
+
+### Weaknesses
+The paper reads more like a concept demo or blog post than a rigorous scientific study, lacking detailed quantitative and ablation studies with proper baselines. Testing on self-curated  "suite of around 40-50 easier problems" and a few case studies falls far short of the evaluation expected of a research paper. 
+
+Hybrid symbolic–neural approaches (e.g., Lean+LLM, AlphaProof, Autoformalization pipelines; see https://arxiv.org/abs/2310.17807, https://ieeexplore.ieee.org/document/10356332) have already explored the same broader paradigm with planning, code generation, formal proof verification, not just heuristic checking. The authors present O-Forge as “revolutionary,” yet it is essentially prompting an LLM for suggestions and sending them to a formal tool - Mathematica. 
+
+Crucial implementation details are omitted. How exactly is the LLM prompted? How is the “in-context feedback” loop structured? How is the decomposition quality evaluated or improved iteratively? Are there failure modes where the LLM produces incorrect decompositions, and how are these handled?
+
+### Questions
+Can you expand experimental evaluation and share quantitative metrics (success rate, runtime, size of inequalities handled) over larger benchmark suite to substantiate performance?
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper introduces a CAS + LLM system for proving asymptotic expressions. The primary thrust of the system is to use a LLM to propose decompositions and then use a CAS to test their correctness.
+
+### Strengths
+Due to the very unusual nature of the weaknesses, I don't have any strengths to comment on.
+
+### Weaknesses
+This paper does not appear to contrain any evidence as to its effectiveness. Section 5 begins "In addition to the above-mentioned case study of hard problems, we tested our tools on an extensive suite of around 40-50 easier problems, in order to study how well it performs on a diverse set of inequalities," but there is no mention of these "hard problems" anywhere in the paper. Additionally, for these easy problems, the paper presents three high level takeaways but no actual evidence.
+
+### Questions
+Did I misunderstand something? Does this paper present evidence of its correctness?
+
+### Soundness
+1
+
+### Presentation
+1
+
+### Contribution
+1
+
+### Rating
+0
+
+### Confidence
+4

human_reviews/SpUXijnBEg.md

@@ -0,0 +1,166 @@
+## Human Reviewer 1
+
+### Summary
+Direct Optimal Action Learning (DOAL) simplifies offline RL policy extraction by decoupling optimal action computation from policy training. It achieves this by generating a synthetic "optimal" action via Q-function ascent, then uses efficient behavior cloning to train the policy to match it. This framework, which also introduces a more stable trust region hyperparameter, consistently outperforms strong baselines across various policy types and benchmarks.
+
+### Strengths
+1. Novel Policy Extraction. DOAL elegantly decouples optimal action computation from policy training, sidestepping complex backpropagation through generative models by reframing optimization as simple behavior cloning.
+2. Improved Hyperparameter Stability. Reinterpreting the BRAC trade-off coefficient $\alpha$ as a normalized trust region $\delta$, backed by Proposition 2 and the Batch-Normalizing Optimizer, offers a more interpretable and stable hyperparameter. Empirical evidence shows $\delta$ can be shared across algorithms, simplifying tuning.
+3. Strong Empirical Results. DOAL achieves consistent and significant performance gains across 15 challenging offline RL tasks, notably improving over meticulously tuned baselines, which enhances the credibility of its effectiveness.
+
+### Weaknesses
+1. Inconsistency in Theorems and practical implementation: Proposition 1 establishes the gradient equivalence by defining $a^{\text {target }}$ using the Q-gradient evaluated at the policy's output, $\pi_\theta(s)$. However, the paper then argues this is conceptually inconsistent and, in the final DOAL objective (Eq. 16), defines the target using the Q-gradient evaluated at the data action, $a$. While this change is key to decoupling, the original equivalence from Proposition 1 no longer strictly holds, leaving a gap in the theoretical justification for the final objective.
+2. The paper introduces the Batch-Normalizing Optimizer and the trust-region parameter $\delta$ as a more stable replacement for the sensitive hyperparameter $\alpha$ from the BRAC objective. Despite this, the final DOAL actor loss in Equation 15 still contains an $\alpha$ parameter, which is described as a controller for the "learning rate of actor" and is copied from a prior work. This is confusing. It is unclear how this $\alpha$ interacts with $\delta$ and why it is still necessary if the optimization trade-off is now managed by the trust region. This ambiguity detracts from the otherwise clean narrative of simplifying hyperparameter tuning.
+3. In Table 1, some experimental results with D- still behind the vanilla algorithms. Also, the reviewer suggests that the authors to better present this table such as showing the final averaged performance.
+
+### Questions
+see weakness
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+2
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper proposes Direct Optimal Action Learning (DOAL), a framework for offline RL that reframes behavior-regularized actor–critic training as directly learning an “optimal” action target for each data point and then imitation-learning that target using losses native to the policy family (Gaussian Policies, flow policies and diffusion policies). This avoids costly backpropagation through iterative sampling chains in expressive policies (diffusion/flow), while keeping value guidance from the learned Q-function. The authors also introduce a Batch-Normalizing Optimizer that sets a dataset-level “trust region” (δ) to control how far targets move from behavior actions, aiming to make tuning more interpretable and consistent across policy distributions. Empirically, across 15 tasks from OGBench and D4RL Adroit, DOAL variants match the performance of strong baselines (IQL/FQL/TrigFlow) with shared, environment-level hyperparameters.
+
+### Strengths
+The paper presents a novel and elegant reformulation of behavior-regularized actor–critic training by directly learning an optimal per-sample action target and then imitating that target with policy-native losses, which removes the need for costly backpropagation through iterative sampling chains in flow and diffusion-based policies. This idea is conceptually clean, improves computational practicality for expressive policy classes, and introduces a trust-region–style batch normalization that provides a more interpretable and stable hyperparameter compared with traditional BRAC scaling. Empirically, across OGBench and D4RL Adroit, the method achieves performance comparable to strong baselines (IQL/FQL/TrigFlow) while using shared environment-level hyperparameters, suggesting that the proposed reformulation maintains effectiveness without extensive tuning
+
+### Weaknesses
+1. Writing & clarity: The paper has noticeable grammatical issues, inconsistent notation, and several typos, which make the narrative harder to follow; The main results table is dense and difficult to scan, hindering quick cross-method comparisons across policy families and environments; consider adding per-method aggregates (e.g., mean/median across tasks) to improve readability.
+
+2. Shallow experimentation: The evaluation covers only 15 tasks total—6 from D4RL Adroit and 9 from OGBench. Moreover, results are averaged over just 4 seeds (D4RL) and 3 seeds (OGBench), which weakens statistical confidence and robustness of the findings.
+
+3. No compute analysis: A central claimed benefit is reduced computational cost by avoiding backpropagation through iterative diffusion/flow sampling chains, yet the paper provides no wall-clock or memory measurements to substantiate this. Even a simple per-epoch time or training-step latency comparison against BRAC-style training would clarify the practical value.
+
+4. Missing ablations: There are no targeted ablations to disentangle contributions from (i) direct optimal-action targets, (ii) the Batch-Normalizing optimizer; the current analysis discusses hyperparameters but doesn’t isolate causal impact on performance.
+
+### Questions
+See Weaknesses
+
+### Soundness
+2
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 3
+
+### Summary
+The paper presents Direct Optimal Action Learning (DOAL), a framework that replaces end-to-end BRAC-style policy gradients through the value network with a two-step policy extraction:
+(1) compute an “optimized” target action per data point via a first-order step on the Q-function;
+(2) train the policy to imitate this target with a behavior-native loss (e.g., Gaussian MSE, flow matching, diffusion reconstruction).
+This yields distribution-agnostic policy extraction without backpropagating through iterative samplers used by diffusion/flow policies. The authors reinterpret the BRAC trade-off as a trust region and introduce a Batch-Normalizing Optimizer that scales the action step by the batch norm of |\∇ₐQ|, controlled by a single hyperparameter δ. Experiments on OGBench and D4RL Adroit show consistent improvements of DOAL variants (Gaussian, flow, diffusion) over their respective baselines, and analyze Max-Q sampling and the role of candidate count 
+𝑛
+sample
+n
+sample
+	​
+
+.
+
+### Strengths
+Simple, general recipe for policy extraction. Shows BRAC’s policy gradient can be reframed as target-matching against an action updated by ∇ₐQ, enabling training with any policy family’s native loss (Gaussian/flow/diffusion) and avoiding gradients through multi-step samplers.
+
+Interpretable trust region. The Batch-Normalizing Optimizer controls expected step size via δ, normalizing by batch statistics of |\∇ₐQ|²—cleaner than tuning BRAC’s 𝛼.
+
+Covers expressive policies. Instantiates DOAL for Gaussian, Flow Matching, and Diffusion policies with concrete objectives (e.g., DIQL, DIFQL, DTrigFlow).
+
+### Weaknesses
+Novelty is mainly a reinterpretation plus a practical recipe.
+The equivalence that turns BRAC into target-matching is neat but technically light. Stronger differentiation from value-guided diffusion/flow, energy guidance, and behavior-regularized objectives would clarify what DOAL achieves beyond engineering convenience.
+
+Value-quality dependence and limited uncertainty handling.
+DOAL targets rely on ∇𝑎𝑄 at data actions; there’s no integrated uncertainty-aware step control (ensembles, variance-based scaling) for OOD safety, which matters in offline regimes.
+
+Benchmark scope and stability.
+While OGBench and Adroit are covered, Adroit volatility and late-training collapse are reported without a full diagnosis, weakening stability claims.
+
+### Questions
+1. Target evaluation point. You compute ∇𝑎𝑄 at data actions to avoid a mismatch with 𝜋𝜃(𝑠). Have you compared evaluating at 𝜋𝜃(𝑠） (or hybrids) and quantified the difference?
+
+2. δ scheduling. Does adaptive or annealed δ improve late-training stability on Adroit where collapses occur?
+
+3. Value backbones. What happens when pairing DOAL with ReBRAC-style critics—do gains persist?
+
+4. High-D actions. For diffusion/flow in high-D action spaces, does anisotropy in ∇𝑎𝑄 hurt target-matching? Any whitening or per-dimension scaling beyond batch normalization? 
+
+5. Max-Q proposals. Beyond sampling from 𝜋𝜃, can proposal refinement (e.g., short Langevin steps guided by Q) reduce noise sensitivity for large 𝑛 samples?
+
+### Soundness
+3
+
+### Presentation
+2
+
+### Contribution
+2
+
+### Rating
+6
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+The paper proposes Direct Optimal Action Learning (DOAL), a framework for offline RL that avoids backpropagating through multi‑step generative policies (diffusion/flow) when using BRAC‑style actor objectives. The key observation is that the BRAC policy gradient is (approximately) equivalent to minimizing the distance between the policy output and a single “optimal action” target derived via a first‑order update from the dataset action. DOAL computes that target directly from \nabla_a Q_\phi(s,a) (evaluated at the dataset action) and then trains the policy to imitate the target using a loss native to the policy family (e.g., flow matching or a diffusion loss), thus decoupling target computation from the policy’s sampling chain. The paper further replaces the sensitive BRAC coefficient \alpha with a Batch‑Normalizing Optimizer that scales the update so that the expected squared step size equals a user‑chosen trust‑region parameter, and presents Algorithm 1 showing integration with IQL for value learning. Experiments on 9 OGBench tasks and 6 D4RL Adroit tasks compare Gaussian, flow, and diffusion policies show mixed results.
+
+### Strengths
+- Clear, unifying insight and simple objective. This paper makes explicit that BRAC’s policy gradient equals the gradient of a squared‑error to a target action. This clarifies what end‑to‑end training is doing and motivates learning the target directly, which is easy to implement and compatible with any policy class.
+- DOAL avoids backprop through iterative sampling chains, which is particularly important for diffusion/flow actors by supervising with native behavior losses. This makes the approach broadly plug‑and‑play with flow/diffusion policies.
+- Hyperparameter reinterpretation via a trust region. The Batch‑Normalizing Optimizer replaces a brittle \alpha search with an interpretable \delta that sets an expected squared update magnitude; the denominator’s batch statistic stabilizes scale across tasks.
+
+### Weaknesses
+- To me, it's unprincipled to replace a’ that which should be sampled from the policy with the dataset action a, the paper does not provide a solid justification for this change (cf. Prop. 1/Eq. 13). In addition, the derivation appears to hinge on an L2 BC loss—how does the argument extend to other losses (e.g., log-likelihood, flow-matching, diffusion objectives)? 
+- Results are not uniformly stronger. The claim that DOAL “consistently improves performance” over strong baselines is only partially supported. For example, DTrigFlow is dramatically worse on cube‑single‑play, and DIQL is broadly weaker than IQL on many tasks. 
+- Efficiency claims lack measurement. A central motivation is avoiding BPTT through iterative samplers, but there are no wall‑clock time, memory, or gradient‑call counts comparing DOAL to TrigFlowQL/FQL under matched hardware.
+- The observation of Max-Q sampling does not constitute a novel contribution. Prior work (e.g., Q-chunking) already formalizes that the candidate count N in Max-Q sampling implicitly sets the level of KL regularization between the induced policy and the proposal/sampling distribution. From this perspective, increasing N weakens the effective regularization and can degrade performance, so it is unsurprising that “the bigger N, the better” does not hold. Please cite these results.. 
+- Scope of value learners. Most experiments couple DOAL with IQL for a controlled study, but since BRAC‑style methods remain strong (e.g., ReBRAC in Table 1). Is this policy extraction approach still improve on other value learners?
+
+### Questions
+see Weaknesses.
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+2
+
+### Confidence
+5

human_reviews/SsUjdSVdUl.md

@@ -0,0 +1,156 @@
+## Human Reviewer 1
+
+### Summary
+The paper proposes Critique-RL, a two-stage reinforcement learning (RL) framework for training critiquing language models. The critic  models that assess the correctness of an actor’s response and provide natural language feedback to guide refinement. The core motivation is to avoid reliance on stronger supervisors for critique annotation, which is costly and hard to scale. The key insight is that optimizing solely via indirect rewards, whether the actor’s refined output is correct, leads to poor discriminability, the critic’s ability to accurately judge whether an initial response is correct. To address this, Critique-RL introduces a two-stage training process. Experiments on mathematical reasoning benchmarksshow consistent improvements over baselines.
+
+### Strengths
+1. The problem setup, motivation, and methodology are clearly articulated with intuitive figures (e.g., Figure 2, 3) and precise definitions of evaluation metrics
+2. Scalable oversight remains a critical bottleneck in LLM alignment and self-improvement. By explicitly disentangling and jointly optimizing two key critique capabilities, the work offers a practical and empirically validated pathway toward more reliable critique models.
+3. The identification of problem due to conflicting objectives under indirect rewards is compelling.
+
+### Weaknesses
+1. The paper claims to avoid “stronger supervisors” for critique data, yet relies heavily on ground-truth answers (i.e., golden labels) to construct both direct (Stage I) and indirect (Stage II) reward signals. These labels are typically human-annotated or derived from curated datasets (e.g., MATH, GSM8K). Thus, the method still depends on human-provided supervision, albeit not in the form of natural language critiques. This undermines the central motivation that existing approaches “rely on stronger supervisors for annotating critique data”—since here, human-labeled answers serve as an equally strong (if not stronger) form of supervision.
+2. Lack of analysis of some highly related works in this paper [1], which also proposes to use the refinment as the additional feedback.
+
+> [1] Training Language Models to Critique With Multi-agent Feedback
+
+### Questions
+No question
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+3
+
+---
+
+## Human Reviewer 2
+
+### Summary
+The paper addresses the challenge of insufficient scalable supervision for LLMs in complex reasoning tasks. It proposes Critique-RL, a two-stage reinforcement learning framework without strong supervision. In the first stage, the critique model’s discriminative ability is optimized using direct rule-based rewards. In the second stage, the framework integrates indirect rewards from agent-corrected response accuracy with regularization to enhance the usefulness of feedback while preserving discriminative capacity. Critique-RL achieves  performance improvements over baselines on several mathematical reasoning and question-answering tasks.
+
+### Strengths
+- The proposed two-stage RL method is effective to provide constructive critique feedback for better refinement and precise filter for effective time-time scaling.
+- The experiments contain several benchmarks across different tasks.
+- This paper is well-written and easy to follow.
+
+### Weaknesses
+My main concern lies in the experimental design, as I am not fully convinced that the current experiments sufficiently demonstrate the proposed method’s advantage on complex reasoning tasks.
+
+- Since the authors explicitly state in the Abstract and Introduction that their method targets complex reasoning tasks, more challenging benchmarks such as AIME and GPQA should have been included in the evaluation.
+
+- Although the main text reports significant improvements on Qwen-3B and Qwen-7B, the appendix reveals that the performance gains on stronger models, such as DeepSeek-R1-Distill-Qwen-7B and Qwen2.5-72B-Instruct, are quite limited. These results should be reported and discussed in the main paper.
+
+- The proposed method requires access to ground-truth answers to compute rewards. Under this setting, it remains unclear what advantages it offers over directly training the model’s reasoning ability using RLVR methods (e.g., GRPO, DAPO). Additional experiments are needed to clarify this distinction.
+
+### Questions
+See weakness
+
+### Soundness
+2
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 3
+
+### Summary
+This paper proposes Critique-RL, a two-stage reinforcement learning (RL) approach to train critiquing language models without requiring stronger supervision. The authors first show that baseline RL methods, which use only indirect rewards from an actor's refinement, fail. This is because they improve the critic's helpfulness (constructive feedback) but not its discriminability (judging correctness), leading to poor performance. Critique-RL solves this by:
+- Stage I: Explicitly optimizing discriminability using direct, rule-based reward signals.
+- Stage II: Optimizing helpfulness using indirect rewards (actor refinement) while using regularization to maintain the discriminability from Stage I.
+
+This two-stage strategy delivers performance improvements on both in-domain and out-of-domain tasks, e.g., +9.02% in-domain and +5.70% OOD for Qwen2.5-7B.
+
+### Strengths
+- The paper's core originality is its clear diagnosis of a key failure mode in training critics: baseline RL methods create a conflict between "discriminability" (judging correctness) and "helpfulness" (providing feedback), optimizing the latter at the expense of the former.
+- The paper's quality is high, with a rigorous methodology. The training dynamics in Figure 3 clearly show the baseline's failure , while decisive ablation studies in Table 3 prove that both stages of Critique-RL and its specific regularization are essential for success.
+- The work significantly contributes to scalable oversight by providing an effective method to train critics without stronger supervisors. Its value is shown through broad applicability, including "weak-to-strong" generalization (a 7B critic improving a 72B actor) and effectiveness on OOD and open-ended tasks .
+
+### Weaknesses
+- The paper's primary motivation is to train critics "without stronger supervision"1. However, the entire method, especially the critical Stage I, is heavily reliant on an "oracle reward function" $r_{oracle}(x,y)$ to compute the direct discrimination reward $r_{dis}$. For the main experiments on math tasks, this oracle is a rule-based verifier that knows the correct answer. This oracle is a form of strong, external supervision.
+- The framework's success, particularly in Stage II, hinges on a critical assumption: the fixed actor model $\pi_{\theta}$ is already a good "refiner". The authors state the actor is pre-trained to be "capable of... faithfully refining them according to critiques". This assumes away a large part of the problem. The helpfulness reward $r_{refine}$ is a convolved signal of both the critique's quality and the actor's ability to understand it. If the actor is a poor refiner, $r_{refine}$ becomes a noisy or meaningless signal, and Stage II would fail to optimize for helpfulness.
+- The paper correctly highlights its inference-time compute-efficiency benefits (e.g., in Figure 1 and Figure 6). However, it completely omits the training-time cost of this complex, two-stage RL pipeline. Stage II, for example, requires at least three model forward passes per training sample (one for the critic $c=\pi_{\phi}(x,y)$, one for the actor's refinement $y^{\prime}=\pi_{\theta}(x,y,c)$, and one for the oracle/RM $r_{refine}=r_{oracle}(x,y^{\prime})$). This is significantly more expensive than the SFT or baseline RL methods it's compared against.
+- The paper's core insight is that final-outcome rewards are insufficient, and a more direct signal is needed. The proposed solution, $r_{dis}$, is a direct reward for judging the outcome of the original response. This overlooks a more direct comparison to Process-based Reward Models (PRMs), which provide supervision at each step of the reasoning. The qualitative examples (Figs. 8, 9) show the critic is evaluating step-by-step, but it is only trained on the final answer's correctness.
+
+### Questions
+The paper is motivated as an approach for training critique models "without stronger supervision". However, the method's crucial first stage relies entirely on a direct reward $r_{dis}$ from an "oracle reward function" $r_{oracle}(x,y)$. This oracle, which knows the ground-truth correctness, seems to be a form of strong supervision.
+
+Could you please clarify this apparent contradiction?
+
+1. How do you formally define the "stronger supervision" (which you avoid) versus the "oracle verifier" (which you use)?
+2. The paper's contribution seems to be a novel way to distill the knowledge of a "weak" binary oracle (answer-checker) into a "strong" generative critic (feedback-generator). Would you agree with this framing?
+3. How does this framework scale to complex domains (e.g., creative writing, complex coding) where no such simple oracle or high-quality reward model exists?
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+2
+
+### Rating
+4
+
+### Confidence
+4
+
+---
+
+## Human Reviewer 4
+
+### Summary
+This paper proposes an online RL approach called Critique-RL for developing critiquing language models without stronger supervision. This approach contains a two-player paradigm, where the actor generates a response, the critic provides feedback, and the actor refines the response accordingly. The authors devise a two-stage optimization strategy, where stage I reinforces the discriminability of the critic with direct rule-based reward signals and stage II introduces indirect rewards based on actor refinement to improve the critic’s helpfulness. Experimental results show the effectiveness of Critique-RL.
+
+### Strengths
+1. The proposed two-stage RL method is sound and well-motivated, which deals with the core problem of critique generation.
+2. Extensive experiments show the effectiveness of the proposed method.
+3. This paper is overall well-written and easy to follow.
+
+### Weaknesses
+1. The design of indirect rewards based on actor refinement is similar to [1], which is not discussed in the current paper. The authors should further clarify the difference between this work and [1] to highlight their novelty.
+
+2. The quality of generated critiques should be individually measured via automatic metrics or human evaluation.
+
+
+[1] Training Language Model to Critique for Better Refinement. ACL 2025 Findings.
+
+### Questions
+I have included my questions in the weaknesses part.
+
+### Soundness
+3
+
+### Presentation
+3
+
+### Contribution
+3
+
+### Rating
+6
+
+### Confidence
+4